diff --git a/.gitattributes b/.gitattributes index 480dbf0d14f39c1271839ca23fcb31e42c049477..4b2b308cf07de134a2ff6156d7e054d3e4371ddc 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1,215 +1,222 @@ -*.7z filter=lfs diff=lfs merge=lfs -text -*.arrow filter=lfs diff=lfs merge=lfs -text -*.bin filter=lfs diff=lfs merge=lfs -text -*.bz2 filter=lfs diff=lfs merge=lfs -text -*.ckpt filter=lfs diff=lfs merge=lfs -text -*.ftz filter=lfs diff=lfs merge=lfs -text -*.gz filter=lfs diff=lfs merge=lfs -text -*.h5 filter=lfs diff=lfs merge=lfs -text -*.joblib filter=lfs diff=lfs merge=lfs -text -*.lfs.* filter=lfs diff=lfs merge=lfs -text -*.mlmodel filter=lfs diff=lfs merge=lfs -text -*.model filter=lfs diff=lfs merge=lfs -text -*.msgpack filter=lfs diff=lfs merge=lfs -text -*.npy filter=lfs diff=lfs merge=lfs -text -*.npz filter=lfs diff=lfs merge=lfs -text -*.onnx filter=lfs diff=lfs merge=lfs -text -*.ot filter=lfs diff=lfs merge=lfs -text -*.parquet filter=lfs diff=lfs merge=lfs -text -*.pb filter=lfs diff=lfs merge=lfs -text -*.pickle filter=lfs diff=lfs merge=lfs -text -*.pkl filter=lfs diff=lfs merge=lfs -text -*.pt filter=lfs diff=lfs merge=lfs -text -*.pth filter=lfs diff=lfs merge=lfs -text -*.rar filter=lfs diff=lfs merge=lfs -text -*.safetensors filter=lfs diff=lfs merge=lfs -text -saved_model/**/* filter=lfs diff=lfs merge=lfs -text -*.tar.* filter=lfs diff=lfs merge=lfs -text -*.tar filter=lfs diff=lfs merge=lfs -text -*.tflite filter=lfs diff=lfs merge=lfs -text -*.tgz filter=lfs diff=lfs merge=lfs -text -*.wasm filter=lfs diff=lfs merge=lfs -text -*.xz filter=lfs diff=lfs merge=lfs -text -*.zip filter=lfs diff=lfs merge=lfs -text -*.zst filter=lfs diff=lfs merge=lfs -text -*tfevents* filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.fconst filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.super filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.super filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.super filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.super filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.super filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.super filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.super filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.fconst filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.super filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.meta filter=lfs diff=lfs merge=lfs -text -common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-canny/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-depth/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-pose/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/controlnet-tile/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.super filter=lfs diff=lfs merge=lfs -text -normal/transformer/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text +*.7z filter=lfs diff=lfs merge=lfs -text +*.arrow filter=lfs diff=lfs merge=lfs -text +*.bin filter=lfs diff=lfs merge=lfs -text +*.bz2 filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text +*.ftz filter=lfs diff=lfs merge=lfs -text +*.gz filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.joblib filter=lfs diff=lfs merge=lfs -text +*.lfs.* filter=lfs diff=lfs merge=lfs -text +*.mlmodel filter=lfs diff=lfs merge=lfs -text +*.model filter=lfs diff=lfs merge=lfs -text +*.msgpack filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text +*.onnx filter=lfs diff=lfs merge=lfs -text +*.ot filter=lfs diff=lfs merge=lfs -text +*.parquet filter=lfs diff=lfs merge=lfs -text +*.pb filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text +*.pt filter=lfs diff=lfs merge=lfs -text +*.pth filter=lfs diff=lfs merge=lfs -text +*.rar filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text +saved_model/**/* filter=lfs diff=lfs merge=lfs -text +*.tar.* filter=lfs diff=lfs merge=lfs -text +*.tar filter=lfs diff=lfs merge=lfs -text +*.tflite filter=lfs diff=lfs merge=lfs -text +*.tgz filter=lfs diff=lfs merge=lfs -text +*.wasm filter=lfs diff=lfs merge=lfs -text +*.xz filter=lfs diff=lfs merge=lfs -text +*.zip filter=lfs diff=lfs merge=lfs -text +*.zst filter=lfs diff=lfs merge=lfs -text +*tfevents* filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.fconst filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.super filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.super filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.super filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.super filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.super filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.super filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.super filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.fconst filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.super filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.meta filter=lfs diff=lfs merge=lfs -text +common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-canny/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-depth/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-pose/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/controlnet-tile/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.ctrlpkt filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.meta filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.super filter=lfs diff=lfs merge=lfs -text +normal/transformer/dynamic/dd/replaced.onnx.data filter=lfs diff=lfs merge=lfs -text +*.ddbundle filter=lfs diff=lfs merge=lfs -text +*.onnx.data filter=lfs diff=lfs merge=lfs -text +*.json filter=lfs diff=lfs merge=lfs -text +*.state filter=lfs diff=lfs merge=lfs -text +inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text +inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text +normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst filter=lfs diff=lfs merge=lfs -text diff --git a/README.md b/README.md index 4847186227809518c48a88941f9fe51175f87969..5e5c9fd7636a24375f4e465e1c0b0ea5ea8737b7 100644 --- a/README.md +++ b/README.md @@ -1,64 +1,64 @@ ---- -license: other -license_name: stabilityai-ai-community -license_link: LICENSE -tags: -- text-to-image -- stable-diffusion -- diffusion-single-file -inference: false -extra_gated_prompt: >- - By clicking "Agree", you agree to the [License - Agreement](https://huggingface.co/stabilityai/stable-diffusion-3-medium-amdnpu/blob/main/LICENSE) - and acknowledge Stability AI's [Privacy - Policy](https://stability.ai/privacy-policy). -extra_gated_fields: - Name: text - Email: text - Country: country - Organization or Affiliation: text - Receive email updates and promotions on Stability AI products, services, and research?: - type: select - options: - - 'Yes' - - 'No' - I agree to the License Agreement and acknowledge Stability AI's Privacy Policy: checkbox -language: -- en -pipeline_tag: text-to-image ---- - -# 🚀 Stable Diffusion 3 Medium on AMD NPU - -"Stable Diffusion 3 Medium is a Multimodal Diffusion Transformer (MMDiT) text-to-image model that features greatly improved performance in image quality, typography, complex prompt understanding, and resource-efficiency." ([stabilityai/stable-diffusion-3-medium](https://huggingface.co/stabilityai/stable-diffusion-3-medium)) - -This version of the model has been optimized to run on AMD NPUs. - -For more information, refer to the original model card under [stabilityai/stable-diffusion-3-medium](https://huggingface.co/stabilityai/stable-diffusion-3-medium). - -## Model Details - -| Model Details | Description | -| ----------- | ----------- | -| Person or organization developing model | [AMD](https://hf.co/amd) | -| Original model authors | [Stability AI](https://huggingface.co/stabilityai/stable-diffusion-3-medium) | -| Model date | February 2026 | -| Model version | 1.7.1 | -| Model type | Diffusion-based text-to-image generation model | -| Information about training algorithms, parameters, fairness constraints or other applied approaches, and features | This is a model that can be used to generate and modify images based on text prompts. | -| License | [Stability AI Community](LICENSE) | -| Where to send questions or comments about the model | [Community Tab](https://hf.co/amd/stable-diffusion-3-medium-amdnpu/discussions) and [AMD Developer Community Discord](https://discord.gg/amd-dev) | - -## ⚡ Intended Use - -#### Getting Started - -To get started with this model, visit [github.com/amd/sd-sandbox](https://github.com/amd/sd-sandbox). - -## ⚓ Ethical Considerations - -AMD is committed to conducting our business in a fair, ethical and honest manner and in compliance with all applicable laws, rules and regulations. You can find out more at the [AMD Ethics and Compliance](https://www.amd.com/en/corporate/corporate-responsibility/ethics-and-compliance.html) page. - -## ⚠️ Caveats and Recommendations - +--- +license: other +license_name: stabilityai-ai-community +license_link: LICENSE +tags: +- text-to-image +- stable-diffusion +- diffusion-single-file +inference: false +extra_gated_prompt: >- + By clicking "Agree", you agree to the [License + Agreement](https://huggingface.co/stabilityai/stable-diffusion-3-medium-amdnpu/blob/main/LICENSE) + and acknowledge Stability AI's [Privacy + Policy](https://stability.ai/privacy-policy). +extra_gated_fields: + Name: text + Email: text + Country: country + Organization or Affiliation: text + Receive email updates and promotions on Stability AI products, services, and research?: + type: select + options: + - 'Yes' + - 'No' + I agree to the License Agreement and acknowledge Stability AI's Privacy Policy: checkbox +language: +- en +pipeline_tag: text-to-image +--- + +# 🚀 Stable Diffusion 3 Medium on AMD NPU + +"Stable Diffusion 3 Medium is a Multimodal Diffusion Transformer (MMDiT) text-to-image model that features greatly improved performance in image quality, typography, complex prompt understanding, and resource-efficiency." ([stabilityai/stable-diffusion-3-medium](https://huggingface.co/stabilityai/stable-diffusion-3-medium)) + +This version of the model has been optimized to run on AMD NPUs. + +For more information, refer to the original model card under [stabilityai/stable-diffusion-3-medium](https://huggingface.co/stabilityai/stable-diffusion-3-medium). + +## Model Details + +| Model Details | Description | +| ----------- | ----------- | +| Person or organization developing model | [AMD](https://hf.co/amd) | +| Original model authors | [Stability AI](https://huggingface.co/stabilityai/stable-diffusion-3-medium) | +| Model date | February 2026 | +| Model version | 1.8.0 | +| Model type | Diffusion-based text-to-image generation model | +| Information about training algorithms, parameters, fairness constraints or other applied approaches, and features | This is a model that can be used to generate and modify images based on text prompts. | +| License | [Stability AI Community](LICENSE) | +| Where to send questions or comments about the model | [Community Tab](https://hf.co/amd/stable-diffusion-3-medium-amdnpu/discussions) and [AMD Developer Community Discord](https://discord.gg/amd-dev) | + +## ⚡ Intended Use + +#### Getting Started + +To get started with this model, visit [github.com/amd/sd-sandbox](https://github.com/amd/sd-sandbox). + +## ⚓ Ethical Considerations + +AMD is committed to conducting our business in a fair, ethical and honest manner and in compliance with all applicable laws, rules and regulations. You can find out more at the [AMD Ethics and Compliance](https://www.amd.com/en/corporate/corporate-responsibility/ethics-and-compliance.html) page. + +## ⚠️ Caveats and Recommendations + Please visit the original model card for more details: [stabilityai/stable-diffusion-3-medium](https://huggingface.co/stabilityai/stable-diffusion-3-medium). \ No newline at end of file diff --git a/common/vae_decoder/dynamic/dd/cache/NhwcConv_0-decoderconv_inConv_meta.json b/common/vae_decoder/dynamic/dd/cache/NhwcConv_0-decoderconv_inConv_meta.json index f23837eea9b25b9c03f9bc4ca8eb27680d029146..0e8207295edaad308355729e1694c8ceb3b3906e 100644 --- a/common/vae_decoder/dynamic/dd/cache/NhwcConv_0-decoderconv_inConv_meta.json +++ b/common/vae_decoder/dynamic/dd/cache/NhwcConv_0-decoderconv_inConv_meta.json @@ -117,7 +117,7 @@ "GroupNorm_0_wts_10_0_0" ], "out_args": [ - "GroupNorm_0.out13_0.out10_0_0" + "GroupNorm_0.out14_0.out10_0_0" ], "attrs": { "epsilon": { @@ -211,7 +211,7 @@ "name": "NhwcConv_1-/decoder/mid_block/resnets.0/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_0.out13_0.out10_0_0" + "GroupNorm_0.out14_0.out10_0_0" ], "const_args": [ "NhwcConv_1_weight_NHWC" @@ -321,7 +321,7 @@ "GroupNorm_1_wts_10_0_1" ], "out_args": [ - "GroupNorm_1.out13_1.out10_0_1" + "GroupNorm_1.out14_1.out10_0_1" ], "attrs": { "epsilon": { @@ -415,7 +415,7 @@ "name": "NhwcConv_2-/decoder/mid_block/resnets.0/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_1.out13_1.out10_0_1" + "GroupNorm_1.out14_1.out10_0_1" ], "const_args": [ "NhwcConv_2_weight_NHWC" @@ -682,10 +682,10 @@ "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" ], "const_args": [ - "onnx::MatMul_914" + "onnx::MatMul_914_15_0_0" ], "out_args": [ - "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0" + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_5_0" ], "attrs": { "input_shape": { @@ -746,10 +746,10 @@ "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" ], "const_args": [ - "onnx::MatMul_913" + "onnx::MatMul_913_15_0_1" ], "out_args": [ - "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1" + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_5_1" ], "attrs": { "input_shape": { @@ -810,10 +810,10 @@ "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" ], "const_args": [ - "onnx::MatMul_915" + "onnx::MatMul_915_15_0_2" ], "out_args": [ - "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2" + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_5_2" ], "attrs": { "input_shape": { @@ -868,15 +868,15 @@ } }, { - "name": "/decoder/mid_block/attentions.0/MatMulmha_4_0_0", + "name": "/decoder/mid_block/attentions.0/MatMulmha_5_0_0", "type": "SDMHA_VAE", "in_args": [ - "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1", - "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0", - "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2" + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_5_1", + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_5_0", + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_5_2" ], "const_args": [ - "/decoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0" + "/decoder/mid_block/attentions.0/MatMulmha_5_0_0_mask.9_0" ], "out_args": [ "/decoder/mid_block/attentions.0/Reshape_4_output_0.out9_0" @@ -937,7 +937,7 @@ "/decoder/mid_block/attentions.0/Reshape_4_output_0.out9_0" ], "const_args": [ - "onnx::MatMul_923" + "onnx::MatMul_923_15_0_3" ], "out_args": [ "/decoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_1_0" @@ -1070,7 +1070,7 @@ "GroupNorm_3_wts_10_0_2" ], "out_args": [ - "GroupNorm_3.out13_3.out10_0_2" + "GroupNorm_3.out14_3.out10_0_2" ], "attrs": { "epsilon": { @@ -1164,7 +1164,7 @@ "name": "NhwcConv_3-/decoder/mid_block/resnets.1/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_3.out13_3.out10_0_2" + "GroupNorm_3.out14_3.out10_0_2" ], "const_args": [ "NhwcConv_3_weight_NHWC" @@ -1274,7 +1274,7 @@ "GroupNorm_4_wts_10_0_3" ], "out_args": [ - "GroupNorm_4.out13_4.out10_0_3" + "GroupNorm_4.out14_4.out10_0_3" ], "attrs": { "epsilon": { @@ -1368,7 +1368,7 @@ "name": "NhwcConv_4-/decoder/mid_block/resnets.1/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_4.out13_4.out10_0_3" + "GroupNorm_4.out14_4.out10_0_3" ], "const_args": [ "NhwcConv_4_weight_NHWC" @@ -1544,7 +1544,7 @@ "GroupNorm_5_wts_10_0_4" ], "out_args": [ - "GroupNorm_5.out13_5.out10_0_4" + "GroupNorm_5.out14_5.out10_0_4" ], "attrs": { "epsilon": { @@ -1638,7 +1638,7 @@ "name": "NhwcConv_5-/decoder/up_blocks.0/resnets.0/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_5.out13_5.out10_0_4" + "GroupNorm_5.out14_5.out10_0_4" ], "const_args": [ "NhwcConv_5_weight_NHWC" @@ -1748,7 +1748,7 @@ "GroupNorm_6_wts_10_0_5" ], "out_args": [ - "GroupNorm_6.out13_6.out10_0_5" + "GroupNorm_6.out14_6.out10_0_5" ], "attrs": { "epsilon": { @@ -1842,7 +1842,7 @@ "name": "NhwcConv_6-/decoder/up_blocks.0/resnets.0/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_6.out13_6.out10_0_5" + "GroupNorm_6.out14_6.out10_0_5" ], "const_args": [ "NhwcConv_6_weight_NHWC" @@ -2018,7 +2018,7 @@ "GroupNorm_7_wts_10_0_6" ], "out_args": [ - "GroupNorm_7.out13_7.out10_0_6" + "GroupNorm_7.out14_7.out10_0_6" ], "attrs": { "epsilon": { @@ -2112,7 +2112,7 @@ "name": "NhwcConv_7-/decoder/up_blocks.0/resnets.1/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_7.out13_7.out10_0_6" + "GroupNorm_7.out14_7.out10_0_6" ], "const_args": [ "NhwcConv_7_weight_NHWC" @@ -2222,7 +2222,7 @@ "GroupNorm_8_wts_10_0_7" ], "out_args": [ - "GroupNorm_8.out13_8.out10_0_7" + "GroupNorm_8.out14_8.out10_0_7" ], "attrs": { "epsilon": { @@ -2316,7 +2316,7 @@ "name": "NhwcConv_8-/decoder/up_blocks.0/resnets.1/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_8.out13_8.out10_0_7" + "GroupNorm_8.out14_8.out10_0_7" ], "const_args": [ "NhwcConv_8_weight_NHWC" @@ -2492,7 +2492,7 @@ "GroupNorm_9_wts_10_0_8" ], "out_args": [ - "GroupNorm_9.out13_9.out10_0_8" + "GroupNorm_9.out14_9.out10_0_8" ], "attrs": { "epsilon": { @@ -2586,7 +2586,7 @@ "name": "NhwcConv_9-/decoder/up_blocks.0/resnets.2/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_9.out13_9.out10_0_8" + "GroupNorm_9.out14_9.out10_0_8" ], "const_args": [ "NhwcConv_9_weight_NHWC" @@ -2696,7 +2696,7 @@ "GroupNorm_10_wts_10_0_9" ], "out_args": [ - "GroupNorm_10.out13_10.out10_0_9" + "GroupNorm_10.out14_10.out10_0_9" ], "attrs": { "epsilon": { @@ -2790,7 +2790,7 @@ "name": "NhwcConv_10-/decoder/up_blocks.0/resnets.2/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_10.out13_10.out10_0_9" + "GroupNorm_10.out14_10.out10_0_9" ], "const_args": [ "NhwcConv_10_weight_NHWC" @@ -2899,7 +2899,7 @@ ], "const_args": [], "out_args": [ - "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5" + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc7_0.out_7_1_5" ], "attrs": { "a_shape": { @@ -2960,13 +2960,13 @@ "name": "/decoder/up_blocks.0/upsamplers.0/Resize", "type": "SDResize", "in_args": [ - "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5" + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc7_0.out_7_1_5" ], "const_args": [ "/decoder/up_blocks.0/upsamplers.0/Resize.weights6_0" ], "out_args": [ - "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0" + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc7_0.out_6_0" ], "attrs": { "a_shape": { @@ -3011,7 +3011,7 @@ "name": "NhwcConv_11-/decoder/up_blocks.0/upsamplers.0/conv/Conv", "type": "SDConv", "in_args": [ - "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0" + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc7_0.out_6_0" ], "const_args": [ "NhwcConv_11_weight_NHWC" @@ -3121,7 +3121,7 @@ "GroupNorm_11_wts_10_0_10" ], "out_args": [ - "GroupNorm_11.out13_11.out10_0_10" + "GroupNorm_11.out14_11.out10_0_10" ], "attrs": { "epsilon": { @@ -3215,7 +3215,7 @@ "name": "NhwcConv_12-/decoder/up_blocks.1/resnets.0/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_11.out13_11.out10_0_10" + "GroupNorm_11.out14_11.out10_0_10" ], "const_args": [ "NhwcConv_12_weight_NHWC" @@ -3325,7 +3325,7 @@ "GroupNorm_12_wts_10_0_11" ], "out_args": [ - "GroupNorm_12.out13_12.out10_0_11" + "GroupNorm_12.out14_12.out10_0_11" ], "attrs": { "epsilon": { @@ -3419,7 +3419,7 @@ "name": "NhwcConv_13-/decoder/up_blocks.1/resnets.0/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_12.out13_12.out10_0_11" + "GroupNorm_12.out14_12.out10_0_11" ], "const_args": [ "NhwcConv_13_weight_NHWC" @@ -3595,7 +3595,7 @@ "GroupNorm_13_wts_10_0_12" ], "out_args": [ - "GroupNorm_13.out13_13.out10_0_12" + "GroupNorm_13.out14_13.out10_0_12" ], "attrs": { "epsilon": { @@ -3689,7 +3689,7 @@ "name": "NhwcConv_14-/decoder/up_blocks.1/resnets.1/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_13.out13_13.out10_0_12" + "GroupNorm_13.out14_13.out10_0_12" ], "const_args": [ "NhwcConv_14_weight_NHWC" @@ -3799,7 +3799,7 @@ "GroupNorm_14_wts_10_0_13" ], "out_args": [ - "GroupNorm_14.out13_14.out10_0_13" + "GroupNorm_14.out14_14.out10_0_13" ], "attrs": { "epsilon": { @@ -3893,7 +3893,7 @@ "name": "NhwcConv_15-/decoder/up_blocks.1/resnets.1/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_14.out13_14.out10_0_13" + "GroupNorm_14.out14_14.out10_0_13" ], "const_args": [ "NhwcConv_15_weight_NHWC" @@ -4069,7 +4069,7 @@ "GroupNorm_15_wts_10_0_14" ], "out_args": [ - "GroupNorm_15.out13_15.out10_0_14" + "GroupNorm_15.out14_15.out10_0_14" ], "attrs": { "epsilon": { @@ -4163,7 +4163,7 @@ "name": "NhwcConv_16-/decoder/up_blocks.1/resnets.2/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_15.out13_15.out10_0_14" + "GroupNorm_15.out14_15.out10_0_14" ], "const_args": [ "NhwcConv_16_weight_NHWC" @@ -4273,7 +4273,7 @@ "GroupNorm_16_wts_10_0_15" ], "out_args": [ - "GroupNorm_16.out13_16.out10_0_15" + "GroupNorm_16.out14_16.out10_0_15" ], "attrs": { "epsilon": { @@ -4367,7 +4367,7 @@ "name": "NhwcConv_17-/decoder/up_blocks.1/resnets.2/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_16.out13_16.out10_0_15" + "GroupNorm_16.out14_16.out10_0_15" ], "const_args": [ "NhwcConv_17_weight_NHWC" @@ -4543,7 +4543,7 @@ "/decoder/up_blocks.1/upsamplers.0/Resize.weights6_1" ], "out_args": [ - "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1" + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc7_1.out_6_1" ], "attrs": { "a_shape": { @@ -4588,7 +4588,7 @@ "name": "NhwcConv_18-/decoder/up_blocks.1/upsamplers.0/conv/Conv", "type": "SDConv", "in_args": [ - "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1" + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc7_1.out_6_1" ], "const_args": [ "NhwcConv_18_weight_NHWC" @@ -4698,7 +4698,7 @@ "GroupNorm_17_wts_10_0_16" ], "out_args": [ - "GroupNorm_17.out13_17.out10_0_16" + "GroupNorm_17.out14_17.out10_0_16" ], "attrs": { "epsilon": { @@ -4896,7 +4896,7 @@ "name": "NhwcConv_20-/decoder/up_blocks.2/resnets.0/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_17.out13_17.out10_0_16" + "GroupNorm_17.out14_17.out10_0_16" ], "const_args": [ "NhwcConv_20_weight_NHWC" @@ -5006,7 +5006,7 @@ "GroupNorm_18_wts_10_0_17" ], "out_args": [ - "GroupNorm_18.out13_18.out10_0_17" + "GroupNorm_18.out14_18.out10_0_17" ], "attrs": { "epsilon": { @@ -5100,7 +5100,7 @@ "name": "NhwcConv_21-/decoder/up_blocks.2/resnets.0/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_18.out13_18.out10_0_17" + "GroupNorm_18.out14_18.out10_0_17" ], "const_args": [ "NhwcConv_21_weight_NHWC" @@ -5276,7 +5276,7 @@ "GroupNorm_19_wts_10_0_18" ], "out_args": [ - "GroupNorm_19.out13_19.out10_0_18" + "GroupNorm_19.out14_19.out10_0_18" ], "attrs": { "epsilon": { @@ -5370,7 +5370,7 @@ "name": "NhwcConv_22-/decoder/up_blocks.2/resnets.1/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_19.out13_19.out10_0_18" + "GroupNorm_19.out14_19.out10_0_18" ], "const_args": [ "NhwcConv_22_weight_NHWC" @@ -5480,7 +5480,7 @@ "GroupNorm_20_wts_10_0_19" ], "out_args": [ - "GroupNorm_20.out13_20.out10_0_19" + "GroupNorm_20.out14_20.out10_0_19" ], "attrs": { "epsilon": { @@ -5574,7 +5574,7 @@ "name": "NhwcConv_23-/decoder/up_blocks.2/resnets.1/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_20.out13_20.out10_0_19" + "GroupNorm_20.out14_20.out10_0_19" ], "const_args": [ "NhwcConv_23_weight_NHWC" @@ -5750,7 +5750,7 @@ "GroupNorm_21_wts_10_0_20" ], "out_args": [ - "GroupNorm_21.out13_21.out10_0_20" + "GroupNorm_21.out14_21.out10_0_20" ], "attrs": { "epsilon": { @@ -5844,7 +5844,7 @@ "name": "NhwcConv_24-/decoder/up_blocks.2/resnets.2/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_21.out13_21.out10_0_20" + "GroupNorm_21.out14_21.out10_0_20" ], "const_args": [ "NhwcConv_24_weight_NHWC" @@ -5954,7 +5954,7 @@ "GroupNorm_22_wts_10_0_21" ], "out_args": [ - "GroupNorm_22.out13_22.out10_0_21" + "GroupNorm_22.out14_22.out10_0_21" ], "attrs": { "epsilon": { @@ -6048,7 +6048,7 @@ "name": "NhwcConv_25-/decoder/up_blocks.2/resnets.2/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_22.out13_22.out10_0_21" + "GroupNorm_22.out14_22.out10_0_21" ], "const_args": [ "NhwcConv_25_weight_NHWC" @@ -6224,7 +6224,7 @@ "/decoder/up_blocks.2/upsamplers.0/Resize.weights6_2" ], "out_args": [ - "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2" + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc7_2.out_6_2" ], "attrs": { "a_shape": { @@ -6269,7 +6269,7 @@ "name": "NhwcConv_26-/decoder/up_blocks.2/upsamplers.0/conv/Conv", "type": "SDConv", "in_args": [ - "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2" + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc7_2.out_6_2" ], "const_args": [ "NhwcConv_26_weight_NHWC" @@ -6379,7 +6379,7 @@ "GroupNorm_23_wts_10_0_22" ], "out_args": [ - "GroupNorm_23.out13_23.out10_0_22" + "GroupNorm_23.out14_23.out10_0_22" ], "attrs": { "epsilon": { @@ -6577,7 +6577,7 @@ "name": "NhwcConv_28-/decoder/up_blocks.3/resnets.0/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_23.out13_23.out10_0_22" + "GroupNorm_23.out14_23.out10_0_22" ], "const_args": [ "NhwcConv_28_weight_NHWC" @@ -6687,7 +6687,7 @@ "GroupNorm_24_wts_10_0_23" ], "out_args": [ - "GroupNorm_24.out13_24.out10_0_23" + "GroupNorm_24.out14_24.out10_0_23" ], "attrs": { "epsilon": { @@ -6781,7 +6781,7 @@ "name": "NhwcConv_29-/decoder/up_blocks.3/resnets.0/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_24.out13_24.out10_0_23" + "GroupNorm_24.out14_24.out10_0_23" ], "const_args": [ "NhwcConv_29_weight_NHWC" @@ -6957,7 +6957,7 @@ "GroupNorm_25_wts_10_0_24" ], "out_args": [ - "GroupNorm_25.out13_25.out10_0_24" + "GroupNorm_25.out14_25.out10_0_24" ], "attrs": { "epsilon": { @@ -7051,7 +7051,7 @@ "name": "NhwcConv_30-/decoder/up_blocks.3/resnets.1/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_25.out13_25.out10_0_24" + "GroupNorm_25.out14_25.out10_0_24" ], "const_args": [ "NhwcConv_30_weight_NHWC" @@ -7161,7 +7161,7 @@ "GroupNorm_26_wts_10_0_25" ], "out_args": [ - "GroupNorm_26.out13_26.out10_0_25" + "GroupNorm_26.out14_26.out10_0_25" ], "attrs": { "epsilon": { @@ -7255,7 +7255,7 @@ "name": "NhwcConv_31-/decoder/up_blocks.3/resnets.1/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_26.out13_26.out10_0_25" + "GroupNorm_26.out14_26.out10_0_25" ], "const_args": [ "NhwcConv_31_weight_NHWC" @@ -7431,7 +7431,7 @@ "GroupNorm_27_wts_10_0_26" ], "out_args": [ - "GroupNorm_27.out13_27.out10_0_26" + "GroupNorm_27.out14_27.out10_0_26" ], "attrs": { "epsilon": { @@ -7525,7 +7525,7 @@ "name": "NhwcConv_32-/decoder/up_blocks.3/resnets.2/conv1/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_27.out13_27.out10_0_26" + "GroupNorm_27.out14_27.out10_0_26" ], "const_args": [ "NhwcConv_32_weight_NHWC" @@ -7635,7 +7635,7 @@ "GroupNorm_28_wts_10_0_27" ], "out_args": [ - "GroupNorm_28.out13_28.out10_0_27" + "GroupNorm_28.out14_28.out10_0_27" ], "attrs": { "epsilon": { @@ -7729,7 +7729,7 @@ "name": "NhwcConv_33-/decoder/up_blocks.3/resnets.2/conv2/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_28.out13_28.out10_0_27" + "GroupNorm_28.out14_28.out10_0_27" ], "const_args": [ "NhwcConv_33_weight_NHWC" @@ -7905,7 +7905,7 @@ "GroupNorm_29_wts_10_0_28" ], "out_args": [ - "GroupNorm_29.out13_29.out10_0_28" + "GroupNorm_29.out14_29.out10_0_28" ], "attrs": { "epsilon": { @@ -7999,7 +7999,7 @@ "name": "NhwcConv_34-/decoder/conv_out/Conv", "type": "SDConv", "in_args": [ - "GroupNorm_29.out13_29.out10_0_28" + "GroupNorm_29.out14_29.out10_0_28" ], "const_args": [ "NhwcConv_34_weight_NHWC" @@ -8120,92 +8120,92 @@ "xrt_arg_id": 2, "packed_tensors": [ "NhwcConv_0_out-/decoder/conv_in/Conv_output_0.out2_3_0", - "GroupNorm_0.out13_0.out10_0_0", + "GroupNorm_0.out14_0.out10_0_0", "NhwcConv_1_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_1", - "GroupNorm_1.out13_1.out10_0_1", + "GroupNorm_1.out14_1.out10_0_1", "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_2", "/decoder/mid_block/resnets.0/Add.out_7_1_0", "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0", - "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0", - "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1", - "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2", + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_5_0", + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_5_1", + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_5_2", "/decoder/mid_block/attentions.0/Reshape_4_output_0.out9_0", "/decoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_1_0", "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_7_1_1", - "GroupNorm_3.out13_3.out10_0_2", + "GroupNorm_3.out14_3.out10_0_2", "NhwcConv_3_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_3", - "GroupNorm_4.out13_4.out10_0_3", + "GroupNorm_4.out14_4.out10_0_3", "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_4", "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_7_1_2", - "GroupNorm_5.out13_5.out10_0_4", + "GroupNorm_5.out14_5.out10_0_4", "NhwcConv_5_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_5", - "GroupNorm_6.out13_6.out10_0_5", + "GroupNorm_6.out14_6.out10_0_5", "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_6", "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_7_1_3", - "GroupNorm_7.out13_7.out10_0_6", + "GroupNorm_7.out14_7.out10_0_6", "NhwcConv_7_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_7", - "GroupNorm_8.out13_8.out10_0_7", + "GroupNorm_8.out14_8.out10_0_7", "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_8", "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_7_1_4", - "GroupNorm_9.out13_9.out10_0_8", + "GroupNorm_9.out14_9.out10_0_8", "NhwcConv_9_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out2_3_9", - "GroupNorm_10.out13_10.out10_0_9", + "GroupNorm_10.out14_10.out10_0_9", "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out2_3_10", - "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5", - "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0", + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc7_0.out_7_1_5", + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc7_0.out_6_0", "NhwcConv_11_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out2_3_11", - "GroupNorm_11.out13_11.out10_0_10", + "GroupNorm_11.out14_11.out10_0_10", "NhwcConv_12_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_12", - "GroupNorm_12.out13_12.out10_0_11", + "GroupNorm_12.out14_12.out10_0_11", "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_13", "/decoder/up_blocks.1/resnets.0/Add.out_7_1_6", - "GroupNorm_13.out13_13.out10_0_12", + "GroupNorm_13.out14_13.out10_0_12", "NhwcConv_14_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_14", - "GroupNorm_14.out13_14.out10_0_13", + "GroupNorm_14.out14_14.out10_0_13", "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_15", "/decoder/up_blocks.1/resnets.1/Add.out_7_1_7", - "GroupNorm_15.out13_15.out10_0_14", + "GroupNorm_15.out14_15.out10_0_14", "NhwcConv_16_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out2_3_16", - "GroupNorm_16.out13_16.out10_0_15", + "GroupNorm_16.out14_16.out10_0_15", "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out2_3_17", "/decoder/up_blocks.1/resnets.2/Add.out_7_1_8", - "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1", + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc7_1.out_6_1", "NhwcConv_18_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out2_3_18", - "GroupNorm_17.out13_17.out10_0_16", + "GroupNorm_17.out14_17.out10_0_16", "NhwcConv_19_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_19", "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_20", - "GroupNorm_18.out13_18.out10_0_17", + "GroupNorm_18.out14_18.out10_0_17", "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_21", "/decoder/up_blocks.2/resnets.0/Add.out_7_1_9", - "GroupNorm_19.out13_19.out10_0_18", + "GroupNorm_19.out14_19.out10_0_18", "NhwcConv_22_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_22", - "GroupNorm_20.out13_20.out10_0_19", + "GroupNorm_20.out14_20.out10_0_19", "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_23", "/decoder/up_blocks.2/resnets.1/Add.out_7_1_10", - "GroupNorm_21.out13_21.out10_0_20", + "GroupNorm_21.out14_21.out10_0_20", "NhwcConv_24_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out2_3_24", - "GroupNorm_22.out13_22.out10_0_21", + "GroupNorm_22.out14_22.out10_0_21", "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out2_3_25", "/decoder/up_blocks.2/resnets.2/Add.out_7_1_11", - "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2", + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc7_2.out_6_2", "NhwcConv_26_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out2_3_26", - "GroupNorm_23.out13_23.out10_0_22", + "GroupNorm_23.out14_23.out10_0_22", "NhwcConv_27_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out2_3_27", "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_28", - "GroupNorm_24.out13_24.out10_0_23", + "GroupNorm_24.out14_24.out10_0_23", "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_29", "/decoder/up_blocks.3/resnets.0/Add.out_7_1_12", - "GroupNorm_25.out13_25.out10_0_24", + "GroupNorm_25.out14_25.out10_0_24", "NhwcConv_30_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_30", - "GroupNorm_26.out13_26.out10_0_25", + "GroupNorm_26.out14_26.out10_0_25", "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_31", "/decoder/up_blocks.3/resnets.1/Add.out_7_1_13", - "GroupNorm_27.out13_27.out10_0_26", + "GroupNorm_27.out14_27.out10_0_26", "NhwcConv_32_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out2_3_32", - "GroupNorm_28.out13_28.out10_0_27", + "GroupNorm_28.out14_28.out10_0_27", "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out2_3_33", "/decoder/up_blocks.3/resnets.2/Add.out_7_1_14", - "GroupNorm_29.out13_29.out10_0_28" + "GroupNorm_29.out14_29.out10_0_28" ] }, "const": { @@ -8218,11 +8218,11 @@ "GroupNorm_1_wts_10_0_1", "NhwcConv_2_weight_NHWC", "GroupNorm_2_wts_10_1_0", - "onnx::MatMul_914", - "onnx::MatMul_913", - "onnx::MatMul_915", - "/decoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0", - "onnx::MatMul_923", + "onnx::MatMul_914_15_0_0", + "onnx::MatMul_913_15_0_1", + "onnx::MatMul_915_15_0_2", + "/decoder/mid_block/attentions.0/MatMulmha_5_0_0_mask.9_0", + "onnx::MatMul_923_15_0_3", "GroupNorm_3_wts_10_0_2", "NhwcConv_3_weight_NHWC", "GroupNorm_4_wts_10_0_3", @@ -8354,7 +8354,7 @@ ], "offset": 0 }, - "GroupNorm_0.out13_0.out10_0_0": { + "GroupNorm_0.out14_0.out10_0_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8394,7 +8394,7 @@ ], "offset": 2048 }, - "GroupNorm_1.out13_1.out10_0_1": { + "GroupNorm_1.out14_1.out10_0_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8472,7 +8472,7 @@ ], "offset": 6144 }, - "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0": { + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_5_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8490,7 +8490,7 @@ ], "offset": 7168 }, - "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1": { + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_5_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8508,7 +8508,7 @@ ], "offset": 8192 }, - "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2": { + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_5_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8584,7 +8584,7 @@ ], "offset": 12288 }, - "GroupNorm_3.out13_3.out10_0_2": { + "GroupNorm_3.out14_3.out10_0_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8624,7 +8624,7 @@ ], "offset": 14336 }, - "GroupNorm_4.out13_4.out10_0_3": { + "GroupNorm_4.out14_4.out10_0_3": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8684,7 +8684,7 @@ ], "offset": 17408 }, - "GroupNorm_5.out13_5.out10_0_4": { + "GroupNorm_5.out14_5.out10_0_4": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8724,7 +8724,7 @@ ], "offset": 19456 }, - "GroupNorm_6.out13_6.out10_0_5": { + "GroupNorm_6.out14_6.out10_0_5": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8784,7 +8784,7 @@ ], "offset": 22528 }, - "GroupNorm_7.out13_7.out10_0_6": { + "GroupNorm_7.out14_7.out10_0_6": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8824,7 +8824,7 @@ ], "offset": 24576 }, - "GroupNorm_8.out13_8.out10_0_7": { + "GroupNorm_8.out14_8.out10_0_7": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8884,7 +8884,7 @@ ], "offset": 27648 }, - "GroupNorm_9.out13_9.out10_0_8": { + "GroupNorm_9.out14_9.out10_0_8": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8924,7 +8924,7 @@ ], "offset": 29696 }, - "GroupNorm_10.out13_10.out10_0_9": { + "GroupNorm_10.out14_10.out10_0_9": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8964,7 +8964,7 @@ ], "offset": 31744 }, - "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5": { + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc7_0.out_7_1_5": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8984,7 +8984,7 @@ ], "offset": 32768 }, - "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0": { + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc7_0.out_6_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9024,7 +9024,7 @@ ], "offset": 34816 }, - "GroupNorm_11.out13_11.out10_0_10": { + "GroupNorm_11.out14_11.out10_0_10": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9064,7 +9064,7 @@ ], "offset": 36864 }, - "GroupNorm_12.out13_12.out10_0_11": { + "GroupNorm_12.out14_12.out10_0_11": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9124,7 +9124,7 @@ ], "offset": 39936 }, - "GroupNorm_13.out13_13.out10_0_12": { + "GroupNorm_13.out14_13.out10_0_12": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9164,7 +9164,7 @@ ], "offset": 41984 }, - "GroupNorm_14.out13_14.out10_0_13": { + "GroupNorm_14.out14_14.out10_0_13": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9224,7 +9224,7 @@ ], "offset": 45056 }, - "GroupNorm_15.out13_15.out10_0_14": { + "GroupNorm_15.out14_15.out10_0_14": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9264,7 +9264,7 @@ ], "offset": 47104 }, - "GroupNorm_16.out13_16.out10_0_15": { + "GroupNorm_16.out14_16.out10_0_15": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9324,7 +9324,7 @@ ], "offset": 50176 }, - "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1": { + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc7_1.out_6_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9364,7 +9364,7 @@ ], "offset": 52224 }, - "GroupNorm_17.out13_17.out10_0_16": { + "GroupNorm_17.out14_17.out10_0_16": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9424,7 +9424,7 @@ ], "offset": 54784 }, - "GroupNorm_18.out13_18.out10_0_17": { + "GroupNorm_18.out14_18.out10_0_17": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9484,7 +9484,7 @@ ], "offset": 56320 }, - "GroupNorm_19.out13_19.out10_0_18": { + "GroupNorm_19.out14_19.out10_0_18": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9524,7 +9524,7 @@ ], "offset": 57344 }, - "GroupNorm_20.out13_20.out10_0_19": { + "GroupNorm_20.out14_20.out10_0_19": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9584,7 +9584,7 @@ ], "offset": 58880 }, - "GroupNorm_21.out13_21.out10_0_20": { + "GroupNorm_21.out14_21.out10_0_20": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9624,7 +9624,7 @@ ], "offset": 59904 }, - "GroupNorm_22.out13_22.out10_0_21": { + "GroupNorm_22.out14_22.out10_0_21": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9684,7 +9684,7 @@ ], "offset": 61440 }, - "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2": { + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc7_2.out_6_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9724,7 +9724,7 @@ ], "offset": 62464 }, - "GroupNorm_23.out13_23.out10_0_22": { + "GroupNorm_23.out14_23.out10_0_22": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9784,7 +9784,7 @@ ], "offset": 63744 }, - "GroupNorm_24.out13_24.out10_0_23": { + "GroupNorm_24.out14_24.out10_0_23": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9844,7 +9844,7 @@ ], "offset": 64512 }, - "GroupNorm_25.out13_25.out10_0_24": { + "GroupNorm_25.out14_25.out10_0_24": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9884,7 +9884,7 @@ ], "offset": 65024 }, - "GroupNorm_26.out13_26.out10_0_25": { + "GroupNorm_26.out14_26.out10_0_25": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9944,7 +9944,7 @@ ], "offset": 65792 }, - "GroupNorm_27.out13_27.out10_0_26": { + "GroupNorm_27.out14_27.out10_0_26": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -9984,7 +9984,7 @@ ], "offset": 66304 }, - "GroupNorm_28.out13_28.out10_0_27": { + "GroupNorm_28.out14_28.out10_0_27": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -10044,7 +10044,7 @@ ], "offset": 67072 }, - "GroupNorm_29.out13_29.out10_0_28": { + "GroupNorm_29.out14_29.out10_0_28": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -10142,7 +10142,7 @@ "file_name": "cache/NhwcConv_0-decoderconv_inConv_5.const", "file_size": 2048 }, - "onnx::MatMul_914": { + "onnx::MatMul_914_15_0_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -10155,7 +10155,7 @@ "file_name": "cache/NhwcConv_0-decoderconv_inConv_6.const", "file_size": 299008 }, - "onnx::MatMul_913": { + "onnx::MatMul_913_15_0_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -10168,7 +10168,7 @@ "file_name": "cache/NhwcConv_0-decoderconv_inConv_7.const", "file_size": 299008 }, - "onnx::MatMul_915": { + "onnx::MatMul_915_15_0_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -10181,7 +10181,7 @@ "file_name": "cache/NhwcConv_0-decoderconv_inConv_8.const", "file_size": 299008 }, - "/decoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0": { + "/decoder/mid_block/attentions.0/MatMulmha_5_0_0_mask.9_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -10194,7 +10194,7 @@ "file_name": "cache/NhwcConv_0-decoderconv_inConv_9.const", "file_size": 32768 }, - "onnx::MatMul_923": { + "onnx::MatMul_923_15_0_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -11017,92 +11017,92 @@ "dynamic_shape_subgraph": true, "dynamic_shape_list": [ { - "floor(2.0*floor(2.0*floor(2.0*width)))": 512, + "floor(2.0*floor(2.0*height))": 256, + "floor(2.0*floor(2.0*floor(2.0*height)))": 512, "width": 64, - "floor(2.0*width)": 128, + "height": 64, "floor(2.0*height)": 128, - "floor(2.0*floor(2.0*floor(2.0*height)))": 512, - "floor(2.0*floor(2.0*height))": 256, + "floor(2.0*floor(2.0*floor(2.0*width)))": 512, "height*width": 4096, - "height": 64, - "floor(2.0*floor(2.0*width))": 256 + "floor(2.0*floor(2.0*width))": 256, + "floor(2.0*width)": 128 }, { - "floor(2.0*floor(2.0*floor(2.0*width)))": 768, + "floor(2.0*floor(2.0*height))": 256, + "floor(2.0*floor(2.0*floor(2.0*height)))": 512, "width": 96, - "floor(2.0*width)": 192, + "height": 64, "floor(2.0*height)": 128, - "floor(2.0*floor(2.0*floor(2.0*height)))": 512, - "floor(2.0*floor(2.0*height))": 256, + "floor(2.0*floor(2.0*floor(2.0*width)))": 768, "height*width": 6144, - "height": 64, - "floor(2.0*floor(2.0*width))": 384 + "floor(2.0*floor(2.0*width))": 384, + "floor(2.0*width)": 192 }, { - "floor(2.0*floor(2.0*floor(2.0*width)))": 512, + "floor(2.0*floor(2.0*height))": 384, + "floor(2.0*floor(2.0*floor(2.0*height)))": 768, "width": 64, - "floor(2.0*width)": 128, + "height": 96, "floor(2.0*height)": 192, - "floor(2.0*floor(2.0*floor(2.0*height)))": 768, - "floor(2.0*floor(2.0*height))": 384, + "floor(2.0*floor(2.0*floor(2.0*width)))": 512, "height*width": 6144, - "height": 96, - "floor(2.0*floor(2.0*width))": 256 + "floor(2.0*floor(2.0*width))": 256, + "floor(2.0*width)": 128 }, { - "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, + "floor(2.0*floor(2.0*height))": 288, + "floor(2.0*floor(2.0*floor(2.0*height)))": 576, "width": 128, - "floor(2.0*width)": 256, + "height": 72, "floor(2.0*height)": 144, - "floor(2.0*floor(2.0*floor(2.0*height)))": 576, - "floor(2.0*floor(2.0*height))": 288, + "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, "height*width": 9216, - "height": 72, - "floor(2.0*floor(2.0*width))": 512 + "floor(2.0*floor(2.0*width))": 512, + "floor(2.0*width)": 256 }, { - "floor(2.0*floor(2.0*floor(2.0*width)))": 576, + "floor(2.0*floor(2.0*height))": 512, + "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, "width": 72, - "floor(2.0*width)": 144, + "height": 128, "floor(2.0*height)": 256, - "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, - "floor(2.0*floor(2.0*height))": 512, + "floor(2.0*floor(2.0*floor(2.0*width)))": 576, "height*width": 9216, - "height": 128, - "floor(2.0*floor(2.0*width))": 288 + "floor(2.0*floor(2.0*width))": 288, + "floor(2.0*width)": 144 }, { - "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, + "floor(2.0*floor(2.0*height))": 384, + "floor(2.0*floor(2.0*floor(2.0*height)))": 768, "width": 128, - "floor(2.0*width)": 256, + "height": 96, "floor(2.0*height)": 192, - "floor(2.0*floor(2.0*floor(2.0*height)))": 768, - "floor(2.0*floor(2.0*height))": 384, + "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, "height*width": 12288, - "height": 96, - "floor(2.0*floor(2.0*width))": 512 + "floor(2.0*floor(2.0*width))": 512, + "floor(2.0*width)": 256 }, { - "floor(2.0*floor(2.0*floor(2.0*width)))": 768, + "floor(2.0*floor(2.0*height))": 512, + "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, "width": 96, - "floor(2.0*width)": 192, + "height": 128, "floor(2.0*height)": 256, - "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, - "floor(2.0*floor(2.0*height))": 512, + "floor(2.0*floor(2.0*floor(2.0*width)))": 768, "height*width": 12288, - "height": 128, - "floor(2.0*floor(2.0*width))": 384 + "floor(2.0*floor(2.0*width))": 384, + "floor(2.0*width)": 192 }, { - "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, + "floor(2.0*floor(2.0*height))": 512, + "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, "width": 128, - "floor(2.0*width)": 256, + "height": 128, "floor(2.0*height)": 256, - "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, - "floor(2.0*floor(2.0*height))": 512, + "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, "height*width": 16384, - "height": 128, - "floor(2.0*floor(2.0*width))": 512 + "floor(2.0*floor(2.0*width))": 512, + "floor(2.0*width)": 256 } ], "aux_info": {} diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.ddbundle b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..b1200c5b5740c2728fbc4f3de3dea0b9fc3a78c8 --- /dev/null +++ b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59e8c0c060ae3ab3ef94077e04db792e56adb804cab254f30299e0a04882ca51 +size 377566197 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.state b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.state index 6dbac8403172f5629c564c1015f052d1140328da..61c2ae4d7edc105cbaec8dbf4565d7de74091b95 100644 Binary files a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.state and b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv.state differ diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.ctrlpkt deleted file mode 100644 index d22fb537038c685bb747d8bd23ee85ef6bbf907c..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:65afa687a04f11864b89f592785298c0e6335f92b3854582b27ab8feeaa12274 -size 58483440 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.meta deleted file mode 100644 index df4b1238f7db717c88d8bb3c519ab65fb46120f2..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:9597fd073851ca4961387908be0dfc460c6c0ae0b7664a047cf1f60f7673103d -size 13673069 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_128_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.ctrlpkt deleted file mode 100644 index b0b24a0a52403bde9ac63cc6387bcfcb545c4642..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:8f5bff4583f7db3f029ddb1e451c25b0d6f5ddef8622f3d7ddb04779a2d19468 -size 45850032 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.meta deleted file mode 100644 index 4fc63beb9760479bc90440358d78a88f27d266db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:82686265c36f9a00550fd0e28e068a0f34fc6050787e3f8ed79b7ceb040fa08c -size 9569730 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_72_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.ctrlpkt deleted file mode 100644 index 2c74967049e40740153456f6c877e4d950ca18a5..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:5f282ccf9f7a024f541c70dcfaea1bb17afa4881422ae36a2b372cf4eb26a8e5 -size 47764560 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.meta deleted file mode 100644 index e1a42bc6f6dd2a7b8090eb17f7aef922f358721d..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:1e82785b9898780484774f2a141f6cb2396872e7e188e0318fc36cac868a6404 -size 8969246 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_128_width_96_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.ctrlpkt deleted file mode 100644 index 90ff4721e4fc7c97f57e536823822b416b94a737..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:4dd386a138bbb93b12f7de49e8b6756e0f8a8fa3b1471e8859a5345ae401f48c -size 12569520 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.meta deleted file mode 100644 index 612d17311eef3ae74a4c59d1ba75f281c5be23ac..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:f0a5be3184699c663bc9be42c23ce6d85f7c23f9416f704c506e2e1aac419847 -size 3362240 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_64_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.ctrlpkt deleted file mode 100644 index 93a049f73feb42a33828e0269691230275804e32..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:d5939fa6d717e49df5c5b2a33648e6a5df6f871b25a67bfbf25b828e7286cb2a -size 25744368 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.meta deleted file mode 100644 index 180998742cdd1ccd48b43ed42c43fba94f89b1ae..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c53bd0be0e25311721a1b1c1b58a513959679872c2f7acc85d0df1a343621153 -size 4873573 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_64_width_96_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.ctrlpkt deleted file mode 100644 index 1c403b2bbd6b0dab15ffcf9578dcd41c0690106f..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:5e13e2500b0878bf27b0e31ed730a488f544a5741864cb6db50d5bb09077dbb8 -size 41365584 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.meta deleted file mode 100644 index c65b1fdcb18eb826b248e9a85388e7939bf1312a..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c6387cd0ff14ef3ca6ee98425d6db98faee69c65c85624901a2e467bf78ad302 -size 8466191 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_72_width_128_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.ctrlpkt deleted file mode 100644 index 617f303a3950c04de8a0c5e9af606f321af0326f..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:7504615da8ac154e87add3ed21abd0ed9b8521fe37c8ddf690acf9cb3c43ed1b -size 52493808 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.meta deleted file mode 100644 index cbd5d117dc4708c21a968a9ca665156645377f9e..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:60da5b2002b51f1ef6341773907cea5c8ed0271bb0e2f498b453aaf208731ef6 -size 9800829 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_128_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.ctrlpkt b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.ctrlpkt deleted file mode 100644 index 5b1b8563fdcdf6bd4e5b53c31de578bebb272c2e..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:5a6b6b34b4652dee9d31423b893e8645973ecb1014eb76f8beff2637f849726d -size 24780288 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.meta b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.meta deleted file mode 100644 index 31e119d7c5fef6a2b9657883d5e1ebd817996dbe..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ef2f2779e632338f4cada20a4c6718d965f8b1a04d824b631d539a1b7fa080ab -size 4619529 diff --git a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.super b/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_decoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-decoderconv_inConv_height_96_width_64_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_decoder/dynamic/dd/replaced.onnx b/common/vae_decoder/dynamic/dd/replaced.onnx index d836819bc7e161ec972c14c662f86465b2795afe..ce914394a26c0f5391561d9141f4445e8b730b1b 100644 --- a/common/vae_decoder/dynamic/dd/replaced.onnx +++ b/common/vae_decoder/dynamic/dd/replaced.onnx @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:109cffc82601519c46201a5091ff55ef3a1615990f31f01736e969ef222eca85 +oid sha256:45866024848f6f84a2e53194e36a7a0f0b39d618943017b78dad9093ce757ddc size 48621 diff --git a/common/vae_encoder/dynamic/dd/cache/NhwcConv_0-encoderconv_inConv_meta.json b/common/vae_encoder/dynamic/dd/cache/NhwcConv_0-encoderconv_inConv_meta.json index 7f311472150d76ea88eb544817b6db1625304074..12d5ce28736ec619396605a89f265bd7ac781c82 100644 --- a/common/vae_encoder/dynamic/dd/cache/NhwcConv_0-encoderconv_inConv_meta.json +++ b/common/vae_encoder/dynamic/dd/cache/NhwcConv_0-encoderconv_inConv_meta.json @@ -4994,10 +4994,10 @@ "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" ], "const_args": [ - "onnx::MatMul_875" + "onnx::MatMul_875_15_0_0" ], "out_args": [ - "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_3_0" + "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0" ], "attrs": { "input_shape": { @@ -5058,10 +5058,10 @@ "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" ], "const_args": [ - "onnx::MatMul_874" + "onnx::MatMul_874_15_0_1" ], "out_args": [ - "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_3_1" + "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1" ], "attrs": { "input_shape": { @@ -5122,10 +5122,10 @@ "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" ], "const_args": [ - "onnx::MatMul_876" + "onnx::MatMul_876_15_0_2" ], "out_args": [ - "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_3_2" + "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2" ], "attrs": { "input_shape": { @@ -5183,9 +5183,9 @@ "name": "/encoder/mid_block/attentions.0/MatMulmha_4_0_0", "type": "SDMHA_VAE", "in_args": [ - "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_3_1", - "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_3_0", - "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_3_2" + "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1", + "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0", + "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2" ], "const_args": [ "/encoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0" @@ -5249,7 +5249,7 @@ "/encoder/mid_block/attentions.0/Reshape_4_output_0.out9_0" ], "const_args": [ - "onnx::MatMul_884" + "onnx::MatMul_884_15_0_3" ], "out_args": [ "/encoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_0_0" @@ -6122,9 +6122,9 @@ "NhwcConv_23_out-/encoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_23", "/encoder/mid_block/resnets.0/Add.out_7_1_8", "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0", - "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_3_0", - "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_3_1", - "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_3_2", + "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0", + "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1", + "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2", "/encoder/mid_block/attentions.0/Reshape_4_output_0.out9_0", "/encoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_0_0", "/encoder/mid_block/attentions.0/Add_output_0_NHWC.out_7_1_9", @@ -6183,11 +6183,11 @@ "GroupNorm_17_wts_10_0_17", "NhwcConv_23_weight_NHWC", "GroupNorm_18_wts_10_1_0", - "onnx::MatMul_875", - "onnx::MatMul_874", - "onnx::MatMul_876", + "onnx::MatMul_875_15_0_0", + "onnx::MatMul_874_15_0_1", + "onnx::MatMul_876_15_0_2", "/encoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0", - "onnx::MatMul_884", + "onnx::MatMul_884_15_0_3", "GroupNorm_19_wts_10_0_18", "NhwcConv_24_weight_NHWC", "GroupNorm_20_wts_10_0_19", @@ -7281,7 +7281,7 @@ ], "offset": 36096 }, - "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_3_0": { + "/encoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -7299,7 +7299,7 @@ ], "offset": 37120 }, - "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_3_1": { + "/encoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -7317,7 +7317,7 @@ ], "offset": 38144 }, - "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_3_2": { + "/encoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -8072,7 +8072,7 @@ "file_name": "cache/NhwcConv_0-encoderconv_inConv_42.const", "file_size": 2048 }, - "onnx::MatMul_875": { + "onnx::MatMul_875_15_0_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -8085,7 +8085,7 @@ "file_name": "cache/NhwcConv_0-encoderconv_inConv_43.const", "file_size": 299008 }, - "onnx::MatMul_874": { + "onnx::MatMul_874_15_0_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -8098,7 +8098,7 @@ "file_name": "cache/NhwcConv_0-encoderconv_inConv_44.const", "file_size": 299008 }, - "onnx::MatMul_876": { + "onnx::MatMul_876_15_0_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -8124,7 +8124,7 @@ "file_name": "cache/NhwcConv_0-encoderconv_inConv_46.const", "file_size": 32768 }, - "onnx::MatMul_884": { + "onnx::MatMul_884_15_0_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -8219,92 +8219,92 @@ "dynamic_shape_subgraph": true, "dynamic_shape_list": [ { - "floor(floor(floor(width/2)/2)/2)": 64, "floor(floor(height/2)/2)": 128, - "floor(height/2)": 256, - "width": 512, - "floor(width/2)": 256, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 4096, "floor(floor(floor(height/2)/2)/2)": 64, - "floor(floor(width/2)/2)": 128, - "height": 512 + "width": 512, + "height": 512, + "floor(height/2)": 256, + "floor(floor(floor(width/2)/2)/2)": 64, + "floor(width/2)": 256, + "floor(floor(width/2)/2)": 128 }, { - "floor(floor(floor(width/2)/2)/2)": 96, "floor(floor(height/2)/2)": 128, - "floor(height/2)": 256, - "width": 768, - "floor(width/2)": 384, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 6144, "floor(floor(floor(height/2)/2)/2)": 64, - "floor(floor(width/2)/2)": 192, - "height": 512 + "width": 768, + "height": 512, + "floor(height/2)": 256, + "floor(floor(floor(width/2)/2)/2)": 96, + "floor(width/2)": 384, + "floor(floor(width/2)/2)": 192 }, { - "floor(floor(floor(width/2)/2)/2)": 64, "floor(floor(height/2)/2)": 192, - "floor(height/2)": 384, - "width": 512, - "floor(width/2)": 256, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 6144, "floor(floor(floor(height/2)/2)/2)": 96, - "floor(floor(width/2)/2)": 128, - "height": 768 + "width": 512, + "height": 768, + "floor(height/2)": 384, + "floor(floor(floor(width/2)/2)/2)": 64, + "floor(width/2)": 256, + "floor(floor(width/2)/2)": 128 }, { - "floor(floor(floor(width/2)/2)/2)": 128, "floor(floor(height/2)/2)": 144, - "floor(height/2)": 288, - "width": 1024, - "floor(width/2)": 512, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 9216, "floor(floor(floor(height/2)/2)/2)": 72, - "floor(floor(width/2)/2)": 256, - "height": 576 + "width": 1024, + "height": 576, + "floor(height/2)": 288, + "floor(floor(floor(width/2)/2)/2)": 128, + "floor(width/2)": 512, + "floor(floor(width/2)/2)": 256 }, { - "floor(floor(floor(width/2)/2)/2)": 72, "floor(floor(height/2)/2)": 256, - "floor(height/2)": 512, - "width": 576, - "floor(width/2)": 288, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 9216, "floor(floor(floor(height/2)/2)/2)": 128, - "floor(floor(width/2)/2)": 144, - "height": 1024 + "width": 576, + "height": 1024, + "floor(height/2)": 512, + "floor(floor(floor(width/2)/2)/2)": 72, + "floor(width/2)": 288, + "floor(floor(width/2)/2)": 144 }, { - "floor(floor(floor(width/2)/2)/2)": 128, "floor(floor(height/2)/2)": 192, - "floor(height/2)": 384, - "width": 1024, - "floor(width/2)": 512, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 12288, "floor(floor(floor(height/2)/2)/2)": 96, - "floor(floor(width/2)/2)": 256, - "height": 768 + "width": 1024, + "height": 768, + "floor(height/2)": 384, + "floor(floor(floor(width/2)/2)/2)": 128, + "floor(width/2)": 512, + "floor(floor(width/2)/2)": 256 }, { - "floor(floor(floor(width/2)/2)/2)": 96, "floor(floor(height/2)/2)": 256, - "floor(height/2)": 512, - "width": 768, - "floor(width/2)": 384, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 12288, "floor(floor(floor(height/2)/2)/2)": 128, - "floor(floor(width/2)/2)": 192, - "height": 1024 + "width": 768, + "height": 1024, + "floor(height/2)": 512, + "floor(floor(floor(width/2)/2)/2)": 96, + "floor(width/2)": 384, + "floor(floor(width/2)/2)": 192 }, { - "floor(floor(floor(width/2)/2)/2)": 128, "floor(floor(height/2)/2)": 256, - "floor(height/2)": 512, - "width": 1024, - "floor(width/2)": 512, "floor(floor(floor(height/2)/2)/2)*floor(floor(floor(width/2)/2)/2)": 16384, "floor(floor(floor(height/2)/2)/2)": 128, - "floor(floor(width/2)/2)": 256, - "height": 1024 + "width": 1024, + "height": 1024, + "floor(height/2)": 512, + "floor(floor(floor(width/2)/2)/2)": 128, + "floor(width/2)": 512, + "floor(floor(width/2)/2)": 256 } ], "aux_info": {} diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.ddbundle b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..2407ba4201d1e550225e8eb4824b1727802cad2c --- /dev/null +++ b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90399aab5e6d902a9b97d50883cf8e53c8e27b87a1dc17e9e5c6aaf51d4ceeea +size 205536673 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.state b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.state index d6b7ebe4471342938eafdf35c04c6f7ba1966bf4..86f438f431b2935741fd2bec020fa0a48fe9a14a 100644 Binary files a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.state and b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv.state differ diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.ctrlpkt deleted file mode 100644 index b1073042a7e95cf83d70860eaf41174d3b10bb73..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:5d6cefaad66a9fa11818def0ca30c6e722a5b698d1d326466858c269f3dfdded -size 31213128 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.meta deleted file mode 100644 index 71c3316ecb69e0521b885d19da9c6510c5abbbb0..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:494b485d60b1135de1c5b7341bbd36bc84a81ac78414788980a31f3d82721aac -size 5755588 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_1024_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.ctrlpkt deleted file mode 100644 index e6b52a152bf163438b539264491e0ce20df58ebf..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:e618ba49b852d30f7e3ee42546a826a151f06be25638b21a3298729b0db89197 -size 24744432 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.meta deleted file mode 100644 index 7298b782f231fa1dc333f1adee64b35074c17f06..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ee6de5822ee9ce8c272c7a37cae731f62da585e5ebb0e750425aea1fdd15e6f3 -size 4709487 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_576_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.ctrlpkt deleted file mode 100644 index dcef026bf5424ee6fc4ddce39a6910109642c546..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:2548fb53eb29fd3e898f2e5407b97df639717315d1d7a1f70aa363f86826e637 -size 25063152 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.meta deleted file mode 100644 index 99d54e4f1ca0ec15a40e656df79a89d0b6c08c92..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:8e88ea74103b76fa44fe266088de0089cb361d79a7aa77e1feb63be3b177e604 -size 4523555 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_1024_width_768_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.ctrlpkt deleted file mode 100644 index 8cd9d6490963834fc76779fbb27051aedeae305d..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ad97d41627eb8b54ecd9b7e1a96652f92e6560c846a21d0d6d0b902f627ea918 -size 8814432 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.meta deleted file mode 100644 index 9e60c72a8ac96675c7204ac7fb3203a79ad6374b..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ded8e44aceeceba9d132fcd9a5df848150e3580f845830ac063742e7fd5bbb37 -size 1671018 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_512_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.ctrlpkt deleted file mode 100644 index ea7c0368fdb6b2046e198fb96c41ff351de25987..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:510c92fa93b1741fb6a5cd53ec4927cdb036b494cb0ce51e6c9ce2562b22b425 -size 13993920 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.meta deleted file mode 100644 index c853bc9323deebea1dc98bba854bc0ce9829af07..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:012231287ef078fb0428bf8592c882925da9d8fee231b9d8132042f08181c0db -size 2662601 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_512_width_768_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.ctrlpkt deleted file mode 100644 index 9a3e00228af6dea4ab15cde545f9326949629a35..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:f8453e1e88ddd0858ecacffda76d710c7760a0b1337dab26455414760186e193 -size 23492208 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.meta deleted file mode 100644 index 35da8bc71c43ec418a7cf6bc5b1ab98d1a86d713..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:b1afa2bd8c81d6a5798de131dbf060eb45d2fef32be14271a6e0772822ff4181 -size 3741981 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_576_width_1024_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.ctrlpkt deleted file mode 100644 index 49d1e0822cb28054aef111b137d21bb8b75a17f6..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:323c8e80ee32630c8272b0657fb8b1f135498fcbabb50443160babf36ccefe12 -size 30167184 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.meta deleted file mode 100644 index e62b87089dd923b28234c0552ce1b4e80611efef..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:f6ba703864f07dc4e8c7fa9a30acc4d951fa3941d069bdfafd90089bfa748ae3 -size 4783156 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_1024_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.ctrlpkt b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.ctrlpkt deleted file mode 100644 index 1c7a956a44f944ed588f79c8164d24e746055311..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:7825bb17ea54baaab54f640e5e6a19d54fe3470701965611ac6b28f890225c09 -size 12699120 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.meta b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.meta deleted file mode 100644 index 4de586104722b14058285eadcfc72166ae3656fe..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:5252b9ab5f21d3b6f83ec12b0b9e4e220fcd6e7dedcee4ea66415893dfad6571 -size 2321377 diff --git a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.super b/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.super deleted file mode 100644 index e0697be38f142e371c23c0b2e9ed23b5d139c2db..0000000000000000000000000000000000000000 --- a/common/vae_encoder/dynamic/dd/dd_metastate_Sd30_VAE_NhwcConv_0-encoderconv_inConv_height_768_width_512_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:230fe28efa807448cf48c4d03c05becc580249daabc6b2a7120f35cdbcca2414 -size 772640 diff --git a/common/vae_encoder/dynamic/dd/replaced.onnx b/common/vae_encoder/dynamic/dd/replaced.onnx index 64cc8a914896d99c5df4cd8a1ea40cd75456f9c9..2303c3aa805eba2153abe6bbdd18201bd4985d0d 100644 --- a/common/vae_encoder/dynamic/dd/replaced.onnx +++ b/common/vae_encoder/dynamic/dd/replaced.onnx @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:06f4609bd513dc63b068d739c84658c3288d94977d9ebff6fd6348dd2c160997 +oid sha256:d1290483f126caeecb977b8ce22a404584e3737c5c6b003f3196444f7ada0517 size 39313 diff --git a/inpainting/controlnet-inpainting/config.json b/inpainting/controlnet-inpainting/config.json new file mode 100644 index 0000000000000000000000000000000000000000..84f2b3cce0dbf3dd7e3eef8dee7ccebd786a4ac0 --- /dev/null +++ b/inpainting/controlnet-inpainting/config.json @@ -0,0 +1,22 @@ +{ + "_class_name": "SD3ControlNetModel", + "_diffusers_version": "0.33.1", + "_name_or_path": "alimama-creative/SD3-Controlnet-Inpainting", + "attention_head_dim": 64, + "caption_projection_dim": 1536, + "dual_attention_layers": [], + "extra_conditioning_channels": 1, + "force_zeros_for_pooled_projection": true, + "in_channels": 16, + "joint_attention_dim": 4096, + "num_attention_heads": 24, + "num_layers": 23, + "out_channels": 16, + "patch_size": 2, + "pooled_projection_dim": 2048, + "pos_embed_max_size": 192, + "pos_embed_type": "sincos", + "qk_norm": null, + "sample_size": 128, + "use_pos_embed": true +} diff --git a/inpainting/controlnet-inpainting/dynamic/dd/cache/pos_embedprojConv_meta.json b/inpainting/controlnet-inpainting/dynamic/dd/cache/pos_embedprojConv_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..951ea2f2ad169cde0399c01f3b1f2c285ed8cbb6 --- /dev/null +++ b/inpainting/controlnet-inpainting/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -0,0 +1,62691 @@ +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 4, + "state_table_updates": [], + "op_list": [ + { + "name": "/pos_embed/proj/Conv", + "type": "SDConv", + "in_args": [ + "hidden_states_nhwc.out5_0_0" + ], + "const_args": [ + "pos_embed.proj.weight" + ], + "out_args": [ + "/pos_embed/Transpose_output_0.out5_0_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "w", + "h", + "16" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(w/2)", + "floor(h/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "2", + "2", + "16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/pos_embed/Add_2", + "type": "SDAdd", + "in_args": [ + "/pos_embed/Transpose_output_0.out5_0_0", + "/pos_embed/Reshape_1_output_0.out_35_1_2" + ], + "const_args": [], + "out_args": [ + "/pos_embed/Add_2_output_0.out_35_1_2" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/pos_embed_input/proj/Conv", + "type": "SDConv", + "in_args": [ + "controlnet_cond_nhwc.out5_0_1" + ], + "const_args": [ + "pos_embed_input.proj.weight" + ], + "out_args": [ + "/pos_embed_input/Transpose_output_0.out5_0_1" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "w", + "h", + "24" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(w/2)", + "floor(h/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "2", + "2", + "24" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add", + "type": "SDAdd", + "in_args": [ + "/pos_embed/Add_2_output_0.out_35_1_2", + "/pos_embed_input/Transpose_output_0.out5_0_1" + ], + "const_args": [], + "out_args": [ + "/Add_output_0.out_35_1_3" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/Cast_output_0.out17_4_3" + ], + "const_args": [ + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/act/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" + ], + "const_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" + ], + "const_args": [ + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "pooled_projections.out17_4_1" + ], + "const_args": [ + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0" + ], + "out_args": [ + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "2048" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/act_1/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" + ], + "const_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0" + ], + "out_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" + ], + "const_args": [ + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1" + ], + "out_args": [ + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/Add", + "type": "SDAdd", + "in_args": [ + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" + ], + "const_args": [], + "out_args": [ + "/time_text_embed/Add_output_0.out_35_1_4" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/silu/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/Add_output_0.out_35_1_4" + ], + "const_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2" + ], + "out_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "encoder_hidden_states.out17_4_0_SDCastBf2Bfp", + "type": "SDCastBf2Bfp", + "in_args": [ + "encoder_hidden_states.out17_4_0" + ], + "const_args": [ + "encoder_hidden_states.out17_4_0_bfp.wts" + ], + "out_args": [ + "encoder_hidden_states.out17_4_0_bfp.out25_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "4096" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "4096" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/context_embedder/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "encoder_hidden_states.out17_4_0_bfp.out25_0" + ], + "const_args": [ + "onnx::MatMul_7927_45_1_0" + ], + "out_args": [ + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "4096" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "4096", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_SDCastBfp2Bf", + "type": "SDCastBfp2Bf", + "in_args": [ + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" + ], + "const_args": [ + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts" + ], + "out_args": [ + "/context_embedder/Add_output_0.out17_4_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/context_embedder/Add_output_0.out17_4_0" + ], + "const_args": [ + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0" + ], + "out_args": [ + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" + ], + "const_args": [ + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0" + ], + "out_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_output_0.out_35_1_3" + ], + "const_args": [ + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1" + ], + "out_args": [ + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" + ], + "const_args": [ + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3" + ], + "out_args": [ + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_90" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_90" + ], + "const_args": [ + "onnx::MatMul_7943_onnx::MatMul_7928" + ], + "out_args": [ + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_90" + ], + "const_args": [ + "onnx::MatMul_7944_onnx::MatMul_7929" + ], + "out_args": [ + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_90" + ], + "const_args": [ + "onnx::MatMul_7945_onnx::MatMul_7930" + ], + "out_args": [ + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/MatMulmha_18_0_0", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0", + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1", + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0" + ], + "const_args": [ + "onnx::MatMul_7953_45_5_1" + ], + "out_args": [ + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", + "/context_embedder/Add_output_0.out17_4_0" + ], + "const_args": [ + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_4_output_0.out10_0" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.0/Add_4_output_0.out10_0" + ], + "const_args": [ + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3" + ], + "out_args": [ + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" + ], + "const_args": [ + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2" + ], + "out_args": [ + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0" + ], + "const_args": [ + "onnx::MatMul_7954_45_5_0" + ], + "out_args": [ + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" + ], + "const_args": [ + "onnx::MatMul_7957_45_1_2" + ], + "out_args": [ + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" + ], + "const_args": [ + "onnx::MatMul_7958_45_1_4" + ], + "out_args": [ + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", + "/transformer_blocks.0/Add_4_output_0.out10_0" + ], + "const_args": [ + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_7_output_0.out10_1" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", + "/Add_output_0.out_35_1_3" + ], + "const_args": [ + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_output_0.out10_88" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.0/Add_output_0.out10_88" + ], + "const_args": [ + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2" + ], + "out_args": [ + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" + ], + "const_args": [ + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0" + ], + "out_args": [ + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_89" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_89" + ], + "const_args": [ + "onnx::MatMul_7955_45_1_1" + ], + "out_args": [ + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" + ], + "const_args": [ + "onnx::MatMul_7956_45_1_3" + ], + "out_args": [ + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", + "/transformer_blocks.0/Add_output_0.out10_88" + ], + "const_args": [ + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_3_output_0.out10_89" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.0/Add_3_output_0.out10_89" + ], + "const_args": [ + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4" + ], + "out_args": [ + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" + ], + "const_args": [ + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1" + ], + "out_args": [ + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.0/Add_3_output_0.out10_89" + ], + "const_args": [ + "onnx::MatMul_8635_45_0_4" + ], + "out_args": [ + "/controlnet_blocks.0/Add_output_0.out17_4_13" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.0/Add_7_output_0.out10_1" + ], + "const_args": [ + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5" + ], + "out_args": [ + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" + ], + "const_args": [ + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1" + ], + "out_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_5_bfp.out1_4" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_5_bfp.out1_4", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2" + ], + "const_args": [ + "onnx::MatMul_7974_onnx::MatMul_7959" + ], + "out_args": [ + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_5_bfp.out1_4", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2" + ], + "const_args": [ + "onnx::MatMul_7975_onnx::MatMul_7960" + ], + "out_args": [ + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_5_bfp.out1_4", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2" + ], + "const_args": [ + "onnx::MatMul_7976_onnx::MatMul_7961" + ], + "out_args": [ + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/MatMulmha_18_0_1", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3", + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4", + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1" + ], + "const_args": [ + "onnx::MatMul_7985_45_5_2" + ], + "out_args": [ + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2", + "/transformer_blocks.0/Add_3_output_0.out10_89" + ], + "const_args": [ + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_output_0.out10_2" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.1/Add_output_0.out10_2" + ], + "const_args": [ + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6" + ], + "out_args": [ + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" + ], + "const_args": [ + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4" + ], + "out_args": [ + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" + ], + "const_args": [ + "onnx::MatMul_7986_45_1_5" + ], + "out_args": [ + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" + ], + "const_args": [ + "onnx::MatMul_7987_45_1_7" + ], + "out_args": [ + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", + "/transformer_blocks.1/Add_output_0.out10_2" + ], + "const_args": [ + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_3_output_0.out10_3" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.1/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.1/Add_3_output_0.out10_3" + ], + "const_args": [ + "onnx::MatMul_8636_45_0_5" + ], + "out_args": [ + "/controlnet_blocks.1/Add_output_0.out17_4_22" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1" + ], + "const_args": [ + "onnx::MatMul_7984_45_5_3" + ], + "out_args": [ + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", + "/transformer_blocks.0/Add_7_output_0.out10_1" + ], + "const_args": [ + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_4_output_0.out10_4" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.1/Add_4_output_0.out10_4" + ], + "const_args": [ + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7" + ], + "out_args": [ + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" + ], + "const_args": [ + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4" + ], + "out_args": [ + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" + ], + "const_args": [ + "onnx::MatMul_7988_45_1_6" + ], + "out_args": [ + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" + ], + "const_args": [ + "onnx::MatMul_7989_45_1_8" + ], + "out_args": [ + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", + "/transformer_blocks.1/Add_4_output_0.out10_4" + ], + "const_args": [ + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_7_output_0.out10_5" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.1/Add_3_output_0.out10_3" + ], + "const_args": [ + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8" + ], + "out_args": [ + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.1/Add_7_output_0.out10_5" + ], + "const_args": [ + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9" + ], + "out_args": [ + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" + ], + "const_args": [ + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1" + ], + "out_args": [ + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_46" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" + ], + "const_args": [ + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1" + ], + "out_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_9_bfp.out1_48" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_9_bfp.out1_48", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_46" + ], + "const_args": [ + "onnx::MatMul_8005_onnx::MatMul_7990" + ], + "out_args": [ + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_9_bfp.out1_48", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_46" + ], + "const_args": [ + "onnx::MatMul_8006_onnx::MatMul_7991" + ], + "out_args": [ + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_9_bfp.out1_48", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_46" + ], + "const_args": [ + "onnx::MatMul_8007_onnx::MatMul_7992" + ], + "out_args": [ + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/MatMulmha_18_0_2", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6", + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7", + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2" + ], + "const_args": [ + "onnx::MatMul_8015_45_5_5" + ], + "out_args": [ + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5", + "/transformer_blocks.1/Add_7_output_0.out10_5" + ], + "const_args": [ + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_4_output_0.out10_48" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2" + ], + "const_args": [ + "onnx::MatMul_8016_45_5_4" + ], + "out_args": [ + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4", + "/transformer_blocks.1/Add_3_output_0.out10_3" + ], + "const_args": [ + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_output_0.out10_46" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.2/Add_output_0.out10_46" + ], + "const_args": [ + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10" + ], + "out_args": [ + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" + ], + "const_args": [ + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4" + ], + "out_args": [ + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47" + ], + "const_args": [ + "onnx::MatMul_8017_45_1_9" + ], + "out_args": [ + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" + ], + "const_args": [ + "onnx::MatMul_8018_45_1_11" + ], + "out_args": [ + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", + "/transformer_blocks.2/Add_output_0.out10_46" + ], + "const_args": [ + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_3_output_0.out10_47" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.2/Add_3_output_0.out10_47" + ], + "const_args": [ + "onnx::MatMul_8637_45_0_6" + ], + "out_args": [ + "/controlnet_blocks.2/Add_output_0.out17_4_31" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.2/Add_4_output_0.out10_48" + ], + "const_args": [ + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11" + ], + "out_args": [ + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" + ], + "const_args": [ + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4" + ], + "out_args": [ + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49" + ], + "const_args": [ + "onnx::MatMul_8019_45_1_10" + ], + "out_args": [ + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" + ], + "const_args": [ + "onnx::MatMul_8020_45_1_12" + ], + "out_args": [ + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", + "/transformer_blocks.2/Add_4_output_0.out10_48" + ], + "const_args": [ + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_7_output_0.out10_49" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.2/Add_3_output_0.out10_47" + ], + "const_args": [ + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12" + ], + "out_args": [ + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.2/Add_7_output_0.out10_49" + ], + "const_args": [ + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13" + ], + "out_args": [ + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" + ], + "const_args": [ + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1" + ], + "out_args": [ + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_61" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" + ], + "const_args": [ + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1" + ], + "out_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_13_bfp.out1_63" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_13_bfp.out1_63", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_61" + ], + "const_args": [ + "onnx::MatMul_8036_onnx::MatMul_8021" + ], + "out_args": [ + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_13_bfp.out1_63", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_61" + ], + "const_args": [ + "onnx::MatMul_8037_onnx::MatMul_8022" + ], + "out_args": [ + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_13_bfp.out1_63", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_61" + ], + "const_args": [ + "onnx::MatMul_8038_onnx::MatMul_8023" + ], + "out_args": [ + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/MatMulmha_18_0_3", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9", + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10", + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3" + ], + "const_args": [ + "onnx::MatMul_8046_45_5_7" + ], + "out_args": [ + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7", + "/transformer_blocks.2/Add_7_output_0.out10_49" + ], + "const_args": [ + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_4_output_0.out10_62" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3" + ], + "const_args": [ + "onnx::MatMul_8047_45_5_6" + ], + "out_args": [ + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6", + "/transformer_blocks.2/Add_3_output_0.out10_47" + ], + "const_args": [ + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_output_0.out10_60" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.3/Add_output_0.out10_60" + ], + "const_args": [ + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14" + ], + "out_args": [ + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" + ], + "const_args": [ + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4" + ], + "out_args": [ + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_62" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_62" + ], + "const_args": [ + "onnx::MatMul_8048_45_1_13" + ], + "out_args": [ + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" + ], + "const_args": [ + "onnx::MatMul_8049_45_1_15" + ], + "out_args": [ + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", + "/transformer_blocks.3/Add_output_0.out10_60" + ], + "const_args": [ + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_3_output_0.out10_61" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.3/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.3/Add_3_output_0.out10_61" + ], + "const_args": [ + "onnx::MatMul_8638_45_0_7" + ], + "out_args": [ + "/controlnet_blocks.3/Add_output_0.out17_4_40" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.3/Add_4_output_0.out10_62" + ], + "const_args": [ + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15" + ], + "out_args": [ + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" + ], + "const_args": [ + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4" + ], + "out_args": [ + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_64" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_64" + ], + "const_args": [ + "onnx::MatMul_8050_45_1_14" + ], + "out_args": [ + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" + ], + "const_args": [ + "onnx::MatMul_8051_45_1_16" + ], + "out_args": [ + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", + "/transformer_blocks.3/Add_4_output_0.out10_62" + ], + "const_args": [ + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_7_output_0.out10_63" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.3/Add_3_output_0.out10_61" + ], + "const_args": [ + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16" + ], + "out_args": [ + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.3/Add_7_output_0.out10_63" + ], + "const_args": [ + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17" + ], + "out_args": [ + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" + ], + "const_args": [ + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1" + ], + "out_args": [ + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_65" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" + ], + "const_args": [ + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1" + ], + "out_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_17_bfp.out1_67" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_17_bfp.out1_67", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_65" + ], + "const_args": [ + "onnx::MatMul_8067_onnx::MatMul_8052" + ], + "out_args": [ + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_17_bfp.out1_67", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_65" + ], + "const_args": [ + "onnx::MatMul_8068_onnx::MatMul_8053" + ], + "out_args": [ + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_17_bfp.out1_67", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_65" + ], + "const_args": [ + "onnx::MatMul_8069_onnx::MatMul_8054" + ], + "out_args": [ + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/MatMulmha_18_0_4", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12", + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13", + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4" + ], + "const_args": [ + "onnx::MatMul_8077_45_5_9" + ], + "out_args": [ + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9", + "/transformer_blocks.3/Add_7_output_0.out10_63" + ], + "const_args": [ + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_4_output_0.out10_66" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4" + ], + "const_args": [ + "onnx::MatMul_8078_45_5_8" + ], + "out_args": [ + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8", + "/transformer_blocks.3/Add_3_output_0.out10_61" + ], + "const_args": [ + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_output_0.out10_64" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.4/Add_output_0.out10_64" + ], + "const_args": [ + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18" + ], + "out_args": [ + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" + ], + "const_args": [ + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4" + ], + "out_args": [ + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_66" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_66" + ], + "const_args": [ + "onnx::MatMul_8079_45_1_17" + ], + "out_args": [ + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" + ], + "const_args": [ + "onnx::MatMul_8080_45_1_19" + ], + "out_args": [ + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", + "/transformer_blocks.4/Add_output_0.out10_64" + ], + "const_args": [ + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_3_output_0.out10_65" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.4/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.4/Add_3_output_0.out10_65" + ], + "const_args": [ + "onnx::MatMul_8639_45_0_8" + ], + "out_args": [ + "/controlnet_blocks.4/Add_output_0.out17_4_49" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.4/Add_4_output_0.out10_66" + ], + "const_args": [ + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19" + ], + "out_args": [ + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" + ], + "const_args": [ + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4" + ], + "out_args": [ + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_68" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_68" + ], + "const_args": [ + "onnx::MatMul_8081_45_1_18" + ], + "out_args": [ + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" + ], + "const_args": [ + "onnx::MatMul_8082_45_1_20" + ], + "out_args": [ + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", + "/transformer_blocks.4/Add_4_output_0.out10_66" + ], + "const_args": [ + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_7_output_0.out10_67" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.4/Add_3_output_0.out10_65" + ], + "const_args": [ + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20" + ], + "out_args": [ + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.4/Add_7_output_0.out10_67" + ], + "const_args": [ + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21" + ], + "out_args": [ + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" + ], + "const_args": [ + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1" + ], + "out_args": [ + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_69" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" + ], + "const_args": [ + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1" + ], + "out_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_21_bfp.out1_71" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_21_bfp.out1_71", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_69" + ], + "const_args": [ + "onnx::MatMul_8098_onnx::MatMul_8083" + ], + "out_args": [ + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_21_bfp.out1_71", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_69" + ], + "const_args": [ + "onnx::MatMul_8099_onnx::MatMul_8084" + ], + "out_args": [ + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_21_bfp.out1_71", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_69" + ], + "const_args": [ + "onnx::MatMul_8100_onnx::MatMul_8085" + ], + "out_args": [ + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/MatMulmha_18_0_5", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15", + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16", + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5" + ], + "const_args": [ + "onnx::MatMul_8108_45_5_11" + ], + "out_args": [ + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11", + "/transformer_blocks.4/Add_7_output_0.out10_67" + ], + "const_args": [ + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_2_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_4_output_0.out10_70" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5" + ], + "const_args": [ + "onnx::MatMul_8109_45_5_10" + ], + "out_args": [ + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10", + "/transformer_blocks.4/Add_3_output_0.out10_65" + ], + "const_args": [ + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_output_0.out10_68" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.5/Add_output_0.out10_68" + ], + "const_args": [ + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22" + ], + "out_args": [ + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" + ], + "const_args": [ + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4" + ], + "out_args": [ + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_70" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_70" + ], + "const_args": [ + "onnx::MatMul_8110_45_1_21" + ], + "out_args": [ + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" + ], + "const_args": [ + "onnx::MatMul_8111_45_1_23" + ], + "out_args": [ + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32", + "/transformer_blocks.5/Add_output_0.out10_68" + ], + "const_args": [ + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_3_output_0.out10_69" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.5/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.5/Add_3_output_0.out10_69" + ], + "const_args": [ + "onnx::MatMul_8640_45_0_9" + ], + "out_args": [ + "/controlnet_blocks.5/Add_output_0.out17_4_58" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.5/Add_4_output_0.out10_70" + ], + "const_args": [ + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23" + ], + "out_args": [ + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23" + ], + "const_args": [ + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_3_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_4" + ], + "out_args": [ + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_72" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_72" + ], + "const_args": [ + "onnx::MatMul_8112_45_1_22" + ], + "out_args": [ + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_31" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_31" + ], + "const_args": [ + "onnx::MatMul_8113_45_1_24" + ], + "out_args": [ + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33", + "/transformer_blocks.5/Add_4_output_0.out10_70" + ], + "const_args": [ + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_5_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_7_output_0.out10_71" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.5/Add_3_output_0.out10_69" + ], + "const_args": [ + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_24" + ], + "out_args": [ + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_24_bfp.out15_24" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.5/Add_7_output_0.out10_71" + ], + "const_args": [ + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_25" + ], + "out_args": [ + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_25_bfp.out15_25" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_24_bfp.out15_24" + ], + "const_args": [ + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_0_transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_1" + ], + "out_args": [ + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_24_bfp.out1_73" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_25_bfp.out15_25" + ], + "const_args": [ + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_0_transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_1" + ], + "out_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_25_bfp.out1_75" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_25_bfp.out1_75", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_24_bfp.out1_73" + ], + "const_args": [ + "onnx::MatMul_8129_onnx::MatMul_8114" + ], + "out_args": [ + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_25_bfp.out1_75", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_24_bfp.out1_73" + ], + "const_args": [ + "onnx::MatMul_8130_onnx::MatMul_8115" + ], + "out_args": [ + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_25_bfp.out1_75", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_24_bfp.out1_73" + ], + "const_args": [ + "onnx::MatMul_8131_onnx::MatMul_8116" + ], + "out_args": [ + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/MatMulmha_18_0_6", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18", + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19", + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6" + ], + "const_args": [ + "onnx::MatMul_8139_45_5_13" + ], + "out_args": [ + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13", + "/transformer_blocks.5/Add_7_output_0.out10_71" + ], + "const_args": [ + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_2_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_4_output_0.out10_74" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6" + ], + "const_args": [ + "onnx::MatMul_8140_45_5_12" + ], + "out_args": [ + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12", + "/transformer_blocks.5/Add_3_output_0.out10_69" + ], + "const_args": [ + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_2_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_output_0.out10_72" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.6/Add_output_0.out10_72" + ], + "const_args": [ + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26" + ], + "out_args": [ + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26" + ], + "const_args": [ + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_3_transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_4" + ], + "out_args": [ + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_74" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_74" + ], + "const_args": [ + "onnx::MatMul_8141_45_1_25" + ], + "out_args": [ + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_35" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_35" + ], + "const_args": [ + "onnx::MatMul_8142_45_1_27" + ], + "out_args": [ + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37", + "/transformer_blocks.6/Add_output_0.out10_72" + ], + "const_args": [ + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_5_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_3_output_0.out10_73" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.6/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.6/Add_3_output_0.out10_73" + ], + "const_args": [ + "onnx::MatMul_8641_45_0_10" + ], + "out_args": [ + "/controlnet_blocks.6/Add_output_0.out17_4_67" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.6/Add_4_output_0.out10_74" + ], + "const_args": [ + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27" + ], + "out_args": [ + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27" + ], + "const_args": [ + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_3_transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_4" + ], + "out_args": [ + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_76" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_76" + ], + "const_args": [ + "onnx::MatMul_8143_45_1_26" + ], + "out_args": [ + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_36" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_36" + ], + "const_args": [ + "onnx::MatMul_8144_45_1_28" + ], + "out_args": [ + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38", + "/transformer_blocks.6/Add_4_output_0.out10_74" + ], + "const_args": [ + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_5_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_7_output_0.out10_75" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.6/Add_3_output_0.out10_73" + ], + "const_args": [ + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_28" + ], + "out_args": [ + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_28_bfp.out15_28" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.6/Add_7_output_0.out10_75" + ], + "const_args": [ + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_29" + ], + "out_args": [ + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_29_bfp.out15_29" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_28_bfp.out15_28" + ], + "const_args": [ + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_0_transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_1" + ], + "out_args": [ + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_28_bfp.out1_77" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_29_bfp.out15_29" + ], + "const_args": [ + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_0_transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_1" + ], + "out_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_29_bfp.out1_79" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_29_bfp.out1_79", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_28_bfp.out1_77" + ], + "const_args": [ + "onnx::MatMul_8160_onnx::MatMul_8145" + ], + "out_args": [ + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_29_bfp.out1_79", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_28_bfp.out1_77" + ], + "const_args": [ + "onnx::MatMul_8161_onnx::MatMul_8146" + ], + "out_args": [ + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_29_bfp.out1_79", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_28_bfp.out1_77" + ], + "const_args": [ + "onnx::MatMul_8162_onnx::MatMul_8147" + ], + "out_args": [ + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/MatMulmha_18_0_7", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21", + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22", + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7" + ], + "const_args": [ + "onnx::MatMul_8170_45_5_15" + ], + "out_args": [ + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15", + "/transformer_blocks.6/Add_7_output_0.out10_75" + ], + "const_args": [ + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_2_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_4_output_0.out10_78" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7" + ], + "const_args": [ + "onnx::MatMul_8171_45_5_14" + ], + "out_args": [ + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14", + "/transformer_blocks.6/Add_3_output_0.out10_73" + ], + "const_args": [ + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_2_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_output_0.out10_76" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.7/Add_output_0.out10_76" + ], + "const_args": [ + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30" + ], + "out_args": [ + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30" + ], + "const_args": [ + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_3_transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_4" + ], + "out_args": [ + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_78" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_78" + ], + "const_args": [ + "onnx::MatMul_8172_45_1_29" + ], + "out_args": [ + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_40" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_40" + ], + "const_args": [ + "onnx::MatMul_8173_45_1_31" + ], + "out_args": [ + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42", + "/transformer_blocks.7/Add_output_0.out10_76" + ], + "const_args": [ + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_5_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_3_output_0.out10_77" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.7/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.7/Add_3_output_0.out10_77" + ], + "const_args": [ + "onnx::MatMul_8642_45_0_11" + ], + "out_args": [ + "/controlnet_blocks.7/Add_output_0.out17_4_76" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.7/Add_4_output_0.out10_78" + ], + "const_args": [ + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31" + ], + "out_args": [ + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31" + ], + "const_args": [ + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_3_transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_4" + ], + "out_args": [ + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_80" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_80" + ], + "const_args": [ + "onnx::MatMul_8174_45_1_30" + ], + "out_args": [ + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_41" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_41" + ], + "const_args": [ + "onnx::MatMul_8175_45_1_32" + ], + "out_args": [ + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43", + "/transformer_blocks.7/Add_4_output_0.out10_78" + ], + "const_args": [ + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_5_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_7_output_0.out10_79" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.7/Add_3_output_0.out10_77" + ], + "const_args": [ + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_32" + ], + "out_args": [ + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_32_bfp.out15_32" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.7/Add_7_output_0.out10_79" + ], + "const_args": [ + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_33" + ], + "out_args": [ + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_33_bfp.out15_33" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_32_bfp.out15_32" + ], + "const_args": [ + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_0_transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_1" + ], + "out_args": [ + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_32_bfp.out1_81" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_33_bfp.out15_33" + ], + "const_args": [ + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_0_transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_1" + ], + "out_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_33_bfp.out1_83" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_33_bfp.out1_83", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_32_bfp.out1_81" + ], + "const_args": [ + "onnx::MatMul_8191_onnx::MatMul_8176" + ], + "out_args": [ + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_33_bfp.out1_83", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_32_bfp.out1_81" + ], + "const_args": [ + "onnx::MatMul_8192_onnx::MatMul_8177" + ], + "out_args": [ + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_33_bfp.out1_83", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_32_bfp.out1_81" + ], + "const_args": [ + "onnx::MatMul_8193_onnx::MatMul_8178" + ], + "out_args": [ + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/MatMulmha_18_0_8", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24", + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25", + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8" + ], + "const_args": [ + "onnx::MatMul_8201_45_5_17" + ], + "out_args": [ + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17", + "/transformer_blocks.7/Add_7_output_0.out10_79" + ], + "const_args": [ + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_2_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_4_output_0.out10_82" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8" + ], + "const_args": [ + "onnx::MatMul_8202_45_5_16" + ], + "out_args": [ + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16", + "/transformer_blocks.7/Add_3_output_0.out10_77" + ], + "const_args": [ + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_2_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_output_0.out10_80" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.8/Add_output_0.out10_80" + ], + "const_args": [ + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34" + ], + "out_args": [ + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34" + ], + "const_args": [ + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_3_transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_4" + ], + "out_args": [ + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_82" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_82" + ], + "const_args": [ + "onnx::MatMul_8203_45_1_33" + ], + "out_args": [ + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_45" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_45" + ], + "const_args": [ + "onnx::MatMul_8204_45_1_35" + ], + "out_args": [ + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47", + "/transformer_blocks.8/Add_output_0.out10_80" + ], + "const_args": [ + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_5_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_3_output_0.out10_81" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.8/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.8/Add_3_output_0.out10_81" + ], + "const_args": [ + "onnx::MatMul_8643_45_0_12" + ], + "out_args": [ + "/controlnet_blocks.8/Add_output_0.out17_4_85" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.8/Add_4_output_0.out10_82" + ], + "const_args": [ + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35" + ], + "out_args": [ + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35" + ], + "const_args": [ + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_3_transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_4" + ], + "out_args": [ + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_84" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_84" + ], + "const_args": [ + "onnx::MatMul_8205_45_1_34" + ], + "out_args": [ + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_46" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_46" + ], + "const_args": [ + "onnx::MatMul_8206_45_1_36" + ], + "out_args": [ + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48", + "/transformer_blocks.8/Add_4_output_0.out10_82" + ], + "const_args": [ + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_5_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_7_output_0.out10_83" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.8/Add_3_output_0.out10_81" + ], + "const_args": [ + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_36" + ], + "out_args": [ + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_36_bfp.out15_36" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.8/Add_7_output_0.out10_83" + ], + "const_args": [ + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_37" + ], + "out_args": [ + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_37_bfp.out15_37" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_36_bfp.out15_36" + ], + "const_args": [ + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_0_transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_1" + ], + "out_args": [ + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_36_bfp.out1_85" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_37_bfp.out15_37" + ], + "const_args": [ + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_0_transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_1" + ], + "out_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_37_bfp.out1_87" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_37_bfp.out1_87", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_36_bfp.out1_85" + ], + "const_args": [ + "onnx::MatMul_8222_onnx::MatMul_8207" + ], + "out_args": [ + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_37_bfp.out1_87", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_36_bfp.out1_85" + ], + "const_args": [ + "onnx::MatMul_8223_onnx::MatMul_8208" + ], + "out_args": [ + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_37_bfp.out1_87", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_36_bfp.out1_85" + ], + "const_args": [ + "onnx::MatMul_8224_onnx::MatMul_8209" + ], + "out_args": [ + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/MatMulmha_18_0_9", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27", + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28", + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9" + ], + "const_args": [ + "onnx::MatMul_8232_45_5_19" + ], + "out_args": [ + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19", + "/transformer_blocks.8/Add_7_output_0.out10_83" + ], + "const_args": [ + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_2_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_4_output_0.out10_86" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9" + ], + "const_args": [ + "onnx::MatMul_8233_45_5_18" + ], + "out_args": [ + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18", + "/transformer_blocks.8/Add_3_output_0.out10_81" + ], + "const_args": [ + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_2_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_output_0.out10_84" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.9/Add_output_0.out10_84" + ], + "const_args": [ + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38" + ], + "out_args": [ + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38" + ], + "const_args": [ + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_3_transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_4" + ], + "out_args": [ + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_86" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_86" + ], + "const_args": [ + "onnx::MatMul_8234_45_1_37" + ], + "out_args": [ + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_50" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_50" + ], + "const_args": [ + "onnx::MatMul_8235_45_1_39" + ], + "out_args": [ + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52", + "/transformer_blocks.9/Add_output_0.out10_84" + ], + "const_args": [ + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_5_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_3_output_0.out10_85" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.9/Add_3_output_0.out10_85" + ], + "const_args": [ + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40" + ], + "out_args": [ + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40" + ], + "const_args": [ + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_0_transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_1" + ], + "out_args": [ + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.9/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.9/Add_3_output_0.out10_85" + ], + "const_args": [ + "onnx::MatMul_8644_45_0_13" + ], + "out_args": [ + "/controlnet_blocks.9/Add_output_0.out17_4_94" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.9/Add_4_output_0.out10_86" + ], + "const_args": [ + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39" + ], + "out_args": [ + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39" + ], + "const_args": [ + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_3_transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_4" + ], + "out_args": [ + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_88" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_88" + ], + "const_args": [ + "onnx::MatMul_8236_45_1_38" + ], + "out_args": [ + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_51" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_51" + ], + "const_args": [ + "onnx::MatMul_8237_45_1_40" + ], + "out_args": [ + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53", + "/transformer_blocks.9/Add_4_output_0.out10_86" + ], + "const_args": [ + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_5_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_7_output_0.out10_87" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.9/Add_7_output_0.out10_87" + ], + "const_args": [ + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41" + ], + "out_args": [ + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41" + ], + "const_args": [ + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_0_transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_1" + ], + "out_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "const_args": [ + "onnx::MatMul_8253_onnx::MatMul_8238" + ], + "out_args": [ + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "const_args": [ + "onnx::MatMul_8254_onnx::MatMul_8239" + ], + "out_args": [ + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "const_args": [ + "onnx::MatMul_8255_onnx::MatMul_8240" + ], + "out_args": [ + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/MatMulmha_18_0_10", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30", + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31", + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10" + ], + "const_args": [ + "onnx::MatMul_8264_45_5_20" + ], + "out_args": [ + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20", + "/transformer_blocks.9/Add_3_output_0.out10_85" + ], + "const_args": [ + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_2_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_output_0.out10_6" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.10/Add_output_0.out10_6" + ], + "const_args": [ + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42" + ], + "out_args": [ + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42" + ], + "const_args": [ + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_3_transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_4" + ], + "out_args": [ + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7" + ], + "const_args": [ + "onnx::MatMul_8265_45_1_41" + ], + "out_args": [ + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_55" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_55" + ], + "const_args": [ + "onnx::MatMul_8266_45_1_43" + ], + "out_args": [ + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57", + "/transformer_blocks.10/Add_output_0.out10_6" + ], + "const_args": [ + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_5_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_3_output_0.out10_7" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.10/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.10/Add_3_output_0.out10_7" + ], + "const_args": [ + "onnx::MatMul_8645_45_0_14" + ], + "out_args": [ + "/controlnet_blocks.10/Add_output_0.out17_4_103" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10" + ], + "const_args": [ + "onnx::MatMul_8263_45_5_21" + ], + "out_args": [ + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21", + "/transformer_blocks.9/Add_7_output_0.out10_87" + ], + "const_args": [ + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_2_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_4_output_0.out10_8" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.10/Add_4_output_0.out10_8" + ], + "const_args": [ + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43" + ], + "out_args": [ + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43" + ], + "const_args": [ + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_3_transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_4" + ], + "out_args": [ + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9" + ], + "const_args": [ + "onnx::MatMul_8267_45_1_42" + ], + "out_args": [ + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_56" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_56" + ], + "const_args": [ + "onnx::MatMul_8268_45_1_44" + ], + "out_args": [ + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58", + "/transformer_blocks.10/Add_4_output_0.out10_8" + ], + "const_args": [ + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_5_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_7_output_0.out10_9" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.10/Add_3_output_0.out10_7" + ], + "const_args": [ + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_44" + ], + "out_args": [ + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_44_bfp.out15_44" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_44_bfp.out15_44" + ], + "const_args": [ + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_0_transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_1" + ], + "out_args": [ + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_44_bfp.out1_10" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.10/Add_7_output_0.out10_9" + ], + "const_args": [ + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_45" + ], + "out_args": [ + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_45_bfp.out15_45" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_45_bfp.out15_45" + ], + "const_args": [ + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_0_transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_1" + ], + "out_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_45_bfp.out1_12" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_45_bfp.out1_12", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_44_bfp.out1_10" + ], + "const_args": [ + "onnx::MatMul_8284_onnx::MatMul_8269" + ], + "out_args": [ + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_45_bfp.out1_12", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_44_bfp.out1_10" + ], + "const_args": [ + "onnx::MatMul_8285_onnx::MatMul_8270" + ], + "out_args": [ + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_45_bfp.out1_12", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_44_bfp.out1_10" + ], + "const_args": [ + "onnx::MatMul_8286_onnx::MatMul_8271" + ], + "out_args": [ + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/MatMulmha_18_0_11", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33", + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34", + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11" + ], + "const_args": [ + "onnx::MatMul_8295_45_5_22" + ], + "out_args": [ + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22", + "/transformer_blocks.10/Add_3_output_0.out10_7" + ], + "const_args": [ + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_2_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_output_0.out10_10" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.11/Add_output_0.out10_10" + ], + "const_args": [ + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46" + ], + "out_args": [ + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46" + ], + "const_args": [ + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_3_transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_4" + ], + "out_args": [ + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11" + ], + "const_args": [ + "onnx::MatMul_8296_45_1_45" + ], + "out_args": [ + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_60" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_60" + ], + "const_args": [ + "onnx::MatMul_8297_45_1_47" + ], + "out_args": [ + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_62" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_62", + "/transformer_blocks.11/Add_output_0.out10_10" + ], + "const_args": [ + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_5_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_3_output_0.out10_11" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.11/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.11/Add_3_output_0.out10_11" + ], + "const_args": [ + "onnx::MatMul_8646_45_0_15" + ], + "out_args": [ + "/controlnet_blocks.11/Add_output_0.out17_4_112" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11" + ], + "const_args": [ + "onnx::MatMul_8294_45_5_23" + ], + "out_args": [ + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23", + "/transformer_blocks.10/Add_7_output_0.out10_9" + ], + "const_args": [ + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_2_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_4_output_0.out10_12" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.11/Add_4_output_0.out10_12" + ], + "const_args": [ + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47", + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47" + ], + "out_args": [ + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47" + ], + "const_args": [ + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_3_transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_4" + ], + "out_args": [ + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13" + ], + "const_args": [ + "onnx::MatMul_8298_45_1_46" + ], + "out_args": [ + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_61" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_61" + ], + "const_args": [ + "onnx::MatMul_8299_45_1_48" + ], + "out_args": [ + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_111_bfp.out25_63" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_111_bfp.out25_63", + "/transformer_blocks.11/Add_4_output_0.out10_12" + ], + "const_args": [ + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_5_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_7_output_0.out10_13" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.11/Add_3_output_0.out10_11" + ], + "const_args": [ + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_48", + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_48" + ], + "out_args": [ + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_48_bfp.out15_48" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_48_bfp.out15_48" + ], + "const_args": [ + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_0_transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_1" + ], + "out_args": [ + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_48_bfp.out1_14" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.11/Add_7_output_0.out10_13" + ], + "const_args": [ + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_49", + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_49" + ], + "out_args": [ + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_49_bfp.out15_49" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_49_bfp.out15_49" + ], + "const_args": [ + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_0_transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_1" + ], + "out_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_49_bfp.out1_16" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_49_bfp.out1_16", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_48_bfp.out1_14" + ], + "const_args": [ + "onnx::MatMul_8315_onnx::MatMul_8300" + ], + "out_args": [ + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_49_bfp.out1_16", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_48_bfp.out1_14" + ], + "const_args": [ + "onnx::MatMul_8316_onnx::MatMul_8301" + ], + "out_args": [ + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_49_bfp.out1_16", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_48_bfp.out1_14" + ], + "const_args": [ + "onnx::MatMul_8317_onnx::MatMul_8302" + ], + "out_args": [ + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/MatMulmha_18_0_12", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36", + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37", + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12" + ], + "const_args": [ + "onnx::MatMul_8326_45_5_24" + ], + "out_args": [ + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24", + "/transformer_blocks.11/Add_3_output_0.out10_11" + ], + "const_args": [ + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_2_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_output_0.out10_14" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.12/Add_output_0.out10_14" + ], + "const_args": [ + "/transformer_blocks.12/norm2/Constant_output_0_14_50", + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50" + ], + "out_args": [ + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50" + ], + "const_args": [ + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_3_transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_4" + ], + "out_args": [ + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15" + ], + "const_args": [ + "onnx::MatMul_8327_45_1_49" + ], + "out_args": [ + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_65" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_65" + ], + "const_args": [ + "onnx::MatMul_8328_45_1_51" + ], + "out_args": [ + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_119_bfp.out25_67" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_119_bfp.out25_67", + "/transformer_blocks.12/Add_output_0.out10_14" + ], + "const_args": [ + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_5_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_3_output_0.out10_15" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.12/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.12/Add_3_output_0.out10_15" + ], + "const_args": [ + "onnx::MatMul_8647_45_0_16" + ], + "out_args": [ + "/controlnet_blocks.12/Add_output_0.out17_4_121" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12" + ], + "const_args": [ + "onnx::MatMul_8325_45_5_25" + ], + "out_args": [ + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25", + "/transformer_blocks.11/Add_7_output_0.out10_13" + ], + "const_args": [ + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_2_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_4_output_0.out10_16" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.12/Add_4_output_0.out10_16" + ], + "const_args": [ + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51", + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51" + ], + "out_args": [ + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51" + ], + "const_args": [ + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_3_transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_4" + ], + "out_args": [ + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17" + ], + "const_args": [ + "onnx::MatMul_8329_45_1_50" + ], + "out_args": [ + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_66" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_66" + ], + "const_args": [ + "onnx::MatMul_8330_45_1_52" + ], + "out_args": [ + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_120_bfp.out25_68" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_120_bfp.out25_68", + "/transformer_blocks.12/Add_4_output_0.out10_16" + ], + "const_args": [ + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_5_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_7_output_0.out10_17" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.12/Add_3_output_0.out10_15" + ], + "const_args": [ + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_52", + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_52" + ], + "out_args": [ + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_52_bfp.out15_52" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_52_bfp.out15_52" + ], + "const_args": [ + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_0_transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_1" + ], + "out_args": [ + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_52_bfp.out1_18" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.12/Add_7_output_0.out10_17" + ], + "const_args": [ + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_53", + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_53" + ], + "out_args": [ + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_53_bfp.out15_53" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_53_bfp.out15_53" + ], + "const_args": [ + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_0_transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_1" + ], + "out_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_53_bfp.out1_20" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_53_bfp.out1_20", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_52_bfp.out1_18" + ], + "const_args": [ + "onnx::MatMul_8346_onnx::MatMul_8331" + ], + "out_args": [ + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_53_bfp.out1_20", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_52_bfp.out1_18" + ], + "const_args": [ + "onnx::MatMul_8347_onnx::MatMul_8332" + ], + "out_args": [ + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_53_bfp.out1_20", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_52_bfp.out1_18" + ], + "const_args": [ + "onnx::MatMul_8348_onnx::MatMul_8333" + ], + "out_args": [ + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/MatMulmha_18_0_13", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39", + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40", + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13" + ], + "const_args": [ + "onnx::MatMul_8357_45_5_26" + ], + "out_args": [ + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26", + "/transformer_blocks.12/Add_3_output_0.out10_15" + ], + "const_args": [ + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_2_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_output_0.out10_18" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.13/Add_output_0.out10_18" + ], + "const_args": [ + "/transformer_blocks.13/norm2/Constant_output_0_14_54", + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54" + ], + "out_args": [ + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54" + ], + "const_args": [ + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_3_transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_4" + ], + "out_args": [ + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19" + ], + "const_args": [ + "onnx::MatMul_8358_45_1_53" + ], + "out_args": [ + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_70" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_70" + ], + "const_args": [ + "onnx::MatMul_8359_45_1_55" + ], + "out_args": [ + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_128_bfp.out25_72" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_128_bfp.out25_72", + "/transformer_blocks.13/Add_output_0.out10_18" + ], + "const_args": [ + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_5_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_3_output_0.out10_19" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.13/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.13/Add_3_output_0.out10_19" + ], + "const_args": [ + "onnx::MatMul_8648_45_0_17" + ], + "out_args": [ + "/controlnet_blocks.13/Add_output_0.out17_4_130" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13" + ], + "const_args": [ + "onnx::MatMul_8356_45_5_27" + ], + "out_args": [ + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27", + "/transformer_blocks.12/Add_7_output_0.out10_17" + ], + "const_args": [ + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_2_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_4_output_0.out10_20" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.13/Add_4_output_0.out10_20" + ], + "const_args": [ + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55", + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55" + ], + "out_args": [ + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55" + ], + "const_args": [ + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_3_transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_4" + ], + "out_args": [ + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21" + ], + "const_args": [ + "onnx::MatMul_8360_45_1_54" + ], + "out_args": [ + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_71" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_71" + ], + "const_args": [ + "onnx::MatMul_8361_45_1_56" + ], + "out_args": [ + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_129_bfp.out25_73" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_129_bfp.out25_73", + "/transformer_blocks.13/Add_4_output_0.out10_20" + ], + "const_args": [ + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_5_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_7_output_0.out10_21" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.13/Add_3_output_0.out10_19" + ], + "const_args": [ + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_56", + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_56" + ], + "out_args": [ + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_56_bfp.out15_56" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_56_bfp.out15_56" + ], + "const_args": [ + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_0_transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_1" + ], + "out_args": [ + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_56_bfp.out1_22" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.13/Add_7_output_0.out10_21" + ], + "const_args": [ + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_57", + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_57" + ], + "out_args": [ + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_57_bfp.out15_57" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_57_bfp.out15_57" + ], + "const_args": [ + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_0_transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_1" + ], + "out_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_57_bfp.out1_24" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_57_bfp.out1_24", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_56_bfp.out1_22" + ], + "const_args": [ + "onnx::MatMul_8377_onnx::MatMul_8362" + ], + "out_args": [ + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_57_bfp.out1_24", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_56_bfp.out1_22" + ], + "const_args": [ + "onnx::MatMul_8378_onnx::MatMul_8363" + ], + "out_args": [ + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_57_bfp.out1_24", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_56_bfp.out1_22" + ], + "const_args": [ + "onnx::MatMul_8379_onnx::MatMul_8364" + ], + "out_args": [ + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/MatMulmha_18_0_14", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42", + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43", + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14" + ], + "const_args": [ + "onnx::MatMul_8388_45_5_28" + ], + "out_args": [ + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28", + "/transformer_blocks.13/Add_3_output_0.out10_19" + ], + "const_args": [ + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_2_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_output_0.out10_22" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.14/Add_output_0.out10_22" + ], + "const_args": [ + "/transformer_blocks.14/norm2/Constant_output_0_14_58", + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58" + ], + "out_args": [ + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58" + ], + "const_args": [ + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_3_transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_4" + ], + "out_args": [ + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23" + ], + "const_args": [ + "onnx::MatMul_8389_45_1_57" + ], + "out_args": [ + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_75" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_75" + ], + "const_args": [ + "onnx::MatMul_8390_45_1_59" + ], + "out_args": [ + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_137_bfp.out25_77" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_137_bfp.out25_77", + "/transformer_blocks.14/Add_output_0.out10_22" + ], + "const_args": [ + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_5_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_3_output_0.out10_23" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.14/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.14/Add_3_output_0.out10_23" + ], + "const_args": [ + "onnx::MatMul_8649_45_0_18" + ], + "out_args": [ + "/controlnet_blocks.14/Add_output_0.out17_4_139" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14" + ], + "const_args": [ + "onnx::MatMul_8387_45_5_29" + ], + "out_args": [ + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29", + "/transformer_blocks.13/Add_7_output_0.out10_21" + ], + "const_args": [ + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_2_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_4_output_0.out10_24" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.14/Add_4_output_0.out10_24" + ], + "const_args": [ + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59", + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59" + ], + "out_args": [ + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59" + ], + "const_args": [ + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_3_transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_4" + ], + "out_args": [ + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25" + ], + "const_args": [ + "onnx::MatMul_8391_45_1_58" + ], + "out_args": [ + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_76" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_76" + ], + "const_args": [ + "onnx::MatMul_8392_45_1_60" + ], + "out_args": [ + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_138_bfp.out25_78" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_138_bfp.out25_78", + "/transformer_blocks.14/Add_4_output_0.out10_24" + ], + "const_args": [ + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_5_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_7_output_0.out10_25" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.14/Add_3_output_0.out10_23" + ], + "const_args": [ + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_60", + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_60" + ], + "out_args": [ + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_60_bfp.out15_60" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_60_bfp.out15_60" + ], + "const_args": [ + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_0_transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_1" + ], + "out_args": [ + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_60_bfp.out1_26" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.14/Add_7_output_0.out10_25" + ], + "const_args": [ + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_61", + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_61" + ], + "out_args": [ + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_61_bfp.out15_61" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_61_bfp.out15_61" + ], + "const_args": [ + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_0_transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_1" + ], + "out_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_61_bfp.out1_28" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_61_bfp.out1_28", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_60_bfp.out1_26" + ], + "const_args": [ + "onnx::MatMul_8408_onnx::MatMul_8393" + ], + "out_args": [ + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_61_bfp.out1_28", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_60_bfp.out1_26" + ], + "const_args": [ + "onnx::MatMul_8409_onnx::MatMul_8394" + ], + "out_args": [ + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_61_bfp.out1_28", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_60_bfp.out1_26" + ], + "const_args": [ + "onnx::MatMul_8410_onnx::MatMul_8395" + ], + "out_args": [ + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/MatMulmha_18_0_15", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45", + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46", + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15" + ], + "const_args": [ + "onnx::MatMul_8419_45_5_30" + ], + "out_args": [ + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30", + "/transformer_blocks.14/Add_3_output_0.out10_23" + ], + "const_args": [ + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_2_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_output_0.out10_26" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.15/Add_output_0.out10_26" + ], + "const_args": [ + "/transformer_blocks.15/norm2/Constant_output_0_14_62", + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62" + ], + "out_args": [ + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62" + ], + "const_args": [ + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_3_transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_4" + ], + "out_args": [ + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27" + ], + "const_args": [ + "onnx::MatMul_8420_45_1_61" + ], + "out_args": [ + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_80" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_80" + ], + "const_args": [ + "onnx::MatMul_8421_45_1_63" + ], + "out_args": [ + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_146_bfp.out25_82" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_146_bfp.out25_82", + "/transformer_blocks.15/Add_output_0.out10_26" + ], + "const_args": [ + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_5_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_3_output_0.out10_27" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.15/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.15/Add_3_output_0.out10_27" + ], + "const_args": [ + "onnx::MatMul_8650_45_0_19" + ], + "out_args": [ + "/controlnet_blocks.15/Add_output_0.out17_4_148" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15" + ], + "const_args": [ + "onnx::MatMul_8418_45_5_31" + ], + "out_args": [ + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31", + "/transformer_blocks.14/Add_7_output_0.out10_25" + ], + "const_args": [ + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_2_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_4_output_0.out10_28" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.15/Add_4_output_0.out10_28" + ], + "const_args": [ + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63", + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63" + ], + "out_args": [ + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63" + ], + "const_args": [ + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_3_transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_4" + ], + "out_args": [ + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29" + ], + "const_args": [ + "onnx::MatMul_8422_45_1_62" + ], + "out_args": [ + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_81" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_81" + ], + "const_args": [ + "onnx::MatMul_8423_45_1_64" + ], + "out_args": [ + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_147_bfp.out25_83" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_147_bfp.out25_83", + "/transformer_blocks.15/Add_4_output_0.out10_28" + ], + "const_args": [ + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_5_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_7_output_0.out10_29" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.15/Add_3_output_0.out10_27" + ], + "const_args": [ + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_64", + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_64" + ], + "out_args": [ + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_64_bfp.out15_64" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_64_bfp.out15_64" + ], + "const_args": [ + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_0_transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_1" + ], + "out_args": [ + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_64_bfp.out1_30" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.15/Add_7_output_0.out10_29" + ], + "const_args": [ + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_65", + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_65" + ], + "out_args": [ + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_65_bfp.out15_65" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_65_bfp.out15_65" + ], + "const_args": [ + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_0_transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_1" + ], + "out_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_65_bfp.out1_32" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_65_bfp.out1_32", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_64_bfp.out1_30" + ], + "const_args": [ + "onnx::MatMul_8439_onnx::MatMul_8424" + ], + "out_args": [ + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_65_bfp.out1_32", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_64_bfp.out1_30" + ], + "const_args": [ + "onnx::MatMul_8440_onnx::MatMul_8425" + ], + "out_args": [ + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_65_bfp.out1_32", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_64_bfp.out1_30" + ], + "const_args": [ + "onnx::MatMul_8441_onnx::MatMul_8426" + ], + "out_args": [ + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/MatMulmha_18_0_16", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48", + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49", + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16" + ], + "const_args": [ + "onnx::MatMul_8450_45_5_32" + ], + "out_args": [ + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32", + "/transformer_blocks.15/Add_3_output_0.out10_27" + ], + "const_args": [ + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_2_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_output_0.out10_30" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.16/Add_output_0.out10_30" + ], + "const_args": [ + "/transformer_blocks.16/norm2/Constant_output_0_14_66", + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66" + ], + "out_args": [ + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66" + ], + "const_args": [ + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_3_transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_4" + ], + "out_args": [ + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31" + ], + "const_args": [ + "onnx::MatMul_8451_45_1_65" + ], + "out_args": [ + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_85" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_85" + ], + "const_args": [ + "onnx::MatMul_8452_45_1_67" + ], + "out_args": [ + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_155_bfp.out25_87" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_155_bfp.out25_87", + "/transformer_blocks.16/Add_output_0.out10_30" + ], + "const_args": [ + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_5_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_3_output_0.out10_31" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.16/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.16/Add_3_output_0.out10_31" + ], + "const_args": [ + "onnx::MatMul_8651_45_0_20" + ], + "out_args": [ + "/controlnet_blocks.16/Add_output_0.out17_4_157" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16" + ], + "const_args": [ + "onnx::MatMul_8449_45_5_33" + ], + "out_args": [ + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33", + "/transformer_blocks.15/Add_7_output_0.out10_29" + ], + "const_args": [ + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_2_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_4_output_0.out10_32" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.16/Add_4_output_0.out10_32" + ], + "const_args": [ + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67", + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67" + ], + "out_args": [ + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67" + ], + "const_args": [ + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_3_transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_4" + ], + "out_args": [ + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33" + ], + "const_args": [ + "onnx::MatMul_8453_45_1_66" + ], + "out_args": [ + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_86" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_86" + ], + "const_args": [ + "onnx::MatMul_8454_45_1_68" + ], + "out_args": [ + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_88" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_88", + "/transformer_blocks.16/Add_4_output_0.out10_32" + ], + "const_args": [ + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_5_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_7_output_0.out10_33" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.16/Add_3_output_0.out10_31" + ], + "const_args": [ + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_68", + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_68" + ], + "out_args": [ + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_68_bfp.out15_68" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_68_bfp.out15_68" + ], + "const_args": [ + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_0_transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_1" + ], + "out_args": [ + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_68_bfp.out1_34" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.16/Add_7_output_0.out10_33" + ], + "const_args": [ + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_69", + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_69" + ], + "out_args": [ + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_69_bfp.out15_69" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_69_bfp.out15_69" + ], + "const_args": [ + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_0_transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_1" + ], + "out_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_69_bfp.out1_36" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_69_bfp.out1_36", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_68_bfp.out1_34" + ], + "const_args": [ + "onnx::MatMul_8470_onnx::MatMul_8455" + ], + "out_args": [ + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_69_bfp.out1_36", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_68_bfp.out1_34" + ], + "const_args": [ + "onnx::MatMul_8471_onnx::MatMul_8456" + ], + "out_args": [ + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_69_bfp.out1_36", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_68_bfp.out1_34" + ], + "const_args": [ + "onnx::MatMul_8472_onnx::MatMul_8457" + ], + "out_args": [ + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/MatMulmha_18_0_17", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51", + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52", + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17" + ], + "const_args": [ + "onnx::MatMul_8481_45_5_34" + ], + "out_args": [ + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34", + "/transformer_blocks.16/Add_3_output_0.out10_31" + ], + "const_args": [ + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_2_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_output_0.out10_34" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.17/Add_output_0.out10_34" + ], + "const_args": [ + "/transformer_blocks.17/norm2/Constant_output_0_14_70", + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70" + ], + "out_args": [ + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70" + ], + "const_args": [ + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_3_transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_4" + ], + "out_args": [ + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35" + ], + "const_args": [ + "onnx::MatMul_8482_45_1_69" + ], + "out_args": [ + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_90" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_90" + ], + "const_args": [ + "onnx::MatMul_8483_45_1_71" + ], + "out_args": [ + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_164_bfp.out25_92" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_164_bfp.out25_92", + "/transformer_blocks.17/Add_output_0.out10_34" + ], + "const_args": [ + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_5_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_3_output_0.out10_35" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.17/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.17/Add_3_output_0.out10_35" + ], + "const_args": [ + "onnx::MatMul_8652_45_0_21" + ], + "out_args": [ + "/controlnet_blocks.17/Add_output_0.out17_4_166" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17" + ], + "const_args": [ + "onnx::MatMul_8480_45_5_35" + ], + "out_args": [ + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35", + "/transformer_blocks.16/Add_7_output_0.out10_33" + ], + "const_args": [ + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_2_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_4_output_0.out10_36" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.17/Add_4_output_0.out10_36" + ], + "const_args": [ + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71", + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71" + ], + "out_args": [ + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71" + ], + "const_args": [ + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_3_transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_4" + ], + "out_args": [ + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37" + ], + "const_args": [ + "onnx::MatMul_8484_45_1_70" + ], + "out_args": [ + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_91" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_91" + ], + "const_args": [ + "onnx::MatMul_8485_45_1_72" + ], + "out_args": [ + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_165_bfp.out25_93" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_165_bfp.out25_93", + "/transformer_blocks.17/Add_4_output_0.out10_36" + ], + "const_args": [ + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_5_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_7_output_0.out10_37" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.17/Add_3_output_0.out10_35" + ], + "const_args": [ + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_72", + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_72" + ], + "out_args": [ + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_72_bfp.out15_72" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_72_bfp.out15_72" + ], + "const_args": [ + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_0_transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_1" + ], + "out_args": [ + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_72_bfp.out1_38" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.17/Add_7_output_0.out10_37" + ], + "const_args": [ + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_73", + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_73" + ], + "out_args": [ + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_73_bfp.out15_73" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_73_bfp.out15_73" + ], + "const_args": [ + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_0_transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_1" + ], + "out_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_73_bfp.out1_40" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_73_bfp.out1_40", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_72_bfp.out1_38" + ], + "const_args": [ + "onnx::MatMul_8501_onnx::MatMul_8486" + ], + "out_args": [ + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_73_bfp.out1_40", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_72_bfp.out1_38" + ], + "const_args": [ + "onnx::MatMul_8502_onnx::MatMul_8487" + ], + "out_args": [ + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_73_bfp.out1_40", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_72_bfp.out1_38" + ], + "const_args": [ + "onnx::MatMul_8503_onnx::MatMul_8488" + ], + "out_args": [ + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/MatMulmha_18_0_18", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54", + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55", + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18" + ], + "const_args": [ + "onnx::MatMul_8512_45_5_36" + ], + "out_args": [ + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36", + "/transformer_blocks.17/Add_3_output_0.out10_35" + ], + "const_args": [ + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_2_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_output_0.out10_38" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.18/Add_output_0.out10_38" + ], + "const_args": [ + "/transformer_blocks.18/norm2/Constant_output_0_14_74", + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74" + ], + "out_args": [ + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74" + ], + "const_args": [ + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_3_transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_4" + ], + "out_args": [ + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39" + ], + "const_args": [ + "onnx::MatMul_8513_45_1_73" + ], + "out_args": [ + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_95" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_95" + ], + "const_args": [ + "onnx::MatMul_8514_45_1_75" + ], + "out_args": [ + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_173_bfp.out25_97" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_173_bfp.out25_97", + "/transformer_blocks.18/Add_output_0.out10_38" + ], + "const_args": [ + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_5_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_3_output_0.out10_39" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.18/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.18/Add_3_output_0.out10_39" + ], + "const_args": [ + "onnx::MatMul_8653_45_0_22" + ], + "out_args": [ + "/controlnet_blocks.18/Add_output_0.out17_4_175" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18" + ], + "const_args": [ + "onnx::MatMul_8511_45_5_37" + ], + "out_args": [ + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37", + "/transformer_blocks.17/Add_7_output_0.out10_37" + ], + "const_args": [ + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_2_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_4_output_0.out10_40" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.18/Add_4_output_0.out10_40" + ], + "const_args": [ + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75", + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75" + ], + "out_args": [ + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75" + ], + "const_args": [ + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_3_transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_4" + ], + "out_args": [ + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41" + ], + "const_args": [ + "onnx::MatMul_8515_45_1_74" + ], + "out_args": [ + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_96" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_96" + ], + "const_args": [ + "onnx::MatMul_8516_45_1_76" + ], + "out_args": [ + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_174_bfp.out25_98" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_174_bfp.out25_98", + "/transformer_blocks.18/Add_4_output_0.out10_40" + ], + "const_args": [ + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_5_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_7_output_0.out10_41" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.18/Add_3_output_0.out10_39" + ], + "const_args": [ + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_76", + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_76" + ], + "out_args": [ + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_76_bfp.out15_76" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_76_bfp.out15_76" + ], + "const_args": [ + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_0_transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_1" + ], + "out_args": [ + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_76_bfp.out1_42" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.18/Add_7_output_0.out10_41" + ], + "const_args": [ + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_77", + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_77" + ], + "out_args": [ + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_77_bfp.out15_77" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_77_bfp.out15_77" + ], + "const_args": [ + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_0_transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_1" + ], + "out_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_77_bfp.out1_44" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_77_bfp.out1_44", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_76_bfp.out1_42" + ], + "const_args": [ + "onnx::MatMul_8532_onnx::MatMul_8517" + ], + "out_args": [ + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_77_bfp.out1_44", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_76_bfp.out1_42" + ], + "const_args": [ + "onnx::MatMul_8533_onnx::MatMul_8518" + ], + "out_args": [ + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_77_bfp.out1_44", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_76_bfp.out1_42" + ], + "const_args": [ + "onnx::MatMul_8534_onnx::MatMul_8519" + ], + "out_args": [ + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/MatMulmha_18_0_19", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57", + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58", + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19" + ], + "const_args": [ + "onnx::MatMul_8543_45_5_38" + ], + "out_args": [ + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38", + "/transformer_blocks.18/Add_3_output_0.out10_39" + ], + "const_args": [ + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_2_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_output_0.out10_42" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.19/Add_output_0.out10_42" + ], + "const_args": [ + "/transformer_blocks.19/norm2/Constant_output_0_14_78", + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78" + ], + "out_args": [ + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78" + ], + "const_args": [ + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_3_transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_4" + ], + "out_args": [ + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43" + ], + "const_args": [ + "onnx::MatMul_8544_45_1_77" + ], + "out_args": [ + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_100" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_100" + ], + "const_args": [ + "onnx::MatMul_8545_45_1_79" + ], + "out_args": [ + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_182_bfp.out25_102" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_182_bfp.out25_102", + "/transformer_blocks.19/Add_output_0.out10_42" + ], + "const_args": [ + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_5_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_3_output_0.out10_43" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.19/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.19/Add_3_output_0.out10_43" + ], + "const_args": [ + "onnx::MatMul_8654_45_0_23" + ], + "out_args": [ + "/controlnet_blocks.19/Add_output_0.out17_4_184" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19" + ], + "const_args": [ + "onnx::MatMul_8542_45_5_39" + ], + "out_args": [ + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39", + "/transformer_blocks.18/Add_7_output_0.out10_41" + ], + "const_args": [ + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_2_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_4_output_0.out10_44" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.19/Add_4_output_0.out10_44" + ], + "const_args": [ + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79", + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79" + ], + "out_args": [ + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79" + ], + "const_args": [ + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_3_transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_4" + ], + "out_args": [ + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45" + ], + "const_args": [ + "onnx::MatMul_8546_45_1_78" + ], + "out_args": [ + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_101" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_101" + ], + "const_args": [ + "onnx::MatMul_8547_45_1_80" + ], + "out_args": [ + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_183_bfp.out25_103" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_183_bfp.out25_103", + "/transformer_blocks.19/Add_4_output_0.out10_44" + ], + "const_args": [ + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_5_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_7_output_0.out10_45" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.19/Add_3_output_0.out10_43" + ], + "const_args": [ + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_80", + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_80" + ], + "out_args": [ + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_80_bfp.out15_80" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.19/Add_7_output_0.out10_45" + ], + "const_args": [ + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_81", + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_81" + ], + "out_args": [ + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_81_bfp.out15_81" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_80_bfp.out15_80" + ], + "const_args": [ + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_0_transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_1" + ], + "out_args": [ + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_80_bfp.out1_50" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_81_bfp.out15_81" + ], + "const_args": [ + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_0_transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_1" + ], + "out_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_81_bfp.out1_52" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_81_bfp.out1_52", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_80_bfp.out1_50" + ], + "const_args": [ + "onnx::MatMul_8563_onnx::MatMul_8548" + ], + "out_args": [ + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_81_bfp.out1_52", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_80_bfp.out1_50" + ], + "const_args": [ + "onnx::MatMul_8564_onnx::MatMul_8549" + ], + "out_args": [ + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_81_bfp.out1_52", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_80_bfp.out1_50" + ], + "const_args": [ + "onnx::MatMul_8565_onnx::MatMul_8550" + ], + "out_args": [ + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/MatMulmha_18_0_20", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60", + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61", + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20" + ], + "const_args": [ + "onnx::MatMul_8574_45_5_40" + ], + "out_args": [ + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40", + "/transformer_blocks.19/Add_3_output_0.out10_43" + ], + "const_args": [ + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_2_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_output_0.out10_50" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.20/Add_output_0.out10_50" + ], + "const_args": [ + "/transformer_blocks.20/norm2/Constant_output_0_14_82", + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82" + ], + "out_args": [ + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82" + ], + "const_args": [ + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_3_transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_4" + ], + "out_args": [ + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51" + ], + "const_args": [ + "onnx::MatMul_8575_45_1_81" + ], + "out_args": [ + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_105" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_105" + ], + "const_args": [ + "onnx::MatMul_8576_45_1_83" + ], + "out_args": [ + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_191_bfp.out25_107" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_191_bfp.out25_107", + "/transformer_blocks.20/Add_output_0.out10_50" + ], + "const_args": [ + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_5_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_3_output_0.out10_51" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.20/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.20/Add_3_output_0.out10_51" + ], + "const_args": [ + "onnx::MatMul_8655_45_0_24" + ], + "out_args": [ + "/controlnet_blocks.20/Add_output_0.out17_4_193" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20" + ], + "const_args": [ + "onnx::MatMul_8573_45_5_41" + ], + "out_args": [ + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41", + "/transformer_blocks.19/Add_7_output_0.out10_45" + ], + "const_args": [ + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_2_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_4_output_0.out10_52" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.20/Add_4_output_0.out10_52" + ], + "const_args": [ + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83", + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83" + ], + "out_args": [ + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83" + ], + "const_args": [ + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_3_transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_4" + ], + "out_args": [ + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53" + ], + "const_args": [ + "onnx::MatMul_8577_45_1_82" + ], + "out_args": [ + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_106" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_106" + ], + "const_args": [ + "onnx::MatMul_8578_45_1_84" + ], + "out_args": [ + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_192_bfp.out25_108" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_192_bfp.out25_108", + "/transformer_blocks.20/Add_4_output_0.out10_52" + ], + "const_args": [ + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_5_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_7_output_0.out10_53" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.20/Add_3_output_0.out10_51" + ], + "const_args": [ + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_84", + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_84" + ], + "out_args": [ + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_84_bfp.out15_84" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_84_bfp.out15_84" + ], + "const_args": [ + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_0_transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_1" + ], + "out_args": [ + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_84_bfp.out1_54" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.20/Add_7_output_0.out10_53" + ], + "const_args": [ + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_85", + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_85" + ], + "out_args": [ + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_85_bfp.out15_85" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_85_bfp.out15_85" + ], + "const_args": [ + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_0_transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_1" + ], + "out_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_85_bfp.out1_56" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_85_bfp.out1_56", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_84_bfp.out1_54" + ], + "const_args": [ + "onnx::MatMul_8594_onnx::MatMul_8579" + ], + "out_args": [ + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_85_bfp.out1_56", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_84_bfp.out1_54" + ], + "const_args": [ + "onnx::MatMul_8595_onnx::MatMul_8580" + ], + "out_args": [ + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_85_bfp.out1_56", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_84_bfp.out1_54" + ], + "const_args": [ + "onnx::MatMul_8596_onnx::MatMul_8581" + ], + "out_args": [ + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/MatMulmha_18_0_21", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63", + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64", + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21" + ], + "const_args": [ + "onnx::MatMul_8605_45_5_42" + ], + "out_args": [ + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42", + "/transformer_blocks.20/Add_3_output_0.out10_51" + ], + "const_args": [ + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_2_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_output_0.out10_54" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.21/Add_output_0.out10_54" + ], + "const_args": [ + "/transformer_blocks.21/norm2/Constant_output_0_14_86", + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86" + ], + "out_args": [ + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86" + ], + "const_args": [ + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_3_transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_4" + ], + "out_args": [ + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55" + ], + "const_args": [ + "onnx::MatMul_8606_45_1_85" + ], + "out_args": [ + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_110" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_110" + ], + "const_args": [ + "onnx::MatMul_8607_45_1_87" + ], + "out_args": [ + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_200_bfp.out25_112" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_200_bfp.out25_112", + "/transformer_blocks.21/Add_output_0.out10_54" + ], + "const_args": [ + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_5_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_3_output_0.out10_55" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.21/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.21/Add_3_output_0.out10_55" + ], + "const_args": [ + "onnx::MatMul_8656_45_0_25" + ], + "out_args": [ + "/controlnet_blocks.21/Add_output_0.out17_4_202" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21" + ], + "const_args": [ + "onnx::MatMul_8604_45_5_43" + ], + "out_args": [ + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43", + "/transformer_blocks.20/Add_7_output_0.out10_53" + ], + "const_args": [ + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_2_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_4_output_0.out10_56" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.21/Add_4_output_0.out10_56" + ], + "const_args": [ + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87", + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87" + ], + "out_args": [ + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87" + ], + "const_args": [ + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_3_transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_4" + ], + "out_args": [ + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57" + ], + "const_args": [ + "onnx::MatMul_8608_45_1_86" + ], + "out_args": [ + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_111" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_111" + ], + "const_args": [ + "onnx::MatMul_8609_45_1_88" + ], + "out_args": [ + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_201_bfp.out25_113" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_201_bfp.out25_113", + "/transformer_blocks.21/Add_4_output_0.out10_56" + ], + "const_args": [ + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_5_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_7_output_0.out10_57" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.21/Add_3_output_0.out10_55" + ], + "const_args": [ + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_88", + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_88" + ], + "out_args": [ + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_88_bfp.out15_88" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_88_bfp.out15_88" + ], + "const_args": [ + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_0_transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_1" + ], + "out_args": [ + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_88_bfp.out1_58" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.21/Add_7_output_0.out10_57" + ], + "const_args": [ + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_89", + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_89" + ], + "out_args": [ + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_89_bfp.out15_89" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_89_bfp.out15_89" + ], + "const_args": [ + "transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_0_transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_1" + ], + "out_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_89_bfp.out1_60" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_89_bfp.out1_60", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_88_bfp.out1_58" + ], + "const_args": [ + "onnx::MatMul_8625_onnx::MatMul_8610" + ], + "out_args": [ + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_89_bfp.out1_60", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_88_bfp.out1_58" + ], + "const_args": [ + "onnx::MatMul_8626_onnx::MatMul_8611" + ], + "out_args": [ + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_89_bfp.out1_60", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_88_bfp.out1_58" + ], + "const_args": [ + "onnx::MatMul_8627_onnx::MatMul_8612" + ], + "out_args": [ + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/MatMulmha_18_0_22", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66", + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67", + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)+max_length", + "64", + "floor(h/2)*floor(w/2)+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22" + ], + "const_args": [ + "onnx::MatMul_8632_45_5_44" + ], + "out_args": [ + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "floor(h/2)*floor(w/2)" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44", + "/transformer_blocks.21/Add_3_output_0.out10_55" + ], + "const_args": [ + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_2_gma" + ], + "out_args": [ + "/transformer_blocks.22/Add_output_0.out10_58" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.22/Add_output_0.out10_58" + ], + "const_args": [ + "/transformer_blocks.22/norm2/Constant_output_0_14_90", + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90" + ], + "out_args": [ + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90" + ], + "const_args": [ + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_3_transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_4" + ], + "out_args": [ + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59" + ], + "const_args": [ + "onnx::MatMul_8633_45_1_89" + ], + "out_args": [ + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_115" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_115" + ], + "const_args": [ + "onnx::MatMul_8634_45_1_90" + ], + "out_args": [ + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_209_bfp.out25_116" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_209_bfp.out25_116", + "/transformer_blocks.22/Add_output_0.out10_58" + ], + "const_args": [ + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_5_gma" + ], + "out_args": [ + "/transformer_blocks.22/Add_3_output_0.out10_59" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/controlnet_blocks.22/MatMul", + "type": "SDGemm", + "in_args": [ + "/transformer_blocks.22/Add_3_output_0.out10_59" + ], + "const_args": [ + "onnx::MatMul_8657_45_0_26" + ], + "out_args": [ + "/controlnet_blocks.22/Add_output_0.out17_4_210" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 15952, + "xrt_arg_id": 0, + "packed_tensors": [ + "hidden_states_nhwc.out5_0_0", + "/pos_embed/Reshape_1_output_0.out_35_1_2", + "controlnet_cond_nhwc.out5_0_1", + "/time_text_embed/Cast_output_0.out17_4_3", + "pooled_projections.out17_4_1", + "encoder_hidden_states.out17_4_0" + ] + }, + "out": { + "buffer_size": 70656, + "xrt_arg_id": 1, + "packed_tensors": [ + "/controlnet_blocks.0/Add_output_0.out17_4_13", + "/controlnet_blocks.1/Add_output_0.out17_4_22", + "/controlnet_blocks.2/Add_output_0.out17_4_31", + "/controlnet_blocks.3/Add_output_0.out17_4_40", + "/controlnet_blocks.4/Add_output_0.out17_4_49", + "/controlnet_blocks.5/Add_output_0.out17_4_58", + "/controlnet_blocks.6/Add_output_0.out17_4_67", + "/controlnet_blocks.7/Add_output_0.out17_4_76", + "/controlnet_blocks.8/Add_output_0.out17_4_85", + "/controlnet_blocks.9/Add_output_0.out17_4_94", + "/controlnet_blocks.10/Add_output_0.out17_4_103", + "/controlnet_blocks.11/Add_output_0.out17_4_112", + "/controlnet_blocks.12/Add_output_0.out17_4_121", + "/controlnet_blocks.13/Add_output_0.out17_4_130", + "/controlnet_blocks.14/Add_output_0.out17_4_139", + "/controlnet_blocks.15/Add_output_0.out17_4_148", + "/controlnet_blocks.16/Add_output_0.out17_4_157", + "/controlnet_blocks.17/Add_output_0.out17_4_166", + "/controlnet_blocks.18/Add_output_0.out17_4_175", + "/controlnet_blocks.19/Add_output_0.out17_4_184", + "/controlnet_blocks.20/Add_output_0.out17_4_193", + "/controlnet_blocks.21/Add_output_0.out17_4_202", + "/controlnet_blocks.22/Add_output_0.out17_4_210" + ] + }, + "scratch": { + "buffer_size": 1157632, + "xrt_arg_id": 2, + "packed_tensors": [ + "/pos_embed/Transpose_output_0.out5_0_0", + "/pos_embed/Add_2_output_0.out_35_1_2", + "/pos_embed_input/Transpose_output_0.out5_0_1", + "/Add_output_0.out_35_1_3", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3", + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1", + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2", + "/time_text_embed/Add_output_0.out_35_1_4", + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "encoder_hidden_states.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0", + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0", + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_90", + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0", + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1", + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2", + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0", + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", + "/transformer_blocks.0/Add_4_output_0.out10_0", + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3", + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1", + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", + "/transformer_blocks.0/Add_7_output_0.out10_1", + "/transformer_blocks.0/Add_output_0.out10_88", + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2", + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_89", + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", + "/transformer_blocks.0/Add_3_output_0.out10_89", + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2", + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5", + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_5_bfp.out1_4", + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3", + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4", + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5", + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1", + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2", + "/transformer_blocks.1/Add_output_0.out10_2", + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6", + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3", + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", + "/transformer_blocks.1/Add_3_output_0.out10_3", + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", + "/transformer_blocks.1/Add_4_output_0.out10_4", + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7", + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5", + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", + "/transformer_blocks.1/Add_7_output_0.out10_5", + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8", + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_46", + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_9_bfp.out1_48", + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6", + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7", + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8", + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2", + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5", + "/transformer_blocks.2/Add_4_output_0.out10_48", + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4", + "/transformer_blocks.2/Add_output_0.out10_46", + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10", + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47", + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", + "/transformer_blocks.2/Add_3_output_0.out10_47", + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11", + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49", + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", + "/transformer_blocks.2/Add_7_output_0.out10_49", + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12", + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_61", + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_13_bfp.out1_63", + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9", + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10", + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11", + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3", + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7", + "/transformer_blocks.3/Add_4_output_0.out10_62", + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6", + "/transformer_blocks.3/Add_output_0.out10_60", + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14", + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_62", + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", + "/transformer_blocks.3/Add_3_output_0.out10_61", + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15", + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_64", + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", + "/transformer_blocks.3/Add_7_output_0.out10_63", + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16", + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_65", + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_17_bfp.out1_67", + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12", + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13", + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14", + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4", + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9", + "/transformer_blocks.4/Add_4_output_0.out10_66", + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8", + "/transformer_blocks.4/Add_output_0.out10_64", + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18", + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_66", + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", + "/transformer_blocks.4/Add_3_output_0.out10_65", + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19", + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_68", + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", + "/transformer_blocks.4/Add_7_output_0.out10_67", + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20", + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_69", + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_21_bfp.out1_71", + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15", + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16", + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17", + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5", + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11", + "/transformer_blocks.5/Add_4_output_0.out10_70", + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10", + "/transformer_blocks.5/Add_output_0.out10_68", + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22", + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_70", + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32", + "/transformer_blocks.5/Add_3_output_0.out10_69", + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23", + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_72", + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_31", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33", + "/transformer_blocks.5/Add_7_output_0.out10_71", + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_24_bfp.out15_24", + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_25_bfp.out15_25", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_24_bfp.out1_73", + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_25_bfp.out1_75", + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18", + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19", + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20", + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6", + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13", + "/transformer_blocks.6/Add_4_output_0.out10_74", + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12", + "/transformer_blocks.6/Add_output_0.out10_72", + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26", + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_74", + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_35", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37", + "/transformer_blocks.6/Add_3_output_0.out10_73", + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27", + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_76", + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_36", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38", + "/transformer_blocks.6/Add_7_output_0.out10_75", + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_28_bfp.out15_28", + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_29_bfp.out15_29", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_28_bfp.out1_77", + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_29_bfp.out1_79", + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21", + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22", + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23", + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7", + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15", + "/transformer_blocks.7/Add_4_output_0.out10_78", + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14", + "/transformer_blocks.7/Add_output_0.out10_76", + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30", + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_78", + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_40", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42", + "/transformer_blocks.7/Add_3_output_0.out10_77", + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31", + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_80", + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_41", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43", + "/transformer_blocks.7/Add_7_output_0.out10_79", + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_32_bfp.out15_32", + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_33_bfp.out15_33", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_32_bfp.out1_81", + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_33_bfp.out1_83", + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24", + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25", + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26", + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8", + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17", + "/transformer_blocks.8/Add_4_output_0.out10_82", + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16", + "/transformer_blocks.8/Add_output_0.out10_80", + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34", + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_82", + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_45", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47", + "/transformer_blocks.8/Add_3_output_0.out10_81", + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35", + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_84", + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_46", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48", + "/transformer_blocks.8/Add_7_output_0.out10_83", + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_36_bfp.out15_36", + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_37_bfp.out15_37", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_36_bfp.out1_85", + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_37_bfp.out1_87", + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27", + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28", + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29", + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9", + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19", + "/transformer_blocks.9/Add_4_output_0.out10_86", + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18", + "/transformer_blocks.9/Add_output_0.out10_84", + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38", + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_86", + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_50", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52", + "/transformer_blocks.9/Add_3_output_0.out10_85", + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6", + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39", + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_88", + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_51", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53", + "/transformer_blocks.9/Add_7_output_0.out10_87", + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41", + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30", + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31", + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32", + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10", + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20", + "/transformer_blocks.10/Add_output_0.out10_6", + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42", + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7", + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_55", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57", + "/transformer_blocks.10/Add_3_output_0.out10_7", + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21", + "/transformer_blocks.10/Add_4_output_0.out10_8", + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43", + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9", + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_56", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58", + "/transformer_blocks.10/Add_7_output_0.out10_9", + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_44_bfp.out15_44", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_44_bfp.out1_10", + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_45_bfp.out15_45", + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_45_bfp.out1_12", + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33", + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34", + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35", + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11", + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22", + "/transformer_blocks.11/Add_output_0.out10_10", + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46", + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11", + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_60", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_62", + "/transformer_blocks.11/Add_3_output_0.out10_11", + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23", + "/transformer_blocks.11/Add_4_output_0.out10_12", + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47", + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13", + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_61", + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_111_bfp.out25_63", + "/transformer_blocks.11/Add_7_output_0.out10_13", + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_48_bfp.out15_48", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_48_bfp.out1_14", + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_49_bfp.out15_49", + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_49_bfp.out1_16", + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36", + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37", + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38", + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12", + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24", + "/transformer_blocks.12/Add_output_0.out10_14", + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50", + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15", + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_65", + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_119_bfp.out25_67", + "/transformer_blocks.12/Add_3_output_0.out10_15", + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25", + "/transformer_blocks.12/Add_4_output_0.out10_16", + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51", + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17", + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_66", + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_120_bfp.out25_68", + "/transformer_blocks.12/Add_7_output_0.out10_17", + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_52_bfp.out15_52", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_52_bfp.out1_18", + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_53_bfp.out15_53", + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_53_bfp.out1_20", + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39", + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40", + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41", + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13", + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26", + "/transformer_blocks.13/Add_output_0.out10_18", + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54", + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19", + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_70", + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_128_bfp.out25_72", + "/transformer_blocks.13/Add_3_output_0.out10_19", + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27", + "/transformer_blocks.13/Add_4_output_0.out10_20", + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55", + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21", + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_71", + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_129_bfp.out25_73", + "/transformer_blocks.13/Add_7_output_0.out10_21", + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_56_bfp.out15_56", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_56_bfp.out1_22", + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_57_bfp.out15_57", + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_57_bfp.out1_24", + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42", + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43", + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44", + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14", + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28", + "/transformer_blocks.14/Add_output_0.out10_22", + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58", + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23", + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_75", + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_137_bfp.out25_77", + "/transformer_blocks.14/Add_3_output_0.out10_23", + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29", + "/transformer_blocks.14/Add_4_output_0.out10_24", + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59", + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25", + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_76", + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_138_bfp.out25_78", + "/transformer_blocks.14/Add_7_output_0.out10_25", + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_60_bfp.out15_60", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_60_bfp.out1_26", + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_61_bfp.out15_61", + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_61_bfp.out1_28", + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45", + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46", + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47", + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15", + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30", + "/transformer_blocks.15/Add_output_0.out10_26", + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62", + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27", + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_80", + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_146_bfp.out25_82", + "/transformer_blocks.15/Add_3_output_0.out10_27", + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31", + "/transformer_blocks.15/Add_4_output_0.out10_28", + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63", + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29", + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_81", + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_147_bfp.out25_83", + "/transformer_blocks.15/Add_7_output_0.out10_29", + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_64_bfp.out15_64", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_64_bfp.out1_30", + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_65_bfp.out15_65", + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_65_bfp.out1_32", + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48", + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49", + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50", + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16", + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32", + "/transformer_blocks.16/Add_output_0.out10_30", + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66", + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31", + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_85", + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_155_bfp.out25_87", + "/transformer_blocks.16/Add_3_output_0.out10_31", + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33", + "/transformer_blocks.16/Add_4_output_0.out10_32", + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67", + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33", + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_86", + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_88", + "/transformer_blocks.16/Add_7_output_0.out10_33", + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_68_bfp.out15_68", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_68_bfp.out1_34", + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_69_bfp.out15_69", + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_69_bfp.out1_36", + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51", + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52", + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53", + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17", + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34", + "/transformer_blocks.17/Add_output_0.out10_34", + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70", + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35", + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_90", + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_164_bfp.out25_92", + "/transformer_blocks.17/Add_3_output_0.out10_35", + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35", + "/transformer_blocks.17/Add_4_output_0.out10_36", + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71", + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37", + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_91", + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_165_bfp.out25_93", + "/transformer_blocks.17/Add_7_output_0.out10_37", + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_72_bfp.out15_72", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_72_bfp.out1_38", + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_73_bfp.out15_73", + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_73_bfp.out1_40", + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54", + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55", + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56", + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18", + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36", + "/transformer_blocks.18/Add_output_0.out10_38", + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74", + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39", + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_95", + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_173_bfp.out25_97", + "/transformer_blocks.18/Add_3_output_0.out10_39", + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37", + "/transformer_blocks.18/Add_4_output_0.out10_40", + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75", + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41", + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_96", + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_174_bfp.out25_98", + "/transformer_blocks.18/Add_7_output_0.out10_41", + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_76_bfp.out15_76", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_76_bfp.out1_42", + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_77_bfp.out15_77", + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_77_bfp.out1_44", + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57", + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58", + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59", + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19", + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38", + "/transformer_blocks.19/Add_output_0.out10_42", + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78", + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43", + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_100", + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_182_bfp.out25_102", + "/transformer_blocks.19/Add_3_output_0.out10_43", + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39", + "/transformer_blocks.19/Add_4_output_0.out10_44", + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79", + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45", + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_101", + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_183_bfp.out25_103", + "/transformer_blocks.19/Add_7_output_0.out10_45", + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_80_bfp.out15_80", + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_81_bfp.out15_81", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_80_bfp.out1_50", + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_81_bfp.out1_52", + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60", + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61", + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62", + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20", + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40", + "/transformer_blocks.20/Add_output_0.out10_50", + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82", + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51", + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_105", + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_191_bfp.out25_107", + "/transformer_blocks.20/Add_3_output_0.out10_51", + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41", + "/transformer_blocks.20/Add_4_output_0.out10_52", + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83", + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53", + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_106", + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_192_bfp.out25_108", + "/transformer_blocks.20/Add_7_output_0.out10_53", + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_84_bfp.out15_84", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_84_bfp.out1_54", + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_85_bfp.out15_85", + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_85_bfp.out1_56", + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63", + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64", + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65", + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21", + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42", + "/transformer_blocks.21/Add_output_0.out10_54", + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86", + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55", + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_110", + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_200_bfp.out25_112", + "/transformer_blocks.21/Add_3_output_0.out10_55", + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43", + "/transformer_blocks.21/Add_4_output_0.out10_56", + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87", + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57", + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_111", + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_201_bfp.out25_113", + "/transformer_blocks.21/Add_7_output_0.out10_57", + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_88_bfp.out15_88", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_88_bfp.out1_58", + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_89_bfp.out15_89", + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_89_bfp.out1_60", + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66", + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67", + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68", + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22", + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44", + "/transformer_blocks.22/Add_output_0.out10_58", + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90", + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59", + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_115", + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_209_bfp.out25_116", + "/transformer_blocks.22/Add_3_output_0.out10_59" + ] + }, + "const": { + "buffer_size": 2272697344, + "xrt_arg_id": 3, + "packed_tensors": [ + "pos_embed.proj.weight", + "pos_embed_input.proj.weight", + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2", + "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1", + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3", + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0", + "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0", + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1", + "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2", + "encoder_hidden_states.out17_4_0_bfp.wts", + "onnx::MatMul_7927_45_1_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts", + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0", + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0", + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1", + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3", + "onnx::MatMul_7943_onnx::MatMul_7928", + "onnx::MatMul_7944_onnx::MatMul_7929", + "onnx::MatMul_7945_onnx::MatMul_7930", + "onnx::MatMul_7953_45_5_1", + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4_gma", + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3", + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2", + "onnx::MatMul_7954_45_5_0", + "onnx::MatMul_7957_45_1_2", + "onnx::MatMul_7958_45_1_4", + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma", + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma", + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2", + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0", + "onnx::MatMul_7955_45_1_1", + "onnx::MatMul_7956_45_1_3", + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma", + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4", + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1", + "onnx::MatMul_8635_45_0_4", + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5", + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1", + "onnx::MatMul_7974_onnx::MatMul_7959", + "onnx::MatMul_7975_onnx::MatMul_7960", + "onnx::MatMul_7976_onnx::MatMul_7961", + "onnx::MatMul_7985_45_5_2", + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma", + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6", + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4", + "onnx::MatMul_7986_45_1_5", + "onnx::MatMul_7987_45_1_7", + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma", + "onnx::MatMul_8636_45_0_5", + "onnx::MatMul_7984_45_5_3", + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma", + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7", + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4", + "onnx::MatMul_7988_45_1_6", + "onnx::MatMul_7989_45_1_8", + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma", + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8", + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9", + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1", + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1", + "onnx::MatMul_8005_onnx::MatMul_7990", + "onnx::MatMul_8006_onnx::MatMul_7991", + "onnx::MatMul_8007_onnx::MatMul_7992", + "onnx::MatMul_8015_45_5_5", + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma", + "onnx::MatMul_8016_45_5_4", + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma", + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10", + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4", + "onnx::MatMul_8017_45_1_9", + "onnx::MatMul_8018_45_1_11", + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma", + "onnx::MatMul_8637_45_0_6", + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11", + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4", + "onnx::MatMul_8019_45_1_10", + "onnx::MatMul_8020_45_1_12", + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma", + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12", + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13", + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1", + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1", + "onnx::MatMul_8036_onnx::MatMul_8021", + "onnx::MatMul_8037_onnx::MatMul_8022", + "onnx::MatMul_8038_onnx::MatMul_8023", + "onnx::MatMul_8046_45_5_7", + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma", + "onnx::MatMul_8047_45_5_6", + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma", + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14", + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4", + "onnx::MatMul_8048_45_1_13", + "onnx::MatMul_8049_45_1_15", + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma", + "onnx::MatMul_8638_45_0_7", + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15", + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4", + "onnx::MatMul_8050_45_1_14", + "onnx::MatMul_8051_45_1_16", + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma", + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16", + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17", + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1", + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1", + "onnx::MatMul_8067_onnx::MatMul_8052", + "onnx::MatMul_8068_onnx::MatMul_8053", + "onnx::MatMul_8069_onnx::MatMul_8054", + "onnx::MatMul_8077_45_5_9", + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma", + "onnx::MatMul_8078_45_5_8", + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma", + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18", + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4", + "onnx::MatMul_8079_45_1_17", + "onnx::MatMul_8080_45_1_19", + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma", + "onnx::MatMul_8639_45_0_8", + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19", + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4", + "onnx::MatMul_8081_45_1_18", + "onnx::MatMul_8082_45_1_20", + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma", + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20", + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21", + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1", + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1", + "onnx::MatMul_8098_onnx::MatMul_8083", + "onnx::MatMul_8099_onnx::MatMul_8084", + "onnx::MatMul_8100_onnx::MatMul_8085", + "onnx::MatMul_8108_45_5_11", + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_2_gma", + "onnx::MatMul_8109_45_5_10", + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma", + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22", + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4", + "onnx::MatMul_8110_45_1_21", + "onnx::MatMul_8111_45_1_23", + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma", + "onnx::MatMul_8640_45_0_9", + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23", + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_3_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_4", + "onnx::MatMul_8112_45_1_22", + "onnx::MatMul_8113_45_1_24", + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_5_gma", + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_24", + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_25", + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_0_transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_1", + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_0_transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_1", + "onnx::MatMul_8129_onnx::MatMul_8114", + "onnx::MatMul_8130_onnx::MatMul_8115", + "onnx::MatMul_8131_onnx::MatMul_8116", + "onnx::MatMul_8139_45_5_13", + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_2_gma", + "onnx::MatMul_8140_45_5_12", + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_2_gma", + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26", + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_3_transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_4", + "onnx::MatMul_8141_45_1_25", + "onnx::MatMul_8142_45_1_27", + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_5_gma", + "onnx::MatMul_8641_45_0_10", + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27", + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_3_transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_4", + "onnx::MatMul_8143_45_1_26", + "onnx::MatMul_8144_45_1_28", + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_5_gma", + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_28", + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_29", + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_0_transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_1", + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_0_transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_1", + "onnx::MatMul_8160_onnx::MatMul_8145", + "onnx::MatMul_8161_onnx::MatMul_8146", + "onnx::MatMul_8162_onnx::MatMul_8147", + "onnx::MatMul_8170_45_5_15", + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_2_gma", + "onnx::MatMul_8171_45_5_14", + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_2_gma", + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30", + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_3_transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_4", + "onnx::MatMul_8172_45_1_29", + "onnx::MatMul_8173_45_1_31", + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_5_gma", + "onnx::MatMul_8642_45_0_11", + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31", + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_3_transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_4", + "onnx::MatMul_8174_45_1_30", + "onnx::MatMul_8175_45_1_32", + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_5_gma", + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_32", + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_33", + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_0_transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_1", + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_0_transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_1", + "onnx::MatMul_8191_onnx::MatMul_8176", + "onnx::MatMul_8192_onnx::MatMul_8177", + "onnx::MatMul_8193_onnx::MatMul_8178", + "onnx::MatMul_8201_45_5_17", + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_2_gma", + "onnx::MatMul_8202_45_5_16", + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_2_gma", + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34", + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_3_transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_4", + "onnx::MatMul_8203_45_1_33", + "onnx::MatMul_8204_45_1_35", + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_5_gma", + "onnx::MatMul_8643_45_0_12", + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35", + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_3_transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_4", + "onnx::MatMul_8205_45_1_34", + "onnx::MatMul_8206_45_1_36", + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_5_gma", + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_36", + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_37", + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_0_transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_1", + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_0_transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_1", + "onnx::MatMul_8222_onnx::MatMul_8207", + "onnx::MatMul_8223_onnx::MatMul_8208", + "onnx::MatMul_8224_onnx::MatMul_8209", + "onnx::MatMul_8232_45_5_19", + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_2_gma", + "onnx::MatMul_8233_45_5_18", + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_2_gma", + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38", + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_3_transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_4", + "onnx::MatMul_8234_45_1_37", + "onnx::MatMul_8235_45_1_39", + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_5_gma", + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40", + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_0_transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_1", + "onnx::MatMul_8644_45_0_13", + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39", + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_3_transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_4", + "onnx::MatMul_8236_45_1_38", + "onnx::MatMul_8237_45_1_40", + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_5_gma", + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41", + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_0_transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_1", + "onnx::MatMul_8253_onnx::MatMul_8238", + "onnx::MatMul_8254_onnx::MatMul_8239", + "onnx::MatMul_8255_onnx::MatMul_8240", + "onnx::MatMul_8264_45_5_20", + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_2_gma", + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42", + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_3_transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_4", + "onnx::MatMul_8265_45_1_41", + "onnx::MatMul_8266_45_1_43", + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_5_gma", + "onnx::MatMul_8645_45_0_14", + "onnx::MatMul_8263_45_5_21", + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_2_gma", + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43", + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_3_transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_4", + "onnx::MatMul_8267_45_1_42", + "onnx::MatMul_8268_45_1_44", + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_5_gma", + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_44", + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_0_transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_1", + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_45", + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_0_transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_1", + "onnx::MatMul_8284_onnx::MatMul_8269", + "onnx::MatMul_8285_onnx::MatMul_8270", + "onnx::MatMul_8286_onnx::MatMul_8271", + "onnx::MatMul_8295_45_5_22", + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_2_gma", + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46", + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_3_transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_4", + "onnx::MatMul_8296_45_1_45", + "onnx::MatMul_8297_45_1_47", + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_5_gma", + "onnx::MatMul_8646_45_0_15", + "onnx::MatMul_8294_45_5_23", + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_2_gma", + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47", + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47", + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_3_transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_4", + "onnx::MatMul_8298_45_1_46", + "onnx::MatMul_8299_45_1_48", + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_5_gma", + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_48", + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_48", + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_0_transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_1", + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_49", + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_49", + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_0_transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_1", + "onnx::MatMul_8315_onnx::MatMul_8300", + "onnx::MatMul_8316_onnx::MatMul_8301", + "onnx::MatMul_8317_onnx::MatMul_8302", + "onnx::MatMul_8326_45_5_24", + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_2_gma", + "/transformer_blocks.12/norm2/Constant_output_0_14_50", + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50", + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_3_transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_4", + "onnx::MatMul_8327_45_1_49", + "onnx::MatMul_8328_45_1_51", + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_5_gma", + "onnx::MatMul_8647_45_0_16", + "onnx::MatMul_8325_45_5_25", + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_2_gma", + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51", + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51", + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_3_transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_4", + "onnx::MatMul_8329_45_1_50", + "onnx::MatMul_8330_45_1_52", + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_5_gma", + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_52", + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_52", + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_0_transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_1", + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_53", + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_53", + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_0_transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_1", + "onnx::MatMul_8346_onnx::MatMul_8331", + "onnx::MatMul_8347_onnx::MatMul_8332", + "onnx::MatMul_8348_onnx::MatMul_8333", + "onnx::MatMul_8357_45_5_26", + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_2_gma", + "/transformer_blocks.13/norm2/Constant_output_0_14_54", + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54", + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_3_transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_4", + "onnx::MatMul_8358_45_1_53", + "onnx::MatMul_8359_45_1_55", + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_5_gma", + "onnx::MatMul_8648_45_0_17", + "onnx::MatMul_8356_45_5_27", + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_2_gma", + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55", + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55", + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_3_transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_4", + "onnx::MatMul_8360_45_1_54", + "onnx::MatMul_8361_45_1_56", + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_5_gma", + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_56", + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_56", + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_0_transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_1", + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_57", + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_57", + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_0_transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_1", + "onnx::MatMul_8377_onnx::MatMul_8362", + "onnx::MatMul_8378_onnx::MatMul_8363", + "onnx::MatMul_8379_onnx::MatMul_8364", + "onnx::MatMul_8388_45_5_28", + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_2_gma", + "/transformer_blocks.14/norm2/Constant_output_0_14_58", + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58", + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_3_transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_4", + "onnx::MatMul_8389_45_1_57", + "onnx::MatMul_8390_45_1_59", + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_5_gma", + "onnx::MatMul_8649_45_0_18", + "onnx::MatMul_8387_45_5_29", + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_2_gma", + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59", + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59", + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_3_transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_4", + "onnx::MatMul_8391_45_1_58", + "onnx::MatMul_8392_45_1_60", + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_5_gma", + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_60", + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_60", + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_0_transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_1", + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_61", + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_61", + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_0_transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_1", + "onnx::MatMul_8408_onnx::MatMul_8393", + "onnx::MatMul_8409_onnx::MatMul_8394", + "onnx::MatMul_8410_onnx::MatMul_8395", + "onnx::MatMul_8419_45_5_30", + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_2_gma", + "/transformer_blocks.15/norm2/Constant_output_0_14_62", + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62", + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_3_transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_4", + "onnx::MatMul_8420_45_1_61", + "onnx::MatMul_8421_45_1_63", + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_5_gma", + "onnx::MatMul_8650_45_0_19", + "onnx::MatMul_8418_45_5_31", + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_2_gma", + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63", + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63", + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_3_transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_4", + "onnx::MatMul_8422_45_1_62", + "onnx::MatMul_8423_45_1_64", + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_5_gma", + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_64", + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_64", + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_0_transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_1", + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_65", + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_65", + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_0_transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_1", + "onnx::MatMul_8439_onnx::MatMul_8424", + "onnx::MatMul_8440_onnx::MatMul_8425", + "onnx::MatMul_8441_onnx::MatMul_8426", + "onnx::MatMul_8450_45_5_32", + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_2_gma", + "/transformer_blocks.16/norm2/Constant_output_0_14_66", + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66", + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_3_transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_4", + "onnx::MatMul_8451_45_1_65", + "onnx::MatMul_8452_45_1_67", + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_5_gma", + "onnx::MatMul_8651_45_0_20", + "onnx::MatMul_8449_45_5_33", + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_2_gma", + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67", + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67", + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_3_transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_4", + "onnx::MatMul_8453_45_1_66", + "onnx::MatMul_8454_45_1_68", + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_5_gma", + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_68", + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_68", + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_0_transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_1", + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_69", + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_69", + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_0_transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_1", + "onnx::MatMul_8470_onnx::MatMul_8455", + "onnx::MatMul_8471_onnx::MatMul_8456", + "onnx::MatMul_8472_onnx::MatMul_8457", + "onnx::MatMul_8481_45_5_34", + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_2_gma", + "/transformer_blocks.17/norm2/Constant_output_0_14_70", + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70", + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_3_transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_4", + "onnx::MatMul_8482_45_1_69", + "onnx::MatMul_8483_45_1_71", + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_5_gma", + "onnx::MatMul_8652_45_0_21", + "onnx::MatMul_8480_45_5_35", + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_2_gma", + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71", + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71", + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_3_transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_4", + "onnx::MatMul_8484_45_1_70", + "onnx::MatMul_8485_45_1_72", + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_5_gma", + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_72", + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_72", + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_0_transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_1", + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_73", + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_73", + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_0_transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_1", + "onnx::MatMul_8501_onnx::MatMul_8486", + "onnx::MatMul_8502_onnx::MatMul_8487", + "onnx::MatMul_8503_onnx::MatMul_8488", + "onnx::MatMul_8512_45_5_36", + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_2_gma", + "/transformer_blocks.18/norm2/Constant_output_0_14_74", + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74", + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_3_transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_4", + "onnx::MatMul_8513_45_1_73", + "onnx::MatMul_8514_45_1_75", + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_5_gma", + "onnx::MatMul_8653_45_0_22", + "onnx::MatMul_8511_45_5_37", + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_2_gma", + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75", + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75", + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_3_transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_4", + "onnx::MatMul_8515_45_1_74", + "onnx::MatMul_8516_45_1_76", + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_5_gma", + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_76", + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_76", + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_0_transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_1", + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_77", + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_77", + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_0_transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_1", + "onnx::MatMul_8532_onnx::MatMul_8517", + "onnx::MatMul_8533_onnx::MatMul_8518", + "onnx::MatMul_8534_onnx::MatMul_8519", + "onnx::MatMul_8543_45_5_38", + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_2_gma", + "/transformer_blocks.19/norm2/Constant_output_0_14_78", + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78", + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_3_transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_4", + "onnx::MatMul_8544_45_1_77", + "onnx::MatMul_8545_45_1_79", + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_5_gma", + "onnx::MatMul_8654_45_0_23", + "onnx::MatMul_8542_45_5_39", + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_2_gma", + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79", + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79", + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_3_transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_4", + "onnx::MatMul_8546_45_1_78", + "onnx::MatMul_8547_45_1_80", + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_5_gma", + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_80", + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_80", + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_81", + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_81", + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_0_transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_1", + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_0_transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_1", + "onnx::MatMul_8563_onnx::MatMul_8548", + "onnx::MatMul_8564_onnx::MatMul_8549", + "onnx::MatMul_8565_onnx::MatMul_8550", + "onnx::MatMul_8574_45_5_40", + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_2_gma", + "/transformer_blocks.20/norm2/Constant_output_0_14_82", + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82", + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_3_transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_4", + "onnx::MatMul_8575_45_1_81", + "onnx::MatMul_8576_45_1_83", + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_5_gma", + "onnx::MatMul_8655_45_0_24", + "onnx::MatMul_8573_45_5_41", + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_2_gma", + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83", + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83", + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_3_transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_4", + "onnx::MatMul_8577_45_1_82", + "onnx::MatMul_8578_45_1_84", + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_5_gma", + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_84", + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_84", + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_0_transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_1", + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_85", + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_85", + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_0_transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_1", + "onnx::MatMul_8594_onnx::MatMul_8579", + "onnx::MatMul_8595_onnx::MatMul_8580", + "onnx::MatMul_8596_onnx::MatMul_8581", + "onnx::MatMul_8605_45_5_42", + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_2_gma", + "/transformer_blocks.21/norm2/Constant_output_0_14_86", + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86", + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_3_transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_4", + "onnx::MatMul_8606_45_1_85", + "onnx::MatMul_8607_45_1_87", + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_5_gma", + "onnx::MatMul_8656_45_0_25", + "onnx::MatMul_8604_45_5_43", + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_2_gma", + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87", + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87", + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_3_transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_4", + "onnx::MatMul_8608_45_1_86", + "onnx::MatMul_8609_45_1_88", + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_5_gma", + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_88", + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_88", + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_0_transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_1", + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_89", + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_89", + "transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_0_transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_1", + "onnx::MatMul_8625_onnx::MatMul_8610", + "onnx::MatMul_8626_onnx::MatMul_8611", + "onnx::MatMul_8627_onnx::MatMul_8612", + "onnx::MatMul_8632_45_5_44", + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_2_gma", + "/transformer_blocks.22/norm2/Constant_output_0_14_90", + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90", + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_3_transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_4", + "onnx::MatMul_8633_45_1_89", + "onnx::MatMul_8634_45_1_90", + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_5_gma", + "onnx::MatMul_8657_45_0_26" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "hidden_states_nhwc.out5_0_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 16 + ], + "size_in_bytes": 32, + "op_tensor_size": 32, + "dynamic_shapes": [ + "batch_size", + "w", + "h", + "False" + ], + "offset": 0 + }, + "/pos_embed/Reshape_1_output_0.out_35_1_2": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "False", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 32 + }, + "controlnet_cond_nhwc.out5_0_1": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 24 + ], + "size_in_bytes": 48, + "op_tensor_size": 48, + "dynamic_shapes": [ + "batch_size", + "w", + "h", + "False" + ], + "offset": 3104 + }, + "/time_text_embed/Cast_output_0.out17_4_3": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 3152 + }, + "pooled_projections.out17_4_1": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 2048 + ], + "size_in_bytes": 4096, + "op_tensor_size": 4096, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 3664 + }, + "encoder_hidden_states.out17_4_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 4096 + ], + "size_in_bytes": 8192, + "op_tensor_size": 8192, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 7760 + }, + "/controlnet_blocks.0/Add_output_0.out17_4_13": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 0 + }, + "/controlnet_blocks.1/Add_output_0.out17_4_22": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 3072 + }, + "/controlnet_blocks.2/Add_output_0.out17_4_31": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 6144 + }, + "/controlnet_blocks.3/Add_output_0.out17_4_40": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 9216 + }, + "/controlnet_blocks.4/Add_output_0.out17_4_49": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 12288 + }, + "/controlnet_blocks.5/Add_output_0.out17_4_58": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 15360 + }, + "/controlnet_blocks.6/Add_output_0.out17_4_67": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 18432 + }, + "/controlnet_blocks.7/Add_output_0.out17_4_76": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 21504 + }, + "/controlnet_blocks.8/Add_output_0.out17_4_85": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 24576 + }, + "/controlnet_blocks.9/Add_output_0.out17_4_94": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 27648 + }, + "/controlnet_blocks.10/Add_output_0.out17_4_103": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 30720 + }, + "/controlnet_blocks.11/Add_output_0.out17_4_112": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 33792 + }, + "/controlnet_blocks.12/Add_output_0.out17_4_121": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 36864 + }, + "/controlnet_blocks.13/Add_output_0.out17_4_130": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 39936 + }, + "/controlnet_blocks.14/Add_output_0.out17_4_139": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 43008 + }, + "/controlnet_blocks.15/Add_output_0.out17_4_148": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 46080 + }, + "/controlnet_blocks.16/Add_output_0.out17_4_157": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 49152 + }, + "/controlnet_blocks.17/Add_output_0.out17_4_166": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 52224 + }, + "/controlnet_blocks.18/Add_output_0.out17_4_175": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 55296 + }, + "/controlnet_blocks.19/Add_output_0.out17_4_184": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 58368 + }, + "/controlnet_blocks.20/Add_output_0.out17_4_193": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 61440 + }, + "/controlnet_blocks.21/Add_output_0.out17_4_202": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 64512 + }, + "/controlnet_blocks.22/Add_output_0.out17_4_210": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 67584 + }, + "/pos_embed/Transpose_output_0.out5_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 0 + }, + "/pos_embed/Add_2_output_0.out_35_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 3072 + }, + "/pos_embed_input/Transpose_output_0.out5_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 6144 + }, + "/Add_output_0.out_35_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 9216 + }, + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 12288 + }, + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 15360 + }, + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 18432 + }, + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 21504 + }, + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 24576 + }, + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 27648 + }, + "/time_text_embed/Add_output_0.out_35_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 30720 + }, + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 33792 + }, + "encoder_hidden_states.out17_4_0_bfp.out25_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 4096 + ], + "size_in_bytes": 4096, + "op_tensor_size": 4096, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 36864 + }, + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 40960 + }, + "/context_embedder/Add_output_0.out17_4_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 42496 + }, + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 45568 + }, + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 47104 + }, + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 48640 + }, + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 50176 + }, + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 51712 + }, + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 53248 + }, + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 54784 + }, + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 56320 + }, + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 57856 + }, + "/transformer_blocks.0/Add_4_output_0.out10_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 59392 + }, + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 62464 + }, + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 64000 + }, + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 65536 + }, + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 67072 + }, + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 73216 + }, + "/transformer_blocks.0/Add_7_output_0.out10_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 74752 + }, + "/transformer_blocks.0/Add_output_0.out10_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 77824 + }, + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 80896 + }, + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 82432 + }, + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 83968 + }, + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 90112 + }, + "/transformer_blocks.0/Add_3_output_0.out10_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 91648 + }, + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 94720 + }, + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 96256 + }, + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 97792 + }, + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_5_bfp.out1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 99328 + }, + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 100864 + }, + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 102400 + }, + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 103936 + }, + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 105472 + }, + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 107008 + }, + "/transformer_blocks.1/Add_output_0.out10_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 108544 + }, + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 111616 + }, + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 113152 + }, + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 114688 + }, + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 120832 + }, + "/transformer_blocks.1/Add_3_output_0.out10_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 122368 + }, + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 125440 + }, + "/transformer_blocks.1/Add_4_output_0.out10_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 126976 + }, + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 130048 + }, + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 131584 + }, + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 133120 + }, + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 139264 + }, + "/transformer_blocks.1/Add_7_output_0.out10_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 140800 + }, + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 143872 + }, + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 145408 + }, + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 146944 + }, + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_9_bfp.out1_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 148480 + }, + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 150016 + }, + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 151552 + }, + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 153088 + }, + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 154624 + }, + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 156160 + }, + "/transformer_blocks.2/Add_4_output_0.out10_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 157696 + }, + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 160768 + }, + "/transformer_blocks.2/Add_output_0.out10_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 162304 + }, + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 165376 + }, + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 166912 + }, + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 168448 + }, + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 174592 + }, + "/transformer_blocks.2/Add_3_output_0.out10_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 176128 + }, + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 179200 + }, + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 180736 + }, + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 182272 + }, + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 188416 + }, + "/transformer_blocks.2/Add_7_output_0.out10_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 189952 + }, + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 193024 + }, + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 194560 + }, + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 196096 + }, + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_13_bfp.out1_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 197632 + }, + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 199168 + }, + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 200704 + }, + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 202240 + }, + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 203776 + }, + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 205312 + }, + "/transformer_blocks.3/Add_4_output_0.out10_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 206848 + }, + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 209920 + }, + "/transformer_blocks.3/Add_output_0.out10_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 211456 + }, + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 214528 + }, + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 216064 + }, + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 217600 + }, + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 223744 + }, + "/transformer_blocks.3/Add_3_output_0.out10_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 225280 + }, + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 228352 + }, + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 229888 + }, + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 231424 + }, + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 237568 + }, + "/transformer_blocks.3/Add_7_output_0.out10_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 239104 + }, + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 242176 + }, + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 243712 + }, + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 245248 + }, + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_17_bfp.out1_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 246784 + }, + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 248320 + }, + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 249856 + }, + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 251392 + }, + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 252928 + }, + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 254464 + }, + "/transformer_blocks.4/Add_4_output_0.out10_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 256000 + }, + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 259072 + }, + "/transformer_blocks.4/Add_output_0.out10_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 260608 + }, + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 263680 + }, + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 265216 + }, + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 266752 + }, + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 272896 + }, + "/transformer_blocks.4/Add_3_output_0.out10_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 274432 + }, + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 277504 + }, + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 279040 + }, + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 280576 + }, + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 286720 + }, + "/transformer_blocks.4/Add_7_output_0.out10_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 288256 + }, + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 291328 + }, + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 292864 + }, + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 294400 + }, + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_21_bfp.out1_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 295936 + }, + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 297472 + }, + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 299008 + }, + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 300544 + }, + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 302080 + }, + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 303616 + }, + "/transformer_blocks.5/Add_4_output_0.out10_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 305152 + }, + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 308224 + }, + "/transformer_blocks.5/Add_output_0.out10_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 309760 + }, + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 312832 + }, + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 314368 + }, + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 315904 + }, + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 322048 + }, + "/transformer_blocks.5/Add_3_output_0.out10_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 323584 + }, + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 326656 + }, + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 328192 + }, + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 329728 + }, + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 335872 + }, + "/transformer_blocks.5/Add_7_output_0.out10_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 337408 + }, + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_24_bfp.out15_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 340480 + }, + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_25_bfp.out15_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 342016 + }, + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_24_bfp.out1_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 343552 + }, + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_25_bfp.out1_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 345088 + }, + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 346624 + }, + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 348160 + }, + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 349696 + }, + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 351232 + }, + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 352768 + }, + "/transformer_blocks.6/Add_4_output_0.out10_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 354304 + }, + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 357376 + }, + "/transformer_blocks.6/Add_output_0.out10_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 358912 + }, + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 361984 + }, + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 363520 + }, + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 365056 + }, + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 371200 + }, + "/transformer_blocks.6/Add_3_output_0.out10_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 372736 + }, + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 375808 + }, + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 377344 + }, + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 378880 + }, + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 385024 + }, + "/transformer_blocks.6/Add_7_output_0.out10_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 386560 + }, + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_28_bfp.out15_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 389632 + }, + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_29_bfp.out15_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 391168 + }, + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_28_bfp.out1_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 392704 + }, + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_29_bfp.out1_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 394240 + }, + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 395776 + }, + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 397312 + }, + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 398848 + }, + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 400384 + }, + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 401920 + }, + "/transformer_blocks.7/Add_4_output_0.out10_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 403456 + }, + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 406528 + }, + "/transformer_blocks.7/Add_output_0.out10_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 408064 + }, + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 411136 + }, + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 412672 + }, + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 414208 + }, + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 420352 + }, + "/transformer_blocks.7/Add_3_output_0.out10_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 421888 + }, + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 424960 + }, + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 426496 + }, + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 428032 + }, + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 434176 + }, + "/transformer_blocks.7/Add_7_output_0.out10_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 435712 + }, + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_32_bfp.out15_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 438784 + }, + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_33_bfp.out15_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 440320 + }, + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_32_bfp.out1_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 441856 + }, + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_33_bfp.out1_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 443392 + }, + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 444928 + }, + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 446464 + }, + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 448000 + }, + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 449536 + }, + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 451072 + }, + "/transformer_blocks.8/Add_4_output_0.out10_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 452608 + }, + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 455680 + }, + "/transformer_blocks.8/Add_output_0.out10_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 457216 + }, + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 460288 + }, + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 461824 + }, + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 463360 + }, + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 469504 + }, + "/transformer_blocks.8/Add_3_output_0.out10_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 471040 + }, + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 474112 + }, + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 475648 + }, + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 477184 + }, + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 483328 + }, + "/transformer_blocks.8/Add_7_output_0.out10_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 484864 + }, + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_36_bfp.out15_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 487936 + }, + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_37_bfp.out15_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 489472 + }, + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_36_bfp.out1_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 491008 + }, + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_37_bfp.out1_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 492544 + }, + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 494080 + }, + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 495616 + }, + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 497152 + }, + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 498688 + }, + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 500224 + }, + "/transformer_blocks.9/Add_4_output_0.out10_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 501760 + }, + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 504832 + }, + "/transformer_blocks.9/Add_output_0.out10_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 506368 + }, + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 509440 + }, + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 510976 + }, + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 512512 + }, + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 518656 + }, + "/transformer_blocks.9/Add_3_output_0.out10_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 520192 + }, + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 523264 + }, + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 524800 + }, + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 526336 + }, + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 527872 + }, + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 529408 + }, + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 535552 + }, + "/transformer_blocks.9/Add_7_output_0.out10_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 537088 + }, + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 540160 + }, + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 541696 + }, + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 543232 + }, + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 544768 + }, + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 546304 + }, + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 547840 + }, + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 549376 + }, + "/transformer_blocks.10/Add_output_0.out10_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 550912 + }, + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 553984 + }, + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 555520 + }, + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 557056 + }, + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 563200 + }, + "/transformer_blocks.10/Add_3_output_0.out10_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 564736 + }, + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 567808 + }, + "/transformer_blocks.10/Add_4_output_0.out10_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 569344 + }, + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 572416 + }, + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 573952 + }, + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 575488 + }, + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 581632 + }, + "/transformer_blocks.10/Add_7_output_0.out10_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 583168 + }, + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_44_bfp.out15_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 586240 + }, + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_44_bfp.out1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 587776 + }, + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_45_bfp.out15_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 589312 + }, + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_45_bfp.out1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 590848 + }, + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 592384 + }, + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 593920 + }, + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 595456 + }, + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 596992 + }, + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 598528 + }, + "/transformer_blocks.11/Add_output_0.out10_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 600064 + }, + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 603136 + }, + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 604672 + }, + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 606208 + }, + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 612352 + }, + "/transformer_blocks.11/Add_3_output_0.out10_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 613888 + }, + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 616960 + }, + "/transformer_blocks.11/Add_4_output_0.out10_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 618496 + }, + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 621568 + }, + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 623104 + }, + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 624640 + }, + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_111_bfp.out25_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 630784 + }, + "/transformer_blocks.11/Add_7_output_0.out10_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 632320 + }, + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_48_bfp.out15_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 635392 + }, + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_48_bfp.out1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 636928 + }, + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_49_bfp.out15_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 638464 + }, + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_49_bfp.out1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 640000 + }, + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 641536 + }, + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 643072 + }, + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 644608 + }, + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 646144 + }, + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 647680 + }, + "/transformer_blocks.12/Add_output_0.out10_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 649216 + }, + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 652288 + }, + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 653824 + }, + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 655360 + }, + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_119_bfp.out25_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 661504 + }, + "/transformer_blocks.12/Add_3_output_0.out10_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 663040 + }, + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 666112 + }, + "/transformer_blocks.12/Add_4_output_0.out10_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 667648 + }, + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 670720 + }, + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 672256 + }, + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 673792 + }, + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_120_bfp.out25_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 679936 + }, + "/transformer_blocks.12/Add_7_output_0.out10_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 681472 + }, + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_52_bfp.out15_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 684544 + }, + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_52_bfp.out1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 686080 + }, + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_53_bfp.out15_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 687616 + }, + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_53_bfp.out1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 689152 + }, + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 690688 + }, + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 692224 + }, + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 693760 + }, + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 695296 + }, + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 696832 + }, + "/transformer_blocks.13/Add_output_0.out10_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 698368 + }, + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 701440 + }, + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 702976 + }, + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 704512 + }, + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_128_bfp.out25_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 710656 + }, + "/transformer_blocks.13/Add_3_output_0.out10_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 712192 + }, + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 715264 + }, + "/transformer_blocks.13/Add_4_output_0.out10_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 716800 + }, + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 719872 + }, + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 721408 + }, + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 722944 + }, + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_129_bfp.out25_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 729088 + }, + "/transformer_blocks.13/Add_7_output_0.out10_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 730624 + }, + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_56_bfp.out15_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 733696 + }, + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_56_bfp.out1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 735232 + }, + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_57_bfp.out15_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 736768 + }, + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_57_bfp.out1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 738304 + }, + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 739840 + }, + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 741376 + }, + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 742912 + }, + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 744448 + }, + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 745984 + }, + "/transformer_blocks.14/Add_output_0.out10_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 747520 + }, + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 750592 + }, + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 752128 + }, + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 753664 + }, + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_137_bfp.out25_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 759808 + }, + "/transformer_blocks.14/Add_3_output_0.out10_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 761344 + }, + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 764416 + }, + "/transformer_blocks.14/Add_4_output_0.out10_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 765952 + }, + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 769024 + }, + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 770560 + }, + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 772096 + }, + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_138_bfp.out25_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 778240 + }, + "/transformer_blocks.14/Add_7_output_0.out10_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 779776 + }, + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_60_bfp.out15_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 782848 + }, + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_60_bfp.out1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 784384 + }, + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_61_bfp.out15_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 785920 + }, + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_61_bfp.out1_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 787456 + }, + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 788992 + }, + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 790528 + }, + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 792064 + }, + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 793600 + }, + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 795136 + }, + "/transformer_blocks.15/Add_output_0.out10_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 796672 + }, + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 799744 + }, + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 801280 + }, + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 802816 + }, + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_146_bfp.out25_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 808960 + }, + "/transformer_blocks.15/Add_3_output_0.out10_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 810496 + }, + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 813568 + }, + "/transformer_blocks.15/Add_4_output_0.out10_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 815104 + }, + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 818176 + }, + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 819712 + }, + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 821248 + }, + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_147_bfp.out25_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 827392 + }, + "/transformer_blocks.15/Add_7_output_0.out10_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 828928 + }, + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_64_bfp.out15_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 832000 + }, + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_64_bfp.out1_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 833536 + }, + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_65_bfp.out15_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 835072 + }, + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_65_bfp.out1_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 836608 + }, + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 838144 + }, + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 839680 + }, + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 841216 + }, + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 842752 + }, + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 844288 + }, + "/transformer_blocks.16/Add_output_0.out10_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 845824 + }, + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 848896 + }, + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 850432 + }, + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 851968 + }, + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_155_bfp.out25_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 858112 + }, + "/transformer_blocks.16/Add_3_output_0.out10_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 859648 + }, + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 862720 + }, + "/transformer_blocks.16/Add_4_output_0.out10_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 864256 + }, + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 867328 + }, + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 868864 + }, + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 870400 + }, + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 876544 + }, + "/transformer_blocks.16/Add_7_output_0.out10_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 878080 + }, + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_68_bfp.out15_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 881152 + }, + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_68_bfp.out1_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 882688 + }, + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_69_bfp.out15_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 884224 + }, + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_69_bfp.out1_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 885760 + }, + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 887296 + }, + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 888832 + }, + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 890368 + }, + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 891904 + }, + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 893440 + }, + "/transformer_blocks.17/Add_output_0.out10_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 894976 + }, + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 898048 + }, + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 899584 + }, + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 901120 + }, + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_164_bfp.out25_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 907264 + }, + "/transformer_blocks.17/Add_3_output_0.out10_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 908800 + }, + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 911872 + }, + "/transformer_blocks.17/Add_4_output_0.out10_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 913408 + }, + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 916480 + }, + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 918016 + }, + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 919552 + }, + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_165_bfp.out25_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 925696 + }, + "/transformer_blocks.17/Add_7_output_0.out10_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 927232 + }, + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_72_bfp.out15_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 930304 + }, + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_72_bfp.out1_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 931840 + }, + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_73_bfp.out15_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 933376 + }, + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_73_bfp.out1_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 934912 + }, + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 936448 + }, + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 937984 + }, + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 939520 + }, + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 941056 + }, + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 942592 + }, + "/transformer_blocks.18/Add_output_0.out10_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 944128 + }, + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 947200 + }, + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 948736 + }, + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 950272 + }, + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_173_bfp.out25_97": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 956416 + }, + "/transformer_blocks.18/Add_3_output_0.out10_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 957952 + }, + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 961024 + }, + "/transformer_blocks.18/Add_4_output_0.out10_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 962560 + }, + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 965632 + }, + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 967168 + }, + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_96": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 968704 + }, + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_174_bfp.out25_98": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 974848 + }, + "/transformer_blocks.18/Add_7_output_0.out10_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 976384 + }, + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_76_bfp.out15_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 979456 + }, + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_76_bfp.out1_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 980992 + }, + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_77_bfp.out15_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 982528 + }, + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_77_bfp.out1_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 984064 + }, + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 985600 + }, + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 987136 + }, + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 988672 + }, + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 990208 + }, + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 991744 + }, + "/transformer_blocks.19/Add_output_0.out10_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 993280 + }, + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 996352 + }, + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 997888 + }, + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_100": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 999424 + }, + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_182_bfp.out25_102": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1005568 + }, + "/transformer_blocks.19/Add_3_output_0.out10_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1007104 + }, + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1010176 + }, + "/transformer_blocks.19/Add_4_output_0.out10_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1011712 + }, + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1014784 + }, + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1016320 + }, + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_101": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1017856 + }, + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_183_bfp.out25_103": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1024000 + }, + "/transformer_blocks.19/Add_7_output_0.out10_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1025536 + }, + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_80_bfp.out15_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1028608 + }, + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_81_bfp.out15_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1030144 + }, + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_80_bfp.out1_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1031680 + }, + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_81_bfp.out1_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1033216 + }, + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1034752 + }, + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1036288 + }, + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1037824 + }, + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1039360 + }, + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1040896 + }, + "/transformer_blocks.20/Add_output_0.out10_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1042432 + }, + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1045504 + }, + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1047040 + }, + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_105": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1048576 + }, + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_191_bfp.out25_107": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1054720 + }, + "/transformer_blocks.20/Add_3_output_0.out10_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1056256 + }, + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1059328 + }, + "/transformer_blocks.20/Add_4_output_0.out10_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1060864 + }, + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1063936 + }, + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1065472 + }, + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_106": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1067008 + }, + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_192_bfp.out25_108": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1073152 + }, + "/transformer_blocks.20/Add_7_output_0.out10_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1074688 + }, + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_84_bfp.out15_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1077760 + }, + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_84_bfp.out1_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1079296 + }, + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_85_bfp.out15_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1080832 + }, + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_85_bfp.out1_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1082368 + }, + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1083904 + }, + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1085440 + }, + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1086976 + }, + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1088512 + }, + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1090048 + }, + "/transformer_blocks.21/Add_output_0.out10_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1091584 + }, + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1094656 + }, + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1096192 + }, + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_110": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1097728 + }, + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_200_bfp.out25_112": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1103872 + }, + "/transformer_blocks.21/Add_3_output_0.out10_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1105408 + }, + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1108480 + }, + "/transformer_blocks.21/Add_4_output_0.out10_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1110016 + }, + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1113088 + }, + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1114624 + }, + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_111": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1116160 + }, + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_201_bfp.out25_113": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1122304 + }, + "/transformer_blocks.21/Add_7_output_0.out10_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1123840 + }, + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_88_bfp.out15_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1126912 + }, + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_88_bfp.out1_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1128448 + }, + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_89_bfp.out15_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1129984 + }, + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_89_bfp.out1_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1131520 + }, + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1133056 + }, + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1134592 + }, + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "floor(h/2)*floor(w/2)+max_length", + "False" + ], + "offset": 1136128 + }, + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1137664 + }, + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1139200 + }, + "/transformer_blocks.22/Add_output_0.out10_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1140736 + }, + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1143808 + }, + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1145344 + }, + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_115": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1146880 + }, + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_209_bfp.out25_116": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1153024 + }, + "/transformer_blocks.22/Add_3_output_0.out10_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "floor(h/2)*floor(w/2)", + "False" + ], + "offset": 1154560 + }, + "pos_embed.proj.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135168 + ], + "size_in_bytes": 135168, + "op_tensor_size": 135168, + "offset": 0, + "file_name": "cache/pos_embedprojConv_0.const", + "file_size": 135168 + }, + "pos_embed_input.proj.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 190464 + ], + "size_in_bytes": 190464, + "op_tensor_size": 190464, + "offset": 135168, + "file_name": "cache/pos_embedprojConv_1.const", + "file_size": 190464 + }, + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 448512 + ], + "size_in_bytes": 448512, + "op_tensor_size": 448512, + "offset": 325632, + "file_name": "cache/pos_embedprojConv_2.const", + "file_size": 448512 + }, + "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 774144, + "file_name": "cache/pos_embedprojConv_3.const", + "file_size": 256 + }, + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 774400, + "file_name": "cache/pos_embedprojConv_4.const", + "file_size": 2691072 + }, + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 3588096 + ], + "size_in_bytes": 3588096, + "op_tensor_size": 3588096, + "offset": 3465472, + "file_name": "cache/pos_embedprojConv_5.const", + "file_size": 3588096 + }, + "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 7053568, + "file_name": "cache/pos_embedprojConv_6.const", + "file_size": 256 + }, + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 7053824, + "file_name": "cache/pos_embedprojConv_7.const", + "file_size": 2691072 + }, + "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 9744896, + "file_name": "cache/pos_embedprojConv_8.const", + "file_size": 256 + }, + "encoder_hidden_states.out17_4_0_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 9745152, + "file_name": "cache/pos_embedprojConv_9.const", + "file_size": 128 + }, + "onnx::MatMul_7927_45_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7274496 + ], + "size_in_bytes": 7274496, + "op_tensor_size": 7274496, + "offset": 9745280, + "file_name": "cache/pos_embedprojConv_10.const", + "file_size": 7274496 + }, + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 17019776, + "file_name": "cache/pos_embedprojConv_11.const", + "file_size": 128 + }, + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 17019904, + "file_name": "cache/pos_embedprojConv_12.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 17022976, + "file_name": "cache/pos_embedprojConv_13.const", + "file_size": 3072 + }, + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 17026048, + "file_name": "cache/pos_embedprojConv_14.const", + "file_size": 5382144 + }, + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 22408192, + "file_name": "cache/pos_embedprojConv_15.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 22411264, + "file_name": "cache/pos_embedprojConv_16.const", + "file_size": 3072 + }, + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 22414336, + "file_name": "cache/pos_embedprojConv_17.const", + "file_size": 5382144 + }, + "onnx::MatMul_7943_onnx::MatMul_7928": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 27796480, + "file_name": "cache/pos_embedprojConv_18.const", + "file_size": 5382144 + }, + "onnx::MatMul_7944_onnx::MatMul_7929": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 33178624, + "file_name": "cache/pos_embedprojConv_19.const", + "file_size": 5382144 + }, + "onnx::MatMul_7945_onnx::MatMul_7930": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 38560768, + "file_name": "cache/pos_embedprojConv_20.const", + "file_size": 5382144 + }, + "onnx::MatMul_7953_45_5_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 43942912, + "file_name": "cache/pos_embedprojConv_21.const", + "file_size": 2691072 + }, + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 46633984, + "file_name": "cache/pos_embedprojConv_22.const", + "file_size": 2691072 + }, + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 49325056, + "file_name": "cache/pos_embedprojConv_23.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 49328128, + "file_name": "cache/pos_embedprojConv_24.const", + "file_size": 3072 + }, + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 49331200, + "file_name": "cache/pos_embedprojConv_25.const", + "file_size": 5382144 + }, + "onnx::MatMul_7954_45_5_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 54713344, + "file_name": "cache/pos_embedprojConv_26.const", + "file_size": 2691072 + }, + "onnx::MatMul_7957_45_1_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 57404416, + "file_name": "cache/pos_embedprojConv_27.const", + "file_size": 10764288 + }, + "onnx::MatMul_7958_45_1_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 68168704, + "file_name": "cache/pos_embedprojConv_28.const", + "file_size": 10764288 + }, + "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 78932992, + "file_name": "cache/pos_embedprojConv_29.const", + "file_size": 2691072 + }, + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 81624064, + "file_name": "cache/pos_embedprojConv_30.const", + "file_size": 2691072 + }, + "/transformer_blocks.0/norm2/Constant_output_0_14_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 84315136, + "file_name": "cache/pos_embedprojConv_31.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 84318208, + "file_name": "cache/pos_embedprojConv_32.const", + "file_size": 3072 + }, + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 84321280, + "file_name": "cache/pos_embedprojConv_33.const", + "file_size": 5382144 + }, + "onnx::MatMul_7955_45_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 89703424, + "file_name": "cache/pos_embedprojConv_34.const", + "file_size": 10764288 + }, + "onnx::MatMul_7956_45_1_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 100467712, + "file_name": "cache/pos_embedprojConv_35.const", + "file_size": 10764288 + }, + "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 111232000, + "file_name": "cache/pos_embedprojConv_36.const", + "file_size": 2691072 + }, + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 113923072, + "file_name": "cache/pos_embedprojConv_37.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 113926144, + "file_name": "cache/pos_embedprojConv_38.const", + "file_size": 3072 + }, + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 113929216, + "file_name": "cache/pos_embedprojConv_39.const", + "file_size": 5382144 + }, + "onnx::MatMul_8635_45_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 119311360, + "file_name": "cache/pos_embedprojConv_40.const", + "file_size": 2691072 + }, + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 122002432, + "file_name": "cache/pos_embedprojConv_41.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 122005504, + "file_name": "cache/pos_embedprojConv_42.const", + "file_size": 3072 + }, + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 122008576, + "file_name": "cache/pos_embedprojConv_43.const", + "file_size": 5382144 + }, + "onnx::MatMul_7974_onnx::MatMul_7959": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 127390720, + "file_name": "cache/pos_embedprojConv_44.const", + "file_size": 5382144 + }, + "onnx::MatMul_7975_onnx::MatMul_7960": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 132772864, + "file_name": "cache/pos_embedprojConv_45.const", + "file_size": 5382144 + }, + "onnx::MatMul_7976_onnx::MatMul_7961": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 138155008, + "file_name": "cache/pos_embedprojConv_46.const", + "file_size": 5382144 + }, + "onnx::MatMul_7985_45_5_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 143537152, + "file_name": "cache/pos_embedprojConv_47.const", + "file_size": 2691072 + }, + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 146228224, + "file_name": "cache/pos_embedprojConv_48.const", + "file_size": 2691072 + }, + "/transformer_blocks.1/norm2/Constant_output_0_14_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 148919296, + "file_name": "cache/pos_embedprojConv_49.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 148922368, + "file_name": "cache/pos_embedprojConv_50.const", + "file_size": 3072 + }, + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 148925440, + "file_name": "cache/pos_embedprojConv_51.const", + "file_size": 5382144 + }, + "onnx::MatMul_7986_45_1_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 154307584, + "file_name": "cache/pos_embedprojConv_52.const", + "file_size": 10764288 + }, + "onnx::MatMul_7987_45_1_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 165071872, + "file_name": "cache/pos_embedprojConv_53.const", + "file_size": 10764288 + }, + "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 175836160, + "file_name": "cache/pos_embedprojConv_54.const", + "file_size": 2691072 + }, + "onnx::MatMul_8636_45_0_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 178527232, + "file_name": "cache/pos_embedprojConv_55.const", + "file_size": 2691072 + }, + "onnx::MatMul_7984_45_5_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 181218304, + "file_name": "cache/pos_embedprojConv_56.const", + "file_size": 2691072 + }, + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 183909376, + "file_name": "cache/pos_embedprojConv_57.const", + "file_size": 2691072 + }, + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 186600448, + "file_name": "cache/pos_embedprojConv_58.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 186603520, + "file_name": "cache/pos_embedprojConv_59.const", + "file_size": 3072 + }, + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 186606592, + "file_name": "cache/pos_embedprojConv_60.const", + "file_size": 5382144 + }, + "onnx::MatMul_7988_45_1_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 191988736, + "file_name": "cache/pos_embedprojConv_61.const", + "file_size": 10764288 + }, + "onnx::MatMul_7989_45_1_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 202753024, + "file_name": "cache/pos_embedprojConv_62.const", + "file_size": 10764288 + }, + "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 213517312, + "file_name": "cache/pos_embedprojConv_63.const", + "file_size": 2691072 + }, + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 216208384, + "file_name": "cache/pos_embedprojConv_64.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 216211456, + "file_name": "cache/pos_embedprojConv_65.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 216214528, + "file_name": "cache/pos_embedprojConv_66.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 216217600, + "file_name": "cache/pos_embedprojConv_67.const", + "file_size": 3072 + }, + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 216220672, + "file_name": "cache/pos_embedprojConv_68.const", + "file_size": 5382144 + }, + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 221602816, + "file_name": "cache/pos_embedprojConv_69.const", + "file_size": 5382144 + }, + "onnx::MatMul_8005_onnx::MatMul_7990": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 226984960, + "file_name": "cache/pos_embedprojConv_70.const", + "file_size": 5382144 + }, + "onnx::MatMul_8006_onnx::MatMul_7991": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 232367104, + "file_name": "cache/pos_embedprojConv_71.const", + "file_size": 5382144 + }, + "onnx::MatMul_8007_onnx::MatMul_7992": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 237749248, + "file_name": "cache/pos_embedprojConv_72.const", + "file_size": 5382144 + }, + "onnx::MatMul_8015_45_5_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 243131392, + "file_name": "cache/pos_embedprojConv_73.const", + "file_size": 2691072 + }, + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 245822464, + "file_name": "cache/pos_embedprojConv_74.const", + "file_size": 2691072 + }, + "onnx::MatMul_8016_45_5_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 248513536, + "file_name": "cache/pos_embedprojConv_75.const", + "file_size": 2691072 + }, + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 251204608, + "file_name": "cache/pos_embedprojConv_76.const", + "file_size": 2691072 + }, + "/transformer_blocks.2/norm2/Constant_output_0_14_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 253895680, + "file_name": "cache/pos_embedprojConv_77.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 253898752, + "file_name": "cache/pos_embedprojConv_78.const", + "file_size": 3072 + }, + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 253901824, + "file_name": "cache/pos_embedprojConv_79.const", + "file_size": 5382144 + }, + "onnx::MatMul_8017_45_1_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 259283968, + "file_name": "cache/pos_embedprojConv_80.const", + "file_size": 10764288 + }, + "onnx::MatMul_8018_45_1_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 270048256, + "file_name": "cache/pos_embedprojConv_81.const", + "file_size": 10764288 + }, + "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 280812544, + "file_name": "cache/pos_embedprojConv_82.const", + "file_size": 2691072 + }, + "onnx::MatMul_8637_45_0_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 283503616, + "file_name": "cache/pos_embedprojConv_83.const", + "file_size": 2691072 + }, + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 286194688, + "file_name": "cache/pos_embedprojConv_84.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 286197760, + "file_name": "cache/pos_embedprojConv_85.const", + "file_size": 3072 + }, + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 286200832, + "file_name": "cache/pos_embedprojConv_86.const", + "file_size": 5382144 + }, + "onnx::MatMul_8019_45_1_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 291582976, + "file_name": "cache/pos_embedprojConv_87.const", + "file_size": 10764288 + }, + "onnx::MatMul_8020_45_1_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 302347264, + "file_name": "cache/pos_embedprojConv_88.const", + "file_size": 10764288 + }, + "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 313111552, + "file_name": "cache/pos_embedprojConv_89.const", + "file_size": 2691072 + }, + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 315802624, + "file_name": "cache/pos_embedprojConv_90.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 315805696, + "file_name": "cache/pos_embedprojConv_91.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 315808768, + "file_name": "cache/pos_embedprojConv_92.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 315811840, + "file_name": "cache/pos_embedprojConv_93.const", + "file_size": 3072 + }, + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 315814912, + "file_name": "cache/pos_embedprojConv_94.const", + "file_size": 5382144 + }, + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 321197056, + "file_name": "cache/pos_embedprojConv_95.const", + "file_size": 5382144 + }, + "onnx::MatMul_8036_onnx::MatMul_8021": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 326579200, + "file_name": "cache/pos_embedprojConv_96.const", + "file_size": 5382144 + }, + "onnx::MatMul_8037_onnx::MatMul_8022": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 331961344, + "file_name": "cache/pos_embedprojConv_97.const", + "file_size": 5382144 + }, + "onnx::MatMul_8038_onnx::MatMul_8023": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 337343488, + "file_name": "cache/pos_embedprojConv_98.const", + "file_size": 5382144 + }, + "onnx::MatMul_8046_45_5_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 342725632, + "file_name": "cache/pos_embedprojConv_99.const", + "file_size": 2691072 + }, + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 345416704, + "file_name": "cache/pos_embedprojConv_100.const", + "file_size": 2691072 + }, + "onnx::MatMul_8047_45_5_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 348107776, + "file_name": "cache/pos_embedprojConv_101.const", + "file_size": 2691072 + }, + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 350798848, + "file_name": "cache/pos_embedprojConv_102.const", + "file_size": 2691072 + }, + "/transformer_blocks.3/norm2/Constant_output_0_14_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 353489920, + "file_name": "cache/pos_embedprojConv_103.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 353492992, + "file_name": "cache/pos_embedprojConv_104.const", + "file_size": 3072 + }, + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 353496064, + "file_name": "cache/pos_embedprojConv_105.const", + "file_size": 5382144 + }, + "onnx::MatMul_8048_45_1_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 358878208, + "file_name": "cache/pos_embedprojConv_106.const", + "file_size": 10764288 + }, + "onnx::MatMul_8049_45_1_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 369642496, + "file_name": "cache/pos_embedprojConv_107.const", + "file_size": 10764288 + }, + "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 380406784, + "file_name": "cache/pos_embedprojConv_108.const", + "file_size": 2691072 + }, + "onnx::MatMul_8638_45_0_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 383097856, + "file_name": "cache/pos_embedprojConv_109.const", + "file_size": 2691072 + }, + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 385788928, + "file_name": "cache/pos_embedprojConv_110.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 385792000, + "file_name": "cache/pos_embedprojConv_111.const", + "file_size": 3072 + }, + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 385795072, + "file_name": "cache/pos_embedprojConv_112.const", + "file_size": 5382144 + }, + "onnx::MatMul_8050_45_1_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 391177216, + "file_name": "cache/pos_embedprojConv_113.const", + "file_size": 10764288 + }, + "onnx::MatMul_8051_45_1_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 401941504, + "file_name": "cache/pos_embedprojConv_114.const", + "file_size": 10764288 + }, + "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 412705792, + "file_name": "cache/pos_embedprojConv_115.const", + "file_size": 2691072 + }, + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 415396864, + "file_name": "cache/pos_embedprojConv_116.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 415399936, + "file_name": "cache/pos_embedprojConv_117.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 415403008, + "file_name": "cache/pos_embedprojConv_118.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 415406080, + "file_name": "cache/pos_embedprojConv_119.const", + "file_size": 3072 + }, + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 415409152, + "file_name": "cache/pos_embedprojConv_120.const", + "file_size": 5382144 + }, + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 420791296, + "file_name": "cache/pos_embedprojConv_121.const", + "file_size": 5382144 + }, + "onnx::MatMul_8067_onnx::MatMul_8052": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 426173440, + "file_name": "cache/pos_embedprojConv_122.const", + "file_size": 5382144 + }, + "onnx::MatMul_8068_onnx::MatMul_8053": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 431555584, + "file_name": "cache/pos_embedprojConv_123.const", + "file_size": 5382144 + }, + "onnx::MatMul_8069_onnx::MatMul_8054": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 436937728, + "file_name": "cache/pos_embedprojConv_124.const", + "file_size": 5382144 + }, + "onnx::MatMul_8077_45_5_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 442319872, + "file_name": "cache/pos_embedprojConv_125.const", + "file_size": 2691072 + }, + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 445010944, + "file_name": "cache/pos_embedprojConv_126.const", + "file_size": 2691072 + }, + "onnx::MatMul_8078_45_5_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 447702016, + "file_name": "cache/pos_embedprojConv_127.const", + "file_size": 2691072 + }, + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 450393088, + "file_name": "cache/pos_embedprojConv_128.const", + "file_size": 2691072 + }, + "/transformer_blocks.4/norm2/Constant_output_0_14_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 453084160, + "file_name": "cache/pos_embedprojConv_129.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 453087232, + "file_name": "cache/pos_embedprojConv_130.const", + "file_size": 3072 + }, + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 453090304, + "file_name": "cache/pos_embedprojConv_131.const", + "file_size": 5382144 + }, + "onnx::MatMul_8079_45_1_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 458472448, + "file_name": "cache/pos_embedprojConv_132.const", + "file_size": 10764288 + }, + "onnx::MatMul_8080_45_1_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 469236736, + "file_name": "cache/pos_embedprojConv_133.const", + "file_size": 10764288 + }, + "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 480001024, + "file_name": "cache/pos_embedprojConv_134.const", + "file_size": 2691072 + }, + "onnx::MatMul_8639_45_0_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 482692096, + "file_name": "cache/pos_embedprojConv_135.const", + "file_size": 2691072 + }, + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 485383168, + "file_name": "cache/pos_embedprojConv_136.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 485386240, + "file_name": "cache/pos_embedprojConv_137.const", + "file_size": 3072 + }, + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 485389312, + "file_name": "cache/pos_embedprojConv_138.const", + "file_size": 5382144 + }, + "onnx::MatMul_8081_45_1_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 490771456, + "file_name": "cache/pos_embedprojConv_139.const", + "file_size": 10764288 + }, + "onnx::MatMul_8082_45_1_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 501535744, + "file_name": "cache/pos_embedprojConv_140.const", + "file_size": 10764288 + }, + "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 512300032, + "file_name": "cache/pos_embedprojConv_141.const", + "file_size": 2691072 + }, + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 514991104, + "file_name": "cache/pos_embedprojConv_142.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 514994176, + "file_name": "cache/pos_embedprojConv_143.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 514997248, + "file_name": "cache/pos_embedprojConv_144.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 515000320, + "file_name": "cache/pos_embedprojConv_145.const", + "file_size": 3072 + }, + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 515003392, + "file_name": "cache/pos_embedprojConv_146.const", + "file_size": 5382144 + }, + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 520385536, + "file_name": "cache/pos_embedprojConv_147.const", + "file_size": 5382144 + }, + "onnx::MatMul_8098_onnx::MatMul_8083": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 525767680, + "file_name": "cache/pos_embedprojConv_148.const", + "file_size": 5382144 + }, + "onnx::MatMul_8099_onnx::MatMul_8084": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 531149824, + "file_name": "cache/pos_embedprojConv_149.const", + "file_size": 5382144 + }, + "onnx::MatMul_8100_onnx::MatMul_8085": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 536531968, + "file_name": "cache/pos_embedprojConv_150.const", + "file_size": 5382144 + }, + "onnx::MatMul_8108_45_5_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 541914112, + "file_name": "cache/pos_embedprojConv_151.const", + "file_size": 2691072 + }, + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 544605184, + "file_name": "cache/pos_embedprojConv_152.const", + "file_size": 2691072 + }, + "onnx::MatMul_8109_45_5_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 547296256, + "file_name": "cache/pos_embedprojConv_153.const", + "file_size": 2691072 + }, + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 549987328, + "file_name": "cache/pos_embedprojConv_154.const", + "file_size": 2691072 + }, + "/transformer_blocks.5/norm2/Constant_output_0_14_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 552678400, + "file_name": "cache/pos_embedprojConv_155.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 552681472, + "file_name": "cache/pos_embedprojConv_156.const", + "file_size": 3072 + }, + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 552684544, + "file_name": "cache/pos_embedprojConv_157.const", + "file_size": 5382144 + }, + "onnx::MatMul_8110_45_1_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 558066688, + "file_name": "cache/pos_embedprojConv_158.const", + "file_size": 10764288 + }, + "onnx::MatMul_8111_45_1_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 568830976, + "file_name": "cache/pos_embedprojConv_159.const", + "file_size": 10764288 + }, + "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 579595264, + "file_name": "cache/pos_embedprojConv_160.const", + "file_size": 2691072 + }, + "onnx::MatMul_8640_45_0_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 582286336, + "file_name": "cache/pos_embedprojConv_161.const", + "file_size": 2691072 + }, + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 584977408, + "file_name": "cache/pos_embedprojConv_162.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 584980480, + "file_name": "cache/pos_embedprojConv_163.const", + "file_size": 3072 + }, + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_3_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 584983552, + "file_name": "cache/pos_embedprojConv_164.const", + "file_size": 5382144 + }, + "onnx::MatMul_8112_45_1_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 590365696, + "file_name": "cache/pos_embedprojConv_165.const", + "file_size": 10764288 + }, + "onnx::MatMul_8113_45_1_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 601129984, + "file_name": "cache/pos_embedprojConv_166.const", + "file_size": 10764288 + }, + "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 611894272, + "file_name": "cache/pos_embedprojConv_167.const", + "file_size": 2691072 + }, + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 614585344, + "file_name": "cache/pos_embedprojConv_168.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 614588416, + "file_name": "cache/pos_embedprojConv_169.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 614591488, + "file_name": "cache/pos_embedprojConv_170.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 614594560, + "file_name": "cache/pos_embedprojConv_171.const", + "file_size": 3072 + }, + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_0_transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 614597632, + "file_name": "cache/pos_embedprojConv_172.const", + "file_size": 5382144 + }, + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_0_transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 619979776, + "file_name": "cache/pos_embedprojConv_173.const", + "file_size": 5382144 + }, + "onnx::MatMul_8129_onnx::MatMul_8114": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 625361920, + "file_name": "cache/pos_embedprojConv_174.const", + "file_size": 5382144 + }, + "onnx::MatMul_8130_onnx::MatMul_8115": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 630744064, + "file_name": "cache/pos_embedprojConv_175.const", + "file_size": 5382144 + }, + "onnx::MatMul_8131_onnx::MatMul_8116": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 636126208, + "file_name": "cache/pos_embedprojConv_176.const", + "file_size": 5382144 + }, + "onnx::MatMul_8139_45_5_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 641508352, + "file_name": "cache/pos_embedprojConv_177.const", + "file_size": 2691072 + }, + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 644199424, + "file_name": "cache/pos_embedprojConv_178.const", + "file_size": 2691072 + }, + "onnx::MatMul_8140_45_5_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 646890496, + "file_name": "cache/pos_embedprojConv_179.const", + "file_size": 2691072 + }, + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 649581568, + "file_name": "cache/pos_embedprojConv_180.const", + "file_size": 2691072 + }, + "/transformer_blocks.6/norm2/Constant_output_0_14_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 652272640, + "file_name": "cache/pos_embedprojConv_181.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 652275712, + "file_name": "cache/pos_embedprojConv_182.const", + "file_size": 3072 + }, + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_3_transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 652278784, + "file_name": "cache/pos_embedprojConv_183.const", + "file_size": 5382144 + }, + "onnx::MatMul_8141_45_1_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 657660928, + "file_name": "cache/pos_embedprojConv_184.const", + "file_size": 10764288 + }, + "onnx::MatMul_8142_45_1_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 668425216, + "file_name": "cache/pos_embedprojConv_185.const", + "file_size": 10764288 + }, + "transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 679189504, + "file_name": "cache/pos_embedprojConv_186.const", + "file_size": 2691072 + }, + "onnx::MatMul_8641_45_0_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 681880576, + "file_name": "cache/pos_embedprojConv_187.const", + "file_size": 2691072 + }, + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 684571648, + "file_name": "cache/pos_embedprojConv_188.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 684574720, + "file_name": "cache/pos_embedprojConv_189.const", + "file_size": 3072 + }, + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_3_transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 684577792, + "file_name": "cache/pos_embedprojConv_190.const", + "file_size": 5382144 + }, + "onnx::MatMul_8143_45_1_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 689959936, + "file_name": "cache/pos_embedprojConv_191.const", + "file_size": 10764288 + }, + "onnx::MatMul_8144_45_1_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 700724224, + "file_name": "cache/pos_embedprojConv_192.const", + "file_size": 10764288 + }, + "transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 711488512, + "file_name": "cache/pos_embedprojConv_193.const", + "file_size": 2691072 + }, + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 714179584, + "file_name": "cache/pos_embedprojConv_194.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 714182656, + "file_name": "cache/pos_embedprojConv_195.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 714185728, + "file_name": "cache/pos_embedprojConv_196.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 714188800, + "file_name": "cache/pos_embedprojConv_197.const", + "file_size": 3072 + }, + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_0_transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 714191872, + "file_name": "cache/pos_embedprojConv_198.const", + "file_size": 5382144 + }, + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_0_transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 719574016, + "file_name": "cache/pos_embedprojConv_199.const", + "file_size": 5382144 + }, + "onnx::MatMul_8160_onnx::MatMul_8145": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 724956160, + "file_name": "cache/pos_embedprojConv_200.const", + "file_size": 5382144 + }, + "onnx::MatMul_8161_onnx::MatMul_8146": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 730338304, + "file_name": "cache/pos_embedprojConv_201.const", + "file_size": 5382144 + }, + "onnx::MatMul_8162_onnx::MatMul_8147": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 735720448, + "file_name": "cache/pos_embedprojConv_202.const", + "file_size": 5382144 + }, + "onnx::MatMul_8170_45_5_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 741102592, + "file_name": "cache/pos_embedprojConv_203.const", + "file_size": 2691072 + }, + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 743793664, + "file_name": "cache/pos_embedprojConv_204.const", + "file_size": 2691072 + }, + "onnx::MatMul_8171_45_5_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 746484736, + "file_name": "cache/pos_embedprojConv_205.const", + "file_size": 2691072 + }, + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 749175808, + "file_name": "cache/pos_embedprojConv_206.const", + "file_size": 2691072 + }, + "/transformer_blocks.7/norm2/Constant_output_0_14_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 751866880, + "file_name": "cache/pos_embedprojConv_207.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 751869952, + "file_name": "cache/pos_embedprojConv_208.const", + "file_size": 3072 + }, + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_3_transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 751873024, + "file_name": "cache/pos_embedprojConv_209.const", + "file_size": 5382144 + }, + "onnx::MatMul_8172_45_1_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 757255168, + "file_name": "cache/pos_embedprojConv_210.const", + "file_size": 10764288 + }, + "onnx::MatMul_8173_45_1_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 768019456, + "file_name": "cache/pos_embedprojConv_211.const", + "file_size": 10764288 + }, + "transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 778783744, + "file_name": "cache/pos_embedprojConv_212.const", + "file_size": 2691072 + }, + "onnx::MatMul_8642_45_0_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 781474816, + "file_name": "cache/pos_embedprojConv_213.const", + "file_size": 2691072 + }, + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 784165888, + "file_name": "cache/pos_embedprojConv_214.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 784168960, + "file_name": "cache/pos_embedprojConv_215.const", + "file_size": 3072 + }, + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_3_transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 784172032, + "file_name": "cache/pos_embedprojConv_216.const", + "file_size": 5382144 + }, + "onnx::MatMul_8174_45_1_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 789554176, + "file_name": "cache/pos_embedprojConv_217.const", + "file_size": 10764288 + }, + "onnx::MatMul_8175_45_1_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 800318464, + "file_name": "cache/pos_embedprojConv_218.const", + "file_size": 10764288 + }, + "transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 811082752, + "file_name": "cache/pos_embedprojConv_219.const", + "file_size": 2691072 + }, + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 813773824, + "file_name": "cache/pos_embedprojConv_220.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 813776896, + "file_name": "cache/pos_embedprojConv_221.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 813779968, + "file_name": "cache/pos_embedprojConv_222.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 813783040, + "file_name": "cache/pos_embedprojConv_223.const", + "file_size": 3072 + }, + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_0_transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 813786112, + "file_name": "cache/pos_embedprojConv_224.const", + "file_size": 5382144 + }, + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_0_transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 819168256, + "file_name": "cache/pos_embedprojConv_225.const", + "file_size": 5382144 + }, + "onnx::MatMul_8191_onnx::MatMul_8176": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 824550400, + "file_name": "cache/pos_embedprojConv_226.const", + "file_size": 5382144 + }, + "onnx::MatMul_8192_onnx::MatMul_8177": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 829932544, + "file_name": "cache/pos_embedprojConv_227.const", + "file_size": 5382144 + }, + "onnx::MatMul_8193_onnx::MatMul_8178": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 835314688, + "file_name": "cache/pos_embedprojConv_228.const", + "file_size": 5382144 + }, + "onnx::MatMul_8201_45_5_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 840696832, + "file_name": "cache/pos_embedprojConv_229.const", + "file_size": 2691072 + }, + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 843387904, + "file_name": "cache/pos_embedprojConv_230.const", + "file_size": 2691072 + }, + "onnx::MatMul_8202_45_5_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 846078976, + "file_name": "cache/pos_embedprojConv_231.const", + "file_size": 2691072 + }, + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 848770048, + "file_name": "cache/pos_embedprojConv_232.const", + "file_size": 2691072 + }, + "/transformer_blocks.8/norm2/Constant_output_0_14_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 851461120, + "file_name": "cache/pos_embedprojConv_233.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 851464192, + "file_name": "cache/pos_embedprojConv_234.const", + "file_size": 3072 + }, + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_3_transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 851467264, + "file_name": "cache/pos_embedprojConv_235.const", + "file_size": 5382144 + }, + "onnx::MatMul_8203_45_1_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 856849408, + "file_name": "cache/pos_embedprojConv_236.const", + "file_size": 10764288 + }, + "onnx::MatMul_8204_45_1_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 867613696, + "file_name": "cache/pos_embedprojConv_237.const", + "file_size": 10764288 + }, + "transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 878377984, + "file_name": "cache/pos_embedprojConv_238.const", + "file_size": 2691072 + }, + "onnx::MatMul_8643_45_0_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 881069056, + "file_name": "cache/pos_embedprojConv_239.const", + "file_size": 2691072 + }, + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 883760128, + "file_name": "cache/pos_embedprojConv_240.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 883763200, + "file_name": "cache/pos_embedprojConv_241.const", + "file_size": 3072 + }, + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_3_transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 883766272, + "file_name": "cache/pos_embedprojConv_242.const", + "file_size": 5382144 + }, + "onnx::MatMul_8205_45_1_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 889148416, + "file_name": "cache/pos_embedprojConv_243.const", + "file_size": 10764288 + }, + "onnx::MatMul_8206_45_1_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 899912704, + "file_name": "cache/pos_embedprojConv_244.const", + "file_size": 10764288 + }, + "transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 910676992, + "file_name": "cache/pos_embedprojConv_245.const", + "file_size": 2691072 + }, + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 913368064, + "file_name": "cache/pos_embedprojConv_246.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 913371136, + "file_name": "cache/pos_embedprojConv_247.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 913374208, + "file_name": "cache/pos_embedprojConv_248.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 913377280, + "file_name": "cache/pos_embedprojConv_249.const", + "file_size": 3072 + }, + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_0_transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 913380352, + "file_name": "cache/pos_embedprojConv_250.const", + "file_size": 5382144 + }, + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_0_transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 918762496, + "file_name": "cache/pos_embedprojConv_251.const", + "file_size": 5382144 + }, + "onnx::MatMul_8222_onnx::MatMul_8207": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 924144640, + "file_name": "cache/pos_embedprojConv_252.const", + "file_size": 5382144 + }, + "onnx::MatMul_8223_onnx::MatMul_8208": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 929526784, + "file_name": "cache/pos_embedprojConv_253.const", + "file_size": 5382144 + }, + "onnx::MatMul_8224_onnx::MatMul_8209": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 934908928, + "file_name": "cache/pos_embedprojConv_254.const", + "file_size": 5382144 + }, + "onnx::MatMul_8232_45_5_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 940291072, + "file_name": "cache/pos_embedprojConv_255.const", + "file_size": 2691072 + }, + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 942982144, + "file_name": "cache/pos_embedprojConv_256.const", + "file_size": 2691072 + }, + "onnx::MatMul_8233_45_5_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 945673216, + "file_name": "cache/pos_embedprojConv_257.const", + "file_size": 2691072 + }, + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 948364288, + "file_name": "cache/pos_embedprojConv_258.const", + "file_size": 2691072 + }, + "/transformer_blocks.9/norm2/Constant_output_0_14_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 951055360, + "file_name": "cache/pos_embedprojConv_259.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 951058432, + "file_name": "cache/pos_embedprojConv_260.const", + "file_size": 3072 + }, + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_3_transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 951061504, + "file_name": "cache/pos_embedprojConv_261.const", + "file_size": 5382144 + }, + "onnx::MatMul_8234_45_1_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 956443648, + "file_name": "cache/pos_embedprojConv_262.const", + "file_size": 10764288 + }, + "onnx::MatMul_8235_45_1_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 967207936, + "file_name": "cache/pos_embedprojConv_263.const", + "file_size": 10764288 + }, + "transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 977972224, + "file_name": "cache/pos_embedprojConv_264.const", + "file_size": 2691072 + }, + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 980663296, + "file_name": "cache/pos_embedprojConv_265.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 980666368, + "file_name": "cache/pos_embedprojConv_266.const", + "file_size": 3072 + }, + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_0_transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 980669440, + "file_name": "cache/pos_embedprojConv_267.const", + "file_size": 5382144 + }, + "onnx::MatMul_8644_45_0_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 986051584, + "file_name": "cache/pos_embedprojConv_268.const", + "file_size": 2691072 + }, + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 988742656, + "file_name": "cache/pos_embedprojConv_269.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 988745728, + "file_name": "cache/pos_embedprojConv_270.const", + "file_size": 3072 + }, + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_3_transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 988748800, + "file_name": "cache/pos_embedprojConv_271.const", + "file_size": 5382144 + }, + "onnx::MatMul_8236_45_1_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 994130944, + "file_name": "cache/pos_embedprojConv_272.const", + "file_size": 10764288 + }, + "onnx::MatMul_8237_45_1_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1004895232, + "file_name": "cache/pos_embedprojConv_273.const", + "file_size": 10764288 + }, + "transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1015659520, + "file_name": "cache/pos_embedprojConv_274.const", + "file_size": 2691072 + }, + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1018350592, + "file_name": "cache/pos_embedprojConv_275.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1018353664, + "file_name": "cache/pos_embedprojConv_276.const", + "file_size": 3072 + }, + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_0_transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1018356736, + "file_name": "cache/pos_embedprojConv_277.const", + "file_size": 5382144 + }, + "onnx::MatMul_8253_onnx::MatMul_8238": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1023738880, + "file_name": "cache/pos_embedprojConv_278.const", + "file_size": 5382144 + }, + "onnx::MatMul_8254_onnx::MatMul_8239": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1029121024, + "file_name": "cache/pos_embedprojConv_279.const", + "file_size": 5382144 + }, + "onnx::MatMul_8255_onnx::MatMul_8240": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1034503168, + "file_name": "cache/pos_embedprojConv_280.const", + "file_size": 5382144 + }, + "onnx::MatMul_8264_45_5_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1039885312, + "file_name": "cache/pos_embedprojConv_281.const", + "file_size": 2691072 + }, + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1042576384, + "file_name": "cache/pos_embedprojConv_282.const", + "file_size": 2691072 + }, + "/transformer_blocks.10/norm2/Constant_output_0_14_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1045267456, + "file_name": "cache/pos_embedprojConv_283.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1045270528, + "file_name": "cache/pos_embedprojConv_284.const", + "file_size": 3072 + }, + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_3_transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1045273600, + "file_name": "cache/pos_embedprojConv_285.const", + "file_size": 5382144 + }, + "onnx::MatMul_8265_45_1_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1050655744, + "file_name": "cache/pos_embedprojConv_286.const", + "file_size": 10764288 + }, + "onnx::MatMul_8266_45_1_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1061420032, + "file_name": "cache/pos_embedprojConv_287.const", + "file_size": 10764288 + }, + "transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1072184320, + "file_name": "cache/pos_embedprojConv_288.const", + "file_size": 2691072 + }, + "onnx::MatMul_8645_45_0_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1074875392, + "file_name": "cache/pos_embedprojConv_289.const", + "file_size": 2691072 + }, + "onnx::MatMul_8263_45_5_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1077566464, + "file_name": "cache/pos_embedprojConv_290.const", + "file_size": 2691072 + }, + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1080257536, + "file_name": "cache/pos_embedprojConv_291.const", + "file_size": 2691072 + }, + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1082948608, + "file_name": "cache/pos_embedprojConv_292.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1082951680, + "file_name": "cache/pos_embedprojConv_293.const", + "file_size": 3072 + }, + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_3_transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1082954752, + "file_name": "cache/pos_embedprojConv_294.const", + "file_size": 5382144 + }, + "onnx::MatMul_8267_45_1_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1088336896, + "file_name": "cache/pos_embedprojConv_295.const", + "file_size": 10764288 + }, + "onnx::MatMul_8268_45_1_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1099101184, + "file_name": "cache/pos_embedprojConv_296.const", + "file_size": 10764288 + }, + "transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1109865472, + "file_name": "cache/pos_embedprojConv_297.const", + "file_size": 2691072 + }, + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1112556544, + "file_name": "cache/pos_embedprojConv_298.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1112559616, + "file_name": "cache/pos_embedprojConv_299.const", + "file_size": 3072 + }, + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_0_transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1112562688, + "file_name": "cache/pos_embedprojConv_300.const", + "file_size": 5382144 + }, + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1117944832, + "file_name": "cache/pos_embedprojConv_301.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1117947904, + "file_name": "cache/pos_embedprojConv_302.const", + "file_size": 3072 + }, + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_0_transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1117950976, + "file_name": "cache/pos_embedprojConv_303.const", + "file_size": 5382144 + }, + "onnx::MatMul_8284_onnx::MatMul_8269": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1123333120, + "file_name": "cache/pos_embedprojConv_304.const", + "file_size": 5382144 + }, + "onnx::MatMul_8285_onnx::MatMul_8270": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1128715264, + "file_name": "cache/pos_embedprojConv_305.const", + "file_size": 5382144 + }, + "onnx::MatMul_8286_onnx::MatMul_8271": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1134097408, + "file_name": "cache/pos_embedprojConv_306.const", + "file_size": 5382144 + }, + "onnx::MatMul_8295_45_5_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1139479552, + "file_name": "cache/pos_embedprojConv_307.const", + "file_size": 2691072 + }, + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1142170624, + "file_name": "cache/pos_embedprojConv_308.const", + "file_size": 2691072 + }, + "/transformer_blocks.11/norm2/Constant_output_0_14_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1144861696, + "file_name": "cache/pos_embedprojConv_309.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1144864768, + "file_name": "cache/pos_embedprojConv_310.const", + "file_size": 3072 + }, + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_3_transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1144867840, + "file_name": "cache/pos_embedprojConv_311.const", + "file_size": 5382144 + }, + "onnx::MatMul_8296_45_1_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1150249984, + "file_name": "cache/pos_embedprojConv_312.const", + "file_size": 10764288 + }, + "onnx::MatMul_8297_45_1_47": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1161014272, + "file_name": "cache/pos_embedprojConv_313.const", + "file_size": 10764288 + }, + "transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1171778560, + "file_name": "cache/pos_embedprojConv_314.const", + "file_size": 2691072 + }, + "onnx::MatMul_8646_45_0_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1174469632, + "file_name": "cache/pos_embedprojConv_315.const", + "file_size": 2691072 + }, + "onnx::MatMul_8294_45_5_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1177160704, + "file_name": "cache/pos_embedprojConv_316.const", + "file_size": 2691072 + }, + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1179851776, + "file_name": "cache/pos_embedprojConv_317.const", + "file_size": 2691072 + }, + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1182542848, + "file_name": "cache/pos_embedprojConv_318.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1182545920, + "file_name": "cache/pos_embedprojConv_319.const", + "file_size": 3072 + }, + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_3_transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1182548992, + "file_name": "cache/pos_embedprojConv_320.const", + "file_size": 5382144 + }, + "onnx::MatMul_8298_45_1_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1187931136, + "file_name": "cache/pos_embedprojConv_321.const", + "file_size": 10764288 + }, + "onnx::MatMul_8299_45_1_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1198695424, + "file_name": "cache/pos_embedprojConv_322.const", + "file_size": 10764288 + }, + "transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1209459712, + "file_name": "cache/pos_embedprojConv_323.const", + "file_size": 2691072 + }, + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1212150784, + "file_name": "cache/pos_embedprojConv_324.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1212153856, + "file_name": "cache/pos_embedprojConv_325.const", + "file_size": 3072 + }, + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_0_transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1212156928, + "file_name": "cache/pos_embedprojConv_326.const", + "file_size": 5382144 + }, + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1217539072, + "file_name": "cache/pos_embedprojConv_327.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1217542144, + "file_name": "cache/pos_embedprojConv_328.const", + "file_size": 3072 + }, + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_0_transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1217545216, + "file_name": "cache/pos_embedprojConv_329.const", + "file_size": 5382144 + }, + "onnx::MatMul_8315_onnx::MatMul_8300": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1222927360, + "file_name": "cache/pos_embedprojConv_330.const", + "file_size": 5382144 + }, + "onnx::MatMul_8316_onnx::MatMul_8301": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1228309504, + "file_name": "cache/pos_embedprojConv_331.const", + "file_size": 5382144 + }, + "onnx::MatMul_8317_onnx::MatMul_8302": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1233691648, + "file_name": "cache/pos_embedprojConv_332.const", + "file_size": 5382144 + }, + "onnx::MatMul_8326_45_5_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1239073792, + "file_name": "cache/pos_embedprojConv_333.const", + "file_size": 2691072 + }, + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1241764864, + "file_name": "cache/pos_embedprojConv_334.const", + "file_size": 2691072 + }, + "/transformer_blocks.12/norm2/Constant_output_0_14_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1244455936, + "file_name": "cache/pos_embedprojConv_335.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1244459008, + "file_name": "cache/pos_embedprojConv_336.const", + "file_size": 3072 + }, + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_3_transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1244462080, + "file_name": "cache/pos_embedprojConv_337.const", + "file_size": 5382144 + }, + "onnx::MatMul_8327_45_1_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1249844224, + "file_name": "cache/pos_embedprojConv_338.const", + "file_size": 10764288 + }, + "onnx::MatMul_8328_45_1_51": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1260608512, + "file_name": "cache/pos_embedprojConv_339.const", + "file_size": 10764288 + }, + "transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1271372800, + "file_name": "cache/pos_embedprojConv_340.const", + "file_size": 2691072 + }, + "onnx::MatMul_8647_45_0_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1274063872, + "file_name": "cache/pos_embedprojConv_341.const", + "file_size": 2691072 + }, + "onnx::MatMul_8325_45_5_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1276754944, + "file_name": "cache/pos_embedprojConv_342.const", + "file_size": 2691072 + }, + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1279446016, + "file_name": "cache/pos_embedprojConv_343.const", + "file_size": 2691072 + }, + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1282137088, + "file_name": "cache/pos_embedprojConv_344.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1282140160, + "file_name": "cache/pos_embedprojConv_345.const", + "file_size": 3072 + }, + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_3_transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1282143232, + "file_name": "cache/pos_embedprojConv_346.const", + "file_size": 5382144 + }, + "onnx::MatMul_8329_45_1_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1287525376, + "file_name": "cache/pos_embedprojConv_347.const", + "file_size": 10764288 + }, + "onnx::MatMul_8330_45_1_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1298289664, + "file_name": "cache/pos_embedprojConv_348.const", + "file_size": 10764288 + }, + "transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1309053952, + "file_name": "cache/pos_embedprojConv_349.const", + "file_size": 2691072 + }, + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1311745024, + "file_name": "cache/pos_embedprojConv_350.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1311748096, + "file_name": "cache/pos_embedprojConv_351.const", + "file_size": 3072 + }, + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_0_transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1311751168, + "file_name": "cache/pos_embedprojConv_352.const", + "file_size": 5382144 + }, + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1317133312, + "file_name": "cache/pos_embedprojConv_353.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1317136384, + "file_name": "cache/pos_embedprojConv_354.const", + "file_size": 3072 + }, + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_0_transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1317139456, + "file_name": "cache/pos_embedprojConv_355.const", + "file_size": 5382144 + }, + "onnx::MatMul_8346_onnx::MatMul_8331": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1322521600, + "file_name": "cache/pos_embedprojConv_356.const", + "file_size": 5382144 + }, + "onnx::MatMul_8347_onnx::MatMul_8332": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1327903744, + "file_name": "cache/pos_embedprojConv_357.const", + "file_size": 5382144 + }, + "onnx::MatMul_8348_onnx::MatMul_8333": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1333285888, + "file_name": "cache/pos_embedprojConv_358.const", + "file_size": 5382144 + }, + "onnx::MatMul_8357_45_5_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1338668032, + "file_name": "cache/pos_embedprojConv_359.const", + "file_size": 2691072 + }, + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1341359104, + "file_name": "cache/pos_embedprojConv_360.const", + "file_size": 2691072 + }, + "/transformer_blocks.13/norm2/Constant_output_0_14_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1344050176, + "file_name": "cache/pos_embedprojConv_361.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1344053248, + "file_name": "cache/pos_embedprojConv_362.const", + "file_size": 3072 + }, + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_3_transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1344056320, + "file_name": "cache/pos_embedprojConv_363.const", + "file_size": 5382144 + }, + "onnx::MatMul_8358_45_1_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1349438464, + "file_name": "cache/pos_embedprojConv_364.const", + "file_size": 10764288 + }, + "onnx::MatMul_8359_45_1_55": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1360202752, + "file_name": "cache/pos_embedprojConv_365.const", + "file_size": 10764288 + }, + "transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1370967040, + "file_name": "cache/pos_embedprojConv_366.const", + "file_size": 2691072 + }, + "onnx::MatMul_8648_45_0_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1373658112, + "file_name": "cache/pos_embedprojConv_367.const", + "file_size": 2691072 + }, + "onnx::MatMul_8356_45_5_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1376349184, + "file_name": "cache/pos_embedprojConv_368.const", + "file_size": 2691072 + }, + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1379040256, + "file_name": "cache/pos_embedprojConv_369.const", + "file_size": 2691072 + }, + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1381731328, + "file_name": "cache/pos_embedprojConv_370.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1381734400, + "file_name": "cache/pos_embedprojConv_371.const", + "file_size": 3072 + }, + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_3_transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1381737472, + "file_name": "cache/pos_embedprojConv_372.const", + "file_size": 5382144 + }, + "onnx::MatMul_8360_45_1_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1387119616, + "file_name": "cache/pos_embedprojConv_373.const", + "file_size": 10764288 + }, + "onnx::MatMul_8361_45_1_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1397883904, + "file_name": "cache/pos_embedprojConv_374.const", + "file_size": 10764288 + }, + "transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1408648192, + "file_name": "cache/pos_embedprojConv_375.const", + "file_size": 2691072 + }, + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1411339264, + "file_name": "cache/pos_embedprojConv_376.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1411342336, + "file_name": "cache/pos_embedprojConv_377.const", + "file_size": 3072 + }, + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_0_transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1411345408, + "file_name": "cache/pos_embedprojConv_378.const", + "file_size": 5382144 + }, + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1416727552, + "file_name": "cache/pos_embedprojConv_379.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1416730624, + "file_name": "cache/pos_embedprojConv_380.const", + "file_size": 3072 + }, + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_0_transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1416733696, + "file_name": "cache/pos_embedprojConv_381.const", + "file_size": 5382144 + }, + "onnx::MatMul_8377_onnx::MatMul_8362": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1422115840, + "file_name": "cache/pos_embedprojConv_382.const", + "file_size": 5382144 + }, + "onnx::MatMul_8378_onnx::MatMul_8363": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1427497984, + "file_name": "cache/pos_embedprojConv_383.const", + "file_size": 5382144 + }, + "onnx::MatMul_8379_onnx::MatMul_8364": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1432880128, + "file_name": "cache/pos_embedprojConv_384.const", + "file_size": 5382144 + }, + "onnx::MatMul_8388_45_5_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1438262272, + "file_name": "cache/pos_embedprojConv_385.const", + "file_size": 2691072 + }, + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1440953344, + "file_name": "cache/pos_embedprojConv_386.const", + "file_size": 2691072 + }, + "/transformer_blocks.14/norm2/Constant_output_0_14_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1443644416, + "file_name": "cache/pos_embedprojConv_387.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1443647488, + "file_name": "cache/pos_embedprojConv_388.const", + "file_size": 3072 + }, + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_3_transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1443650560, + "file_name": "cache/pos_embedprojConv_389.const", + "file_size": 5382144 + }, + "onnx::MatMul_8389_45_1_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1449032704, + "file_name": "cache/pos_embedprojConv_390.const", + "file_size": 10764288 + }, + "onnx::MatMul_8390_45_1_59": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1459796992, + "file_name": "cache/pos_embedprojConv_391.const", + "file_size": 10764288 + }, + "transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1470561280, + "file_name": "cache/pos_embedprojConv_392.const", + "file_size": 2691072 + }, + "onnx::MatMul_8649_45_0_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1473252352, + "file_name": "cache/pos_embedprojConv_393.const", + "file_size": 2691072 + }, + "onnx::MatMul_8387_45_5_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1475943424, + "file_name": "cache/pos_embedprojConv_394.const", + "file_size": 2691072 + }, + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1478634496, + "file_name": "cache/pos_embedprojConv_395.const", + "file_size": 2691072 + }, + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1481325568, + "file_name": "cache/pos_embedprojConv_396.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1481328640, + "file_name": "cache/pos_embedprojConv_397.const", + "file_size": 3072 + }, + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_3_transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1481331712, + "file_name": "cache/pos_embedprojConv_398.const", + "file_size": 5382144 + }, + "onnx::MatMul_8391_45_1_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1486713856, + "file_name": "cache/pos_embedprojConv_399.const", + "file_size": 10764288 + }, + "onnx::MatMul_8392_45_1_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1497478144, + "file_name": "cache/pos_embedprojConv_400.const", + "file_size": 10764288 + }, + "transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1508242432, + "file_name": "cache/pos_embedprojConv_401.const", + "file_size": 2691072 + }, + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1510933504, + "file_name": "cache/pos_embedprojConv_402.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1510936576, + "file_name": "cache/pos_embedprojConv_403.const", + "file_size": 3072 + }, + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_0_transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1510939648, + "file_name": "cache/pos_embedprojConv_404.const", + "file_size": 5382144 + }, + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_61": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1516321792, + "file_name": "cache/pos_embedprojConv_405.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_61": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1516324864, + "file_name": "cache/pos_embedprojConv_406.const", + "file_size": 3072 + }, + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_0_transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1516327936, + "file_name": "cache/pos_embedprojConv_407.const", + "file_size": 5382144 + }, + "onnx::MatMul_8408_onnx::MatMul_8393": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1521710080, + "file_name": "cache/pos_embedprojConv_408.const", + "file_size": 5382144 + }, + "onnx::MatMul_8409_onnx::MatMul_8394": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1527092224, + "file_name": "cache/pos_embedprojConv_409.const", + "file_size": 5382144 + }, + "onnx::MatMul_8410_onnx::MatMul_8395": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1532474368, + "file_name": "cache/pos_embedprojConv_410.const", + "file_size": 5382144 + }, + "onnx::MatMul_8419_45_5_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1537856512, + "file_name": "cache/pos_embedprojConv_411.const", + "file_size": 2691072 + }, + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1540547584, + "file_name": "cache/pos_embedprojConv_412.const", + "file_size": 2691072 + }, + "/transformer_blocks.15/norm2/Constant_output_0_14_62": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1543238656, + "file_name": "cache/pos_embedprojConv_413.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1543241728, + "file_name": "cache/pos_embedprojConv_414.const", + "file_size": 3072 + }, + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_3_transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1543244800, + "file_name": "cache/pos_embedprojConv_415.const", + "file_size": 5382144 + }, + "onnx::MatMul_8420_45_1_61": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1548626944, + "file_name": "cache/pos_embedprojConv_416.const", + "file_size": 10764288 + }, + "onnx::MatMul_8421_45_1_63": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1559391232, + "file_name": "cache/pos_embedprojConv_417.const", + "file_size": 10764288 + }, + "transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1570155520, + "file_name": "cache/pos_embedprojConv_418.const", + "file_size": 2691072 + }, + "onnx::MatMul_8650_45_0_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1572846592, + "file_name": "cache/pos_embedprojConv_419.const", + "file_size": 2691072 + }, + "onnx::MatMul_8418_45_5_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1575537664, + "file_name": "cache/pos_embedprojConv_420.const", + "file_size": 2691072 + }, + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1578228736, + "file_name": "cache/pos_embedprojConv_421.const", + "file_size": 2691072 + }, + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1580919808, + "file_name": "cache/pos_embedprojConv_422.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1580922880, + "file_name": "cache/pos_embedprojConv_423.const", + "file_size": 3072 + }, + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_3_transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1580925952, + "file_name": "cache/pos_embedprojConv_424.const", + "file_size": 5382144 + }, + "onnx::MatMul_8422_45_1_62": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1586308096, + "file_name": "cache/pos_embedprojConv_425.const", + "file_size": 10764288 + }, + "onnx::MatMul_8423_45_1_64": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1597072384, + "file_name": "cache/pos_embedprojConv_426.const", + "file_size": 10764288 + }, + "transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1607836672, + "file_name": "cache/pos_embedprojConv_427.const", + "file_size": 2691072 + }, + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_64": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1610527744, + "file_name": "cache/pos_embedprojConv_428.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_64": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1610530816, + "file_name": "cache/pos_embedprojConv_429.const", + "file_size": 3072 + }, + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_0_transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1610533888, + "file_name": "cache/pos_embedprojConv_430.const", + "file_size": 5382144 + }, + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_65": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1615916032, + "file_name": "cache/pos_embedprojConv_431.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_65": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1615919104, + "file_name": "cache/pos_embedprojConv_432.const", + "file_size": 3072 + }, + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_0_transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1615922176, + "file_name": "cache/pos_embedprojConv_433.const", + "file_size": 5382144 + }, + "onnx::MatMul_8439_onnx::MatMul_8424": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1621304320, + "file_name": "cache/pos_embedprojConv_434.const", + "file_size": 5382144 + }, + "onnx::MatMul_8440_onnx::MatMul_8425": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1626686464, + "file_name": "cache/pos_embedprojConv_435.const", + "file_size": 5382144 + }, + "onnx::MatMul_8441_onnx::MatMul_8426": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1632068608, + "file_name": "cache/pos_embedprojConv_436.const", + "file_size": 5382144 + }, + "onnx::MatMul_8450_45_5_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1637450752, + "file_name": "cache/pos_embedprojConv_437.const", + "file_size": 2691072 + }, + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1640141824, + "file_name": "cache/pos_embedprojConv_438.const", + "file_size": 2691072 + }, + "/transformer_blocks.16/norm2/Constant_output_0_14_66": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1642832896, + "file_name": "cache/pos_embedprojConv_439.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1642835968, + "file_name": "cache/pos_embedprojConv_440.const", + "file_size": 3072 + }, + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_3_transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1642839040, + "file_name": "cache/pos_embedprojConv_441.const", + "file_size": 5382144 + }, + "onnx::MatMul_8451_45_1_65": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1648221184, + "file_name": "cache/pos_embedprojConv_442.const", + "file_size": 10764288 + }, + "onnx::MatMul_8452_45_1_67": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1658985472, + "file_name": "cache/pos_embedprojConv_443.const", + "file_size": 10764288 + }, + "transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1669749760, + "file_name": "cache/pos_embedprojConv_444.const", + "file_size": 2691072 + }, + "onnx::MatMul_8651_45_0_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1672440832, + "file_name": "cache/pos_embedprojConv_445.const", + "file_size": 2691072 + }, + "onnx::MatMul_8449_45_5_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1675131904, + "file_name": "cache/pos_embedprojConv_446.const", + "file_size": 2691072 + }, + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1677822976, + "file_name": "cache/pos_embedprojConv_447.const", + "file_size": 2691072 + }, + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1680514048, + "file_name": "cache/pos_embedprojConv_448.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1680517120, + "file_name": "cache/pos_embedprojConv_449.const", + "file_size": 3072 + }, + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_3_transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1680520192, + "file_name": "cache/pos_embedprojConv_450.const", + "file_size": 5382144 + }, + "onnx::MatMul_8453_45_1_66": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1685902336, + "file_name": "cache/pos_embedprojConv_451.const", + "file_size": 10764288 + }, + "onnx::MatMul_8454_45_1_68": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1696666624, + "file_name": "cache/pos_embedprojConv_452.const", + "file_size": 10764288 + }, + "transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1707430912, + "file_name": "cache/pos_embedprojConv_453.const", + "file_size": 2691072 + }, + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_68": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1710121984, + "file_name": "cache/pos_embedprojConv_454.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_68": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1710125056, + "file_name": "cache/pos_embedprojConv_455.const", + "file_size": 3072 + }, + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_0_transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1710128128, + "file_name": "cache/pos_embedprojConv_456.const", + "file_size": 5382144 + }, + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_69": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1715510272, + "file_name": "cache/pos_embedprojConv_457.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_69": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1715513344, + "file_name": "cache/pos_embedprojConv_458.const", + "file_size": 3072 + }, + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_0_transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1715516416, + "file_name": "cache/pos_embedprojConv_459.const", + "file_size": 5382144 + }, + "onnx::MatMul_8470_onnx::MatMul_8455": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1720898560, + "file_name": "cache/pos_embedprojConv_460.const", + "file_size": 5382144 + }, + "onnx::MatMul_8471_onnx::MatMul_8456": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1726280704, + "file_name": "cache/pos_embedprojConv_461.const", + "file_size": 5382144 + }, + "onnx::MatMul_8472_onnx::MatMul_8457": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1731662848, + "file_name": "cache/pos_embedprojConv_462.const", + "file_size": 5382144 + }, + "onnx::MatMul_8481_45_5_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1737044992, + "file_name": "cache/pos_embedprojConv_463.const", + "file_size": 2691072 + }, + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1739736064, + "file_name": "cache/pos_embedprojConv_464.const", + "file_size": 2691072 + }, + "/transformer_blocks.17/norm2/Constant_output_0_14_70": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1742427136, + "file_name": "cache/pos_embedprojConv_465.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1742430208, + "file_name": "cache/pos_embedprojConv_466.const", + "file_size": 3072 + }, + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_3_transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1742433280, + "file_name": "cache/pos_embedprojConv_467.const", + "file_size": 5382144 + }, + "onnx::MatMul_8482_45_1_69": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1747815424, + "file_name": "cache/pos_embedprojConv_468.const", + "file_size": 10764288 + }, + "onnx::MatMul_8483_45_1_71": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1758579712, + "file_name": "cache/pos_embedprojConv_469.const", + "file_size": 10764288 + }, + "transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1769344000, + "file_name": "cache/pos_embedprojConv_470.const", + "file_size": 2691072 + }, + "onnx::MatMul_8652_45_0_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1772035072, + "file_name": "cache/pos_embedprojConv_471.const", + "file_size": 2691072 + }, + "onnx::MatMul_8480_45_5_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1774726144, + "file_name": "cache/pos_embedprojConv_472.const", + "file_size": 2691072 + }, + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1777417216, + "file_name": "cache/pos_embedprojConv_473.const", + "file_size": 2691072 + }, + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1780108288, + "file_name": "cache/pos_embedprojConv_474.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1780111360, + "file_name": "cache/pos_embedprojConv_475.const", + "file_size": 3072 + }, + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_3_transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1780114432, + "file_name": "cache/pos_embedprojConv_476.const", + "file_size": 5382144 + }, + "onnx::MatMul_8484_45_1_70": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1785496576, + "file_name": "cache/pos_embedprojConv_477.const", + "file_size": 10764288 + }, + "onnx::MatMul_8485_45_1_72": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1796260864, + "file_name": "cache/pos_embedprojConv_478.const", + "file_size": 10764288 + }, + "transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1807025152, + "file_name": "cache/pos_embedprojConv_479.const", + "file_size": 2691072 + }, + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_72": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1809716224, + "file_name": "cache/pos_embedprojConv_480.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_72": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1809719296, + "file_name": "cache/pos_embedprojConv_481.const", + "file_size": 3072 + }, + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_0_transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1809722368, + "file_name": "cache/pos_embedprojConv_482.const", + "file_size": 5382144 + }, + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_73": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1815104512, + "file_name": "cache/pos_embedprojConv_483.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_73": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1815107584, + "file_name": "cache/pos_embedprojConv_484.const", + "file_size": 3072 + }, + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_0_transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1815110656, + "file_name": "cache/pos_embedprojConv_485.const", + "file_size": 5382144 + }, + "onnx::MatMul_8501_onnx::MatMul_8486": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1820492800, + "file_name": "cache/pos_embedprojConv_486.const", + "file_size": 5382144 + }, + "onnx::MatMul_8502_onnx::MatMul_8487": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1825874944, + "file_name": "cache/pos_embedprojConv_487.const", + "file_size": 5382144 + }, + "onnx::MatMul_8503_onnx::MatMul_8488": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1831257088, + "file_name": "cache/pos_embedprojConv_488.const", + "file_size": 5382144 + }, + "onnx::MatMul_8512_45_5_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1836639232, + "file_name": "cache/pos_embedprojConv_489.const", + "file_size": 2691072 + }, + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1839330304, + "file_name": "cache/pos_embedprojConv_490.const", + "file_size": 2691072 + }, + "/transformer_blocks.18/norm2/Constant_output_0_14_74": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1842021376, + "file_name": "cache/pos_embedprojConv_491.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1842024448, + "file_name": "cache/pos_embedprojConv_492.const", + "file_size": 3072 + }, + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_3_transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1842027520, + "file_name": "cache/pos_embedprojConv_493.const", + "file_size": 5382144 + }, + "onnx::MatMul_8513_45_1_73": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1847409664, + "file_name": "cache/pos_embedprojConv_494.const", + "file_size": 10764288 + }, + "onnx::MatMul_8514_45_1_75": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1858173952, + "file_name": "cache/pos_embedprojConv_495.const", + "file_size": 10764288 + }, + "transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1868938240, + "file_name": "cache/pos_embedprojConv_496.const", + "file_size": 2691072 + }, + "onnx::MatMul_8653_45_0_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1871629312, + "file_name": "cache/pos_embedprojConv_497.const", + "file_size": 2691072 + }, + "onnx::MatMul_8511_45_5_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1874320384, + "file_name": "cache/pos_embedprojConv_498.const", + "file_size": 2691072 + }, + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1877011456, + "file_name": "cache/pos_embedprojConv_499.const", + "file_size": 2691072 + }, + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1879702528, + "file_name": "cache/pos_embedprojConv_500.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1879705600, + "file_name": "cache/pos_embedprojConv_501.const", + "file_size": 3072 + }, + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_3_transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1879708672, + "file_name": "cache/pos_embedprojConv_502.const", + "file_size": 5382144 + }, + "onnx::MatMul_8515_45_1_74": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1885090816, + "file_name": "cache/pos_embedprojConv_503.const", + "file_size": 10764288 + }, + "onnx::MatMul_8516_45_1_76": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1895855104, + "file_name": "cache/pos_embedprojConv_504.const", + "file_size": 10764288 + }, + "transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1906619392, + "file_name": "cache/pos_embedprojConv_505.const", + "file_size": 2691072 + }, + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_76": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1909310464, + "file_name": "cache/pos_embedprojConv_506.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_76": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1909313536, + "file_name": "cache/pos_embedprojConv_507.const", + "file_size": 3072 + }, + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_0_transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1909316608, + "file_name": "cache/pos_embedprojConv_508.const", + "file_size": 5382144 + }, + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_77": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1914698752, + "file_name": "cache/pos_embedprojConv_509.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_77": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1914701824, + "file_name": "cache/pos_embedprojConv_510.const", + "file_size": 3072 + }, + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_0_transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1914704896, + "file_name": "cache/pos_embedprojConv_511.const", + "file_size": 5382144 + }, + "onnx::MatMul_8532_onnx::MatMul_8517": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1920087040, + "file_name": "cache/pos_embedprojConv_512.const", + "file_size": 5382144 + }, + "onnx::MatMul_8533_onnx::MatMul_8518": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1925469184, + "file_name": "cache/pos_embedprojConv_513.const", + "file_size": 5382144 + }, + "onnx::MatMul_8534_onnx::MatMul_8519": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1930851328, + "file_name": "cache/pos_embedprojConv_514.const", + "file_size": 5382144 + }, + "onnx::MatMul_8543_45_5_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1936233472, + "file_name": "cache/pos_embedprojConv_515.const", + "file_size": 2691072 + }, + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1938924544, + "file_name": "cache/pos_embedprojConv_516.const", + "file_size": 2691072 + }, + "/transformer_blocks.19/norm2/Constant_output_0_14_78": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1941615616, + "file_name": "cache/pos_embedprojConv_517.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1941618688, + "file_name": "cache/pos_embedprojConv_518.const", + "file_size": 3072 + }, + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_3_transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1941621760, + "file_name": "cache/pos_embedprojConv_519.const", + "file_size": 5382144 + }, + "onnx::MatMul_8544_45_1_77": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1947003904, + "file_name": "cache/pos_embedprojConv_520.const", + "file_size": 10764288 + }, + "onnx::MatMul_8545_45_1_79": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1957768192, + "file_name": "cache/pos_embedprojConv_521.const", + "file_size": 10764288 + }, + "transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1968532480, + "file_name": "cache/pos_embedprojConv_522.const", + "file_size": 2691072 + }, + "onnx::MatMul_8654_45_0_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1971223552, + "file_name": "cache/pos_embedprojConv_523.const", + "file_size": 2691072 + }, + "onnx::MatMul_8542_45_5_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1973914624, + "file_name": "cache/pos_embedprojConv_524.const", + "file_size": 2691072 + }, + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1976605696, + "file_name": "cache/pos_embedprojConv_525.const", + "file_size": 2691072 + }, + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1979296768, + "file_name": "cache/pos_embedprojConv_526.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1979299840, + "file_name": "cache/pos_embedprojConv_527.const", + "file_size": 3072 + }, + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_3_transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1979302912, + "file_name": "cache/pos_embedprojConv_528.const", + "file_size": 5382144 + }, + "onnx::MatMul_8546_45_1_78": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1984685056, + "file_name": "cache/pos_embedprojConv_529.const", + "file_size": 10764288 + }, + "onnx::MatMul_8547_45_1_80": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1995449344, + "file_name": "cache/pos_embedprojConv_530.const", + "file_size": 10764288 + }, + "transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2006213632, + "file_name": "cache/pos_embedprojConv_531.const", + "file_size": 2691072 + }, + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_80": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2008904704, + "file_name": "cache/pos_embedprojConv_532.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_80": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2008907776, + "file_name": "cache/pos_embedprojConv_533.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_81": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2008910848, + "file_name": "cache/pos_embedprojConv_534.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_81": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2008913920, + "file_name": "cache/pos_embedprojConv_535.const", + "file_size": 3072 + }, + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_0_transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2008916992, + "file_name": "cache/pos_embedprojConv_536.const", + "file_size": 5382144 + }, + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_0_transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2014299136, + "file_name": "cache/pos_embedprojConv_537.const", + "file_size": 5382144 + }, + "onnx::MatMul_8563_onnx::MatMul_8548": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2019681280, + "file_name": "cache/pos_embedprojConv_538.const", + "file_size": 5382144 + }, + "onnx::MatMul_8564_onnx::MatMul_8549": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2025063424, + "file_name": "cache/pos_embedprojConv_539.const", + "file_size": 5382144 + }, + "onnx::MatMul_8565_onnx::MatMul_8550": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2030445568, + "file_name": "cache/pos_embedprojConv_540.const", + "file_size": 5382144 + }, + "onnx::MatMul_8574_45_5_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2035827712, + "file_name": "cache/pos_embedprojConv_541.const", + "file_size": 2691072 + }, + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2038518784, + "file_name": "cache/pos_embedprojConv_542.const", + "file_size": 2691072 + }, + "/transformer_blocks.20/norm2/Constant_output_0_14_82": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2041209856, + "file_name": "cache/pos_embedprojConv_543.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2041212928, + "file_name": "cache/pos_embedprojConv_544.const", + "file_size": 3072 + }, + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_3_transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2041216000, + "file_name": "cache/pos_embedprojConv_545.const", + "file_size": 5382144 + }, + "onnx::MatMul_8575_45_1_81": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2046598144, + "file_name": "cache/pos_embedprojConv_546.const", + "file_size": 10764288 + }, + "onnx::MatMul_8576_45_1_83": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2057362432, + "file_name": "cache/pos_embedprojConv_547.const", + "file_size": 10764288 + }, + "transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2068126720, + "file_name": "cache/pos_embedprojConv_548.const", + "file_size": 2691072 + }, + "onnx::MatMul_8655_45_0_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2070817792, + "file_name": "cache/pos_embedprojConv_549.const", + "file_size": 2691072 + }, + "onnx::MatMul_8573_45_5_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2073508864, + "file_name": "cache/pos_embedprojConv_550.const", + "file_size": 2691072 + }, + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2076199936, + "file_name": "cache/pos_embedprojConv_551.const", + "file_size": 2691072 + }, + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2078891008, + "file_name": "cache/pos_embedprojConv_552.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2078894080, + "file_name": "cache/pos_embedprojConv_553.const", + "file_size": 3072 + }, + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_3_transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2078897152, + "file_name": "cache/pos_embedprojConv_554.const", + "file_size": 5382144 + }, + "onnx::MatMul_8577_45_1_82": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2084279296, + "file_name": "cache/pos_embedprojConv_555.const", + "file_size": 10764288 + }, + "onnx::MatMul_8578_45_1_84": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2095043584, + "file_name": "cache/pos_embedprojConv_556.const", + "file_size": 10764288 + }, + "transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2105807872, + "file_name": "cache/pos_embedprojConv_557.const", + "file_size": 2691072 + }, + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_84": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2108498944, + "file_name": "cache/pos_embedprojConv_558.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_84": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2108502016, + "file_name": "cache/pos_embedprojConv_559.const", + "file_size": 3072 + }, + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_0_transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2108505088, + "file_name": "cache/pos_embedprojConv_560.const", + "file_size": 5382144 + }, + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_85": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2113887232, + "file_name": "cache/pos_embedprojConv_561.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_85": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2113890304, + "file_name": "cache/pos_embedprojConv_562.const", + "file_size": 3072 + }, + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_0_transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2113893376, + "file_name": "cache/pos_embedprojConv_563.const", + "file_size": 5382144 + }, + "onnx::MatMul_8594_onnx::MatMul_8579": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2119275520, + "file_name": "cache/pos_embedprojConv_564.const", + "file_size": 5382144 + }, + "onnx::MatMul_8595_onnx::MatMul_8580": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2124657664, + "file_name": "cache/pos_embedprojConv_565.const", + "file_size": 5382144 + }, + "onnx::MatMul_8596_onnx::MatMul_8581": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2130039808, + "file_name": "cache/pos_embedprojConv_566.const", + "file_size": 5382144 + }, + "onnx::MatMul_8605_45_5_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2135421952, + "file_name": "cache/pos_embedprojConv_567.const", + "file_size": 2691072 + }, + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2138113024, + "file_name": "cache/pos_embedprojConv_568.const", + "file_size": 2691072 + }, + "/transformer_blocks.21/norm2/Constant_output_0_14_86": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2140804096, + "file_name": "cache/pos_embedprojConv_569.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2140807168, + "file_name": "cache/pos_embedprojConv_570.const", + "file_size": 3072 + }, + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_3_transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2140810240, + "file_name": "cache/pos_embedprojConv_571.const", + "file_size": 5382144 + }, + "onnx::MatMul_8606_45_1_85": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2146192384, + "file_name": "cache/pos_embedprojConv_572.const", + "file_size": 10764288 + }, + "onnx::MatMul_8607_45_1_87": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2156956672, + "file_name": "cache/pos_embedprojConv_573.const", + "file_size": 10764288 + }, + "transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2167720960, + "file_name": "cache/pos_embedprojConv_574.const", + "file_size": 2691072 + }, + "onnx::MatMul_8656_45_0_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2170412032, + "file_name": "cache/pos_embedprojConv_575.const", + "file_size": 2691072 + }, + "onnx::MatMul_8604_45_5_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2173103104, + "file_name": "cache/pos_embedprojConv_576.const", + "file_size": 2691072 + }, + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2175794176, + "file_name": "cache/pos_embedprojConv_577.const", + "file_size": 2691072 + }, + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2178485248, + "file_name": "cache/pos_embedprojConv_578.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2178488320, + "file_name": "cache/pos_embedprojConv_579.const", + "file_size": 3072 + }, + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_3_transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2178491392, + "file_name": "cache/pos_embedprojConv_580.const", + "file_size": 5382144 + }, + "onnx::MatMul_8608_45_1_86": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2183873536, + "file_name": "cache/pos_embedprojConv_581.const", + "file_size": 10764288 + }, + "onnx::MatMul_8609_45_1_88": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2194637824, + "file_name": "cache/pos_embedprojConv_582.const", + "file_size": 10764288 + }, + "transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2205402112, + "file_name": "cache/pos_embedprojConv_583.const", + "file_size": 2691072 + }, + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_88": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2208093184, + "file_name": "cache/pos_embedprojConv_584.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_88": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2208096256, + "file_name": "cache/pos_embedprojConv_585.const", + "file_size": 3072 + }, + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_0_transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2208099328, + "file_name": "cache/pos_embedprojConv_586.const", + "file_size": 5382144 + }, + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_89": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2213481472, + "file_name": "cache/pos_embedprojConv_587.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_89": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2213484544, + "file_name": "cache/pos_embedprojConv_588.const", + "file_size": 3072 + }, + "transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_0_transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2213487616, + "file_name": "cache/pos_embedprojConv_589.const", + "file_size": 5382144 + }, + "onnx::MatMul_8625_onnx::MatMul_8610": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2218869760, + "file_name": "cache/pos_embedprojConv_590.const", + "file_size": 5382144 + }, + "onnx::MatMul_8626_onnx::MatMul_8611": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2224251904, + "file_name": "cache/pos_embedprojConv_591.const", + "file_size": 5382144 + }, + "onnx::MatMul_8627_onnx::MatMul_8612": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2229634048, + "file_name": "cache/pos_embedprojConv_592.const", + "file_size": 5382144 + }, + "onnx::MatMul_8632_45_5_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2235016192, + "file_name": "cache/pos_embedprojConv_593.const", + "file_size": 2691072 + }, + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2237707264, + "file_name": "cache/pos_embedprojConv_594.const", + "file_size": 2691072 + }, + "/transformer_blocks.22/norm2/Constant_output_0_14_90": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2240398336, + "file_name": "cache/pos_embedprojConv_595.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2240401408, + "file_name": "cache/pos_embedprojConv_596.const", + "file_size": 3072 + }, + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_3_transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2240404480, + "file_name": "cache/pos_embedprojConv_597.const", + "file_size": 5382144 + }, + "onnx::MatMul_8633_45_1_89": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2245786624, + "file_name": "cache/pos_embedprojConv_598.const", + "file_size": 10764288 + }, + "onnx::MatMul_8634_45_1_90": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2256550912, + "file_name": "cache/pos_embedprojConv_599.const", + "file_size": 10764288 + }, + "transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2267315200, + "file_name": "cache/pos_embedprojConv_600.const", + "file_size": 2691072 + }, + "onnx::MatMul_8657_45_0_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2270006272, + "file_name": "cache/pos_embedprojConv_601.const", + "file_size": 2691072 + } + }, + "dynamic_shape_subgraph": true, + "dynamic_shape_list": [ + { + "floor(w/2)": 64, + "w": 128, + "max_length + floor(h/2)*floor(w/2)": 4256, + "max_length": 160, + "h": 128, + "batch_size": 2, + "floor(h/2)*floor(w/2)+max_length": 4256, + "floor(h/2)*floor(w/2)": 4096, + "floor(h/2)": 64 + } + ], + "aux_info": {} +} \ No newline at end of file diff --git a/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..8bfe068e97152f191694cad2675100b2947f13e4 --- /dev/null +++ b/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cdb1b50e1b4dec6d0557c6a211dfc7cd8f2df59c31a56fd90b1627ffc7a1d65 +size 197375827 diff --git a/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst b/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst new file mode 100644 index 0000000000000000000000000000000000000000..c230595ac9531f5a0bd14945650fef5c63ccf77e --- /dev/null +++ b/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13c16301d23e99a2b2853f1babdaf712ac992e89e3aff9a1e11d39241a2cdd47 +size 2272696960 diff --git a/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state new file mode 100644 index 0000000000000000000000000000000000000000..486366e063a3059e711f059a5593de2d7d8ce2f0 Binary files /dev/null and b/inpainting/controlnet-inpainting/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/inpainting/controlnet-inpainting/dynamic/dd/replaced.onnx b/inpainting/controlnet-inpainting/dynamic/dd/replaced.onnx new file mode 100644 index 0000000000000000000000000000000000000000..354489fa9617c86b99d72a5b2e19f5a9bb540a30 --- /dev/null +++ b/inpainting/controlnet-inpainting/dynamic/dd/replaced.onnx @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d92d6ddc5709085358971a413ea04bfda5a24c3961bc187f8fae3acf1917eb97 +size 1028294 diff --git a/inpainting/controlnet-inpainting/dynamic/dd/replaced.onnx.data b/inpainting/controlnet-inpainting/dynamic/dd/replaced.onnx.data new file mode 100644 index 0000000000000000000000000000000000000000..7530b7257c754e617c16e4cfcbfa7ee310fe7811 --- /dev/null +++ b/inpainting/controlnet-inpainting/dynamic/dd/replaced.onnx.data @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:850d66e980fbceb846959c026f2ad96fc5fda8c0402866170acf95dfa534be13 +size 113246208 diff --git a/inpainting/transformer/config.json b/inpainting/transformer/config.json new file mode 100644 index 0000000000000000000000000000000000000000..32d9400026677525c71337b7d26d6212c0c3ddda --- /dev/null +++ b/inpainting/transformer/config.json @@ -0,0 +1,18 @@ +{ + "_class_name": "SD3Transformer2DModel", + "_diffusers_version": "0.33.0", + "_name_or_path": "C:\\Users\\AMD\\.cache\\huggingface\\hub\\models--stabilityai--stable-diffusion-3-medium-diffusers\\snapshots\\ea42f8cef0f178587cf766dc8129abd379c90671\\transformer", + "attention_head_dim": 64, + "caption_projection_dim": 1536, + "dual_attention_layers": [], + "in_channels": 16, + "joint_attention_dim": 4096, + "num_attention_heads": 24, + "num_layers": 24, + "out_channels": 16, + "patch_size": 2, + "pooled_projection_dim": 2048, + "pos_embed_max_size": 192, + "qk_norm": null, + "sample_size": 128 +} diff --git a/inpainting/transformer/dynamic/dd/cache/pos_embedprojConv_meta.json b/inpainting/transformer/dynamic/dd/cache/pos_embedprojConv_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..ded2b415a74e0b768a5987c4d11a7888ec848b65 --- /dev/null +++ b/inpainting/transformer/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -0,0 +1,65564 @@ +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 4, + "state_table_updates": [], + "op_list": [ + { + "name": "/pos_embed/proj/Conv", + "type": "SDConv", + "in_args": [ + "hidden_states_nhwc.out5_0_0" + ], + "const_args": [ + "existing_model.pos_embed.proj.weight" + ], + "out_args": [ + "/pos_embed/Transpose_output_0.out5_0_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "w", + "h", + "16" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "floor(w/2)", + "floor(h/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "2", + "2", + "16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/pos_embed/Add_2", + "type": "SDAdd", + "in_args": [ + "/pos_embed/Transpose_output_0.out5_0_0", + "/pos_embed/Reshape_1_output_0.out_35_1_2" + ], + "const_args": [], + "out_args": [ + "/pos_embed/Add_2_output_0.out_35_1_2" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/Cast_output_0.out17_4_3" + ], + "const_args": [ + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_45_0_2" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/act/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" + ], + "const_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" + ], + "const_args": [ + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_45_0_3" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "pooled_projections.out17_4_1" + ], + "const_args": [ + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_45_0_0" + ], + "out_args": [ + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "2048" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/act_1/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" + ], + "const_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0" + ], + "out_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" + ], + "const_args": [ + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_45_0_1" + ], + "out_args": [ + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/Add", + "type": "SDAdd", + "in_args": [ + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" + ], + "const_args": [], + "out_args": [ + "/time_text_embed/Add_output_0.out_35_1_3" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/silu/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/Add_output_0.out_35_1_3" + ], + "const_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2" + ], + "out_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "encoder_hidden_states.out17_4_0_SDCastBf2Bfp", + "type": "SDCastBf2Bfp", + "in_args": [ + "encoder_hidden_states.out17_4_0" + ], + "const_args": [ + "encoder_hidden_states.out17_4_0_bfp.wts" + ], + "out_args": [ + "encoder_hidden_states.out17_4_0_bfp.out25_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "4096" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "4096" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/context_embedder/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "encoder_hidden_states.out17_4_0_bfp.out25_0" + ], + "const_args": [ + "onnx::MatMul_8733_45_1_0" + ], + "out_args": [ + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "4096" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "4096", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_SDCastBfp2Bf", + "type": "SDCastBfp2Bf", + "in_args": [ + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" + ], + "const_args": [ + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts" + ], + "out_args": [ + "/context_embedder/Add_output_0.out17_4_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/context_embedder/Add_output_0.out17_4_0" + ], + "const_args": [ + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0" + ], + "out_args": [ + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1" + ], + "out_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/pos_embed/Add_2_output_0.out_35_1_2" + ], + "const_args": [ + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1" + ], + "out_args": [ + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1" + ], + "out_args": [ + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_94" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_94" + ], + "const_args": [ + "onnx::MatMul_8749_onnx::MatMul_8734" + ], + "out_args": [ + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_94" + ], + "const_args": [ + "onnx::MatMul_8750_onnx::MatMul_8735" + ], + "out_args": [ + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_94" + ], + "const_args": [ + "onnx::MatMul_8751_onnx::MatMul_8736" + ], + "out_args": [ + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/MatMulmha_18_0_0", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0", + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1", + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0" + ], + "const_args": [ + "onnx::MatMul_8759_45_5_1" + ], + "out_args": [ + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", + "/context_embedder/Add_output_0.out17_4_0" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_4_output_0.out10_0" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.0/Add_4_output_0.out10_0" + ], + "const_args": [ + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3" + ], + "out_args": [ + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4" + ], + "out_args": [ + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" + ], + "const_args": [ + "onnx::MatMul_8763_45_1_2" + ], + "out_args": [ + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" + ], + "const_args": [ + "onnx::MatMul_8764_45_1_4" + ], + "out_args": [ + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", + "/transformer_blocks.0/Add_4_output_0.out10_0" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_7_output_0.out10_1" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0" + ], + "const_args": [ + "onnx::MatMul_8760_45_5_0" + ], + "out_args": [ + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", + "/pos_embed/Add_2_output_0.out_35_1_2" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_output_0.out10_92" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.0/Add_output_0.out10_92" + ], + "const_args": [ + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2" + ], + "out_args": [ + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4" + ], + "out_args": [ + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_95" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_95" + ], + "const_args": [ + "onnx::MatMul_8761_45_1_1" + ], + "out_args": [ + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" + ], + "const_args": [ + "onnx::MatMul_8762_45_1_3" + ], + "out_args": [ + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", + "/transformer_blocks.0/Add_output_0.out10_92" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma" + ], + "out_args": [ + "/transformer_blocks.0/Add_3_output_0.out10_93" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.0/Add_7_output_0.out10_1" + ], + "const_args": [ + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_4" + ], + "out_args": [ + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.0/Add_3_output_0.out10_93", + "block_controlnet_hidden_states_0.out_35_1_4" + ], + "const_args": [], + "out_args": [ + "/Add_output_0.out_35_1_4" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_output_0.out_35_1_4" + ], + "const_args": [ + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_5" + ], + "out_args": [ + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1" + ], + "out_args": [ + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_5_bfp.out1_2" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1" + ], + "out_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_4_bfp.out1_4" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_4_bfp.out1_4", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_5_bfp.out1_2" + ], + "const_args": [ + "onnx::MatMul_8780_onnx::MatMul_8765" + ], + "out_args": [ + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_4_bfp.out1_4", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_5_bfp.out1_2" + ], + "const_args": [ + "onnx::MatMul_8781_onnx::MatMul_8766" + ], + "out_args": [ + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_4_bfp.out1_4", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_5_bfp.out1_2" + ], + "const_args": [ + "onnx::MatMul_8782_onnx::MatMul_8767" + ], + "out_args": [ + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/MatMulmha_18_0_1", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3", + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4", + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1" + ], + "const_args": [ + "onnx::MatMul_8791_45_5_2" + ], + "out_args": [ + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2", + "/Add_output_0.out_35_1_4" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_output_0.out10_2" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.1/Add_output_0.out10_2" + ], + "const_args": [ + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6" + ], + "out_args": [ + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4" + ], + "out_args": [ + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" + ], + "const_args": [ + "onnx::MatMul_8792_45_1_5" + ], + "out_args": [ + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_9" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_9" + ], + "const_args": [ + "onnx::MatMul_8793_45_1_7" + ], + "out_args": [ + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11", + "/transformer_blocks.1/Add_output_0.out10_2" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_3_output_0.out10_3" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1" + ], + "const_args": [ + "onnx::MatMul_8790_45_5_3" + ], + "out_args": [ + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", + "/transformer_blocks.0/Add_7_output_0.out10_1" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_4_output_0.out10_4" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.1/Add_4_output_0.out10_4" + ], + "const_args": [ + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7" + ], + "out_args": [ + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4" + ], + "out_args": [ + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" + ], + "const_args": [ + "onnx::MatMul_8794_45_1_6" + ], + "out_args": [ + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_10" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_10" + ], + "const_args": [ + "onnx::MatMul_8795_45_1_8" + ], + "out_args": [ + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.1/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12", + "/transformer_blocks.1/Add_4_output_0.out10_4" + ], + "const_args": [ + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma" + ], + "out_args": [ + "/transformer_blocks.1/Add_7_output_0.out10_5" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.1/Add_7_output_0.out10_5" + ], + "const_args": [ + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_8" + ], + "out_args": [ + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_1", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.1/Add_3_output_0.out10_3", + "block_controlnet_hidden_states_1.out_35_1_5" + ], + "const_args": [], + "out_args": [ + "/Add_1_output_0.out_35_1_5" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_1_output_0.out_35_1_5" + ], + "const_args": [ + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_9" + ], + "out_args": [ + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1" + ], + "out_args": [ + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_9_bfp.out1_46" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1" + ], + "out_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_8_bfp.out1_48" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_8_bfp.out1_48", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_9_bfp.out1_46" + ], + "const_args": [ + "onnx::MatMul_8811_onnx::MatMul_8796" + ], + "out_args": [ + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_8_bfp.out1_48", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_9_bfp.out1_46" + ], + "const_args": [ + "onnx::MatMul_8812_onnx::MatMul_8797" + ], + "out_args": [ + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_8_bfp.out1_48", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_9_bfp.out1_46" + ], + "const_args": [ + "onnx::MatMul_8813_onnx::MatMul_8798" + ], + "out_args": [ + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/MatMulmha_18_0_2", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6", + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7", + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2" + ], + "const_args": [ + "onnx::MatMul_8821_45_5_5" + ], + "out_args": [ + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5", + "/transformer_blocks.1/Add_7_output_0.out10_5" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_4_output_0.out10_48" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2" + ], + "const_args": [ + "onnx::MatMul_8822_45_5_4" + ], + "out_args": [ + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4", + "/Add_1_output_0.out_35_1_5" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_output_0.out10_46" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.2/Add_output_0.out10_46" + ], + "const_args": [ + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10" + ], + "out_args": [ + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4" + ], + "out_args": [ + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47" + ], + "const_args": [ + "onnx::MatMul_8823_45_1_9" + ], + "out_args": [ + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_13" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_13" + ], + "const_args": [ + "onnx::MatMul_8824_45_1_11" + ], + "out_args": [ + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.2/Add_4_output_0.out10_48" + ], + "const_args": [ + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11" + ], + "out_args": [ + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4" + ], + "out_args": [ + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49" + ], + "const_args": [ + "onnx::MatMul_8825_45_1_10" + ], + "out_args": [ + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_14" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_14" + ], + "const_args": [ + "onnx::MatMul_8826_45_1_12" + ], + "out_args": [ + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16", + "/transformer_blocks.2/Add_4_output_0.out10_48" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_7_output_0.out10_49" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.2/Add_7_output_0.out10_49" + ], + "const_args": [ + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_12" + ], + "out_args": [ + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.2/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15", + "/transformer_blocks.2/Add_output_0.out10_46" + ], + "const_args": [ + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma" + ], + "out_args": [ + "/transformer_blocks.2/Add_3_output_0.out10_47" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_2", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.2/Add_3_output_0.out10_47", + "block_controlnet_hidden_states_2.out_35_1_6" + ], + "const_args": [], + "out_args": [ + "/Add_2_output_0.out_35_1_6" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_2_output_0.out_35_1_6" + ], + "const_args": [ + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_13" + ], + "out_args": [ + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1" + ], + "out_args": [ + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_13_bfp.out1_65" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1" + ], + "out_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_12_bfp.out1_67" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_12_bfp.out1_67", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_13_bfp.out1_65" + ], + "const_args": [ + "onnx::MatMul_8842_onnx::MatMul_8827" + ], + "out_args": [ + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_12_bfp.out1_67", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_13_bfp.out1_65" + ], + "const_args": [ + "onnx::MatMul_8843_onnx::MatMul_8828" + ], + "out_args": [ + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_12_bfp.out1_67", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_13_bfp.out1_65" + ], + "const_args": [ + "onnx::MatMul_8844_onnx::MatMul_8829" + ], + "out_args": [ + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/MatMulmha_18_0_3", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9", + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10", + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3" + ], + "const_args": [ + "onnx::MatMul_8852_45_5_7" + ], + "out_args": [ + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7", + "/transformer_blocks.2/Add_7_output_0.out10_49" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_4_output_0.out10_66" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3" + ], + "const_args": [ + "onnx::MatMul_8853_45_5_6" + ], + "out_args": [ + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6", + "/Add_2_output_0.out_35_1_6" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_output_0.out10_64" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.3/Add_output_0.out10_64" + ], + "const_args": [ + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14" + ], + "out_args": [ + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4" + ], + "out_args": [ + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_66" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_66" + ], + "const_args": [ + "onnx::MatMul_8854_45_1_13" + ], + "out_args": [ + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_17" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_17" + ], + "const_args": [ + "onnx::MatMul_8855_45_1_15" + ], + "out_args": [ + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.3/Add_4_output_0.out10_66" + ], + "const_args": [ + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15" + ], + "out_args": [ + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4" + ], + "out_args": [ + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_68" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_68" + ], + "const_args": [ + "onnx::MatMul_8856_45_1_14" + ], + "out_args": [ + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_18" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_18" + ], + "const_args": [ + "onnx::MatMul_8857_45_1_16" + ], + "out_args": [ + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20", + "/transformer_blocks.3/Add_4_output_0.out10_66" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_7_output_0.out10_67" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.3/Add_7_output_0.out10_67" + ], + "const_args": [ + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_16" + ], + "out_args": [ + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.3/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19", + "/transformer_blocks.3/Add_output_0.out10_64" + ], + "const_args": [ + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma" + ], + "out_args": [ + "/transformer_blocks.3/Add_3_output_0.out10_65" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_3", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.3/Add_3_output_0.out10_65", + "block_controlnet_hidden_states_3.out_35_1_7" + ], + "const_args": [], + "out_args": [ + "/Add_3_output_0.out_35_1_7" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_3_output_0.out_35_1_7" + ], + "const_args": [ + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_17" + ], + "out_args": [ + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1" + ], + "out_args": [ + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_17_bfp.out1_69" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1" + ], + "out_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_16_bfp.out1_71" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_16_bfp.out1_71", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_17_bfp.out1_69" + ], + "const_args": [ + "onnx::MatMul_8873_onnx::MatMul_8858" + ], + "out_args": [ + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_16_bfp.out1_71", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_17_bfp.out1_69" + ], + "const_args": [ + "onnx::MatMul_8874_onnx::MatMul_8859" + ], + "out_args": [ + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_16_bfp.out1_71", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_17_bfp.out1_69" + ], + "const_args": [ + "onnx::MatMul_8875_onnx::MatMul_8860" + ], + "out_args": [ + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/MatMulmha_18_0_4", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12", + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13", + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4" + ], + "const_args": [ + "onnx::MatMul_8883_45_5_9" + ], + "out_args": [ + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9", + "/transformer_blocks.3/Add_7_output_0.out10_67" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_4_output_0.out10_70" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4" + ], + "const_args": [ + "onnx::MatMul_8884_45_5_8" + ], + "out_args": [ + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8", + "/Add_3_output_0.out_35_1_7" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_output_0.out10_68" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.4/Add_output_0.out10_68" + ], + "const_args": [ + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18" + ], + "out_args": [ + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4" + ], + "out_args": [ + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_70" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_70" + ], + "const_args": [ + "onnx::MatMul_8885_45_1_17" + ], + "out_args": [ + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_21" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_21" + ], + "const_args": [ + "onnx::MatMul_8886_45_1_19" + ], + "out_args": [ + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.4/Add_4_output_0.out10_70" + ], + "const_args": [ + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19" + ], + "out_args": [ + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4" + ], + "out_args": [ + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_72" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_72" + ], + "const_args": [ + "onnx::MatMul_8887_45_1_18" + ], + "out_args": [ + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_22" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_22" + ], + "const_args": [ + "onnx::MatMul_8888_45_1_20" + ], + "out_args": [ + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24", + "/transformer_blocks.4/Add_4_output_0.out10_70" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_7_output_0.out10_71" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.4/Add_7_output_0.out10_71" + ], + "const_args": [ + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_20" + ], + "out_args": [ + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.4/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23", + "/transformer_blocks.4/Add_output_0.out10_68" + ], + "const_args": [ + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma" + ], + "out_args": [ + "/transformer_blocks.4/Add_3_output_0.out10_69" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_4", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.4/Add_3_output_0.out10_69", + "block_controlnet_hidden_states_4.out_35_1_8" + ], + "const_args": [], + "out_args": [ + "/Add_4_output_0.out_35_1_8" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_4_output_0.out_35_1_8" + ], + "const_args": [ + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_21" + ], + "out_args": [ + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1" + ], + "out_args": [ + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_21_bfp.out1_73" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1" + ], + "out_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_20_bfp.out1_75" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_20_bfp.out1_75", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_21_bfp.out1_73" + ], + "const_args": [ + "onnx::MatMul_8904_onnx::MatMul_8889" + ], + "out_args": [ + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_20_bfp.out1_75", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_21_bfp.out1_73" + ], + "const_args": [ + "onnx::MatMul_8905_onnx::MatMul_8890" + ], + "out_args": [ + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_20_bfp.out1_75", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_21_bfp.out1_73" + ], + "const_args": [ + "onnx::MatMul_8906_onnx::MatMul_8891" + ], + "out_args": [ + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/MatMulmha_18_0_5", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15", + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16", + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5" + ], + "const_args": [ + "onnx::MatMul_8914_45_5_11" + ], + "out_args": [ + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11", + "/transformer_blocks.4/Add_7_output_0.out10_71" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_2_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_4_output_0.out10_74" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5" + ], + "const_args": [ + "onnx::MatMul_8915_45_5_10" + ], + "out_args": [ + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10", + "/Add_4_output_0.out_35_1_8" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_output_0.out10_72" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.5/Add_output_0.out10_72" + ], + "const_args": [ + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22" + ], + "out_args": [ + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4" + ], + "out_args": [ + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_74" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_74" + ], + "const_args": [ + "onnx::MatMul_8916_45_1_21" + ], + "out_args": [ + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_25" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_25" + ], + "const_args": [ + "onnx::MatMul_8917_45_1_23" + ], + "out_args": [ + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.5/Add_4_output_0.out10_74" + ], + "const_args": [ + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23" + ], + "out_args": [ + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_3_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_4" + ], + "out_args": [ + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_76" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_76" + ], + "const_args": [ + "onnx::MatMul_8918_45_1_22" + ], + "out_args": [ + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_26" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_26" + ], + "const_args": [ + "onnx::MatMul_8919_45_1_24" + ], + "out_args": [ + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28", + "/transformer_blocks.5/Add_4_output_0.out10_74" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_5_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_7_output_0.out10_75" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.5/Add_7_output_0.out10_75" + ], + "const_args": [ + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_24" + ], + "out_args": [ + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_24_bfp.out15_24" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.5/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27", + "/transformer_blocks.5/Add_output_0.out10_72" + ], + "const_args": [ + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma" + ], + "out_args": [ + "/transformer_blocks.5/Add_3_output_0.out10_73" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_5", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.5/Add_3_output_0.out10_73", + "block_controlnet_hidden_states_5.out_35_1_9" + ], + "const_args": [], + "out_args": [ + "/Add_5_output_0.out_35_1_9" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_5_output_0.out_35_1_9" + ], + "const_args": [ + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_25" + ], + "out_args": [ + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_25_bfp.out15_25" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_25_bfp.out15_25" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_0_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_1" + ], + "out_args": [ + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_25_bfp.out1_77" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_24_bfp.out15_24" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_0_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_1" + ], + "out_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_24_bfp.out1_79" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_24_bfp.out1_79", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_25_bfp.out1_77" + ], + "const_args": [ + "onnx::MatMul_8935_onnx::MatMul_8920" + ], + "out_args": [ + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_24_bfp.out1_79", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_25_bfp.out1_77" + ], + "const_args": [ + "onnx::MatMul_8936_onnx::MatMul_8921" + ], + "out_args": [ + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_24_bfp.out1_79", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_25_bfp.out1_77" + ], + "const_args": [ + "onnx::MatMul_8937_onnx::MatMul_8922" + ], + "out_args": [ + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/MatMulmha_18_0_6", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18", + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19", + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6" + ], + "const_args": [ + "onnx::MatMul_8945_45_5_13" + ], + "out_args": [ + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13", + "/transformer_blocks.5/Add_7_output_0.out10_75" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_2_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_4_output_0.out10_78" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6" + ], + "const_args": [ + "onnx::MatMul_8946_45_5_12" + ], + "out_args": [ + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12", + "/Add_5_output_0.out_35_1_9" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_2_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_output_0.out10_76" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.6/Add_output_0.out10_76" + ], + "const_args": [ + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26" + ], + "out_args": [ + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_3_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_4" + ], + "out_args": [ + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_78" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_78" + ], + "const_args": [ + "onnx::MatMul_8947_45_1_25" + ], + "out_args": [ + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_29" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_29" + ], + "const_args": [ + "onnx::MatMul_8948_45_1_27" + ], + "out_args": [ + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.6/Add_4_output_0.out10_78" + ], + "const_args": [ + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27" + ], + "out_args": [ + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_3_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_4" + ], + "out_args": [ + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_80" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_80" + ], + "const_args": [ + "onnx::MatMul_8949_45_1_26" + ], + "out_args": [ + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_30" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_30" + ], + "const_args": [ + "onnx::MatMul_8950_45_1_28" + ], + "out_args": [ + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32", + "/transformer_blocks.6/Add_4_output_0.out10_78" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_5_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_7_output_0.out10_79" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.6/Add_7_output_0.out10_79" + ], + "const_args": [ + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_28" + ], + "out_args": [ + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_28_bfp.out15_28" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.6/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31", + "/transformer_blocks.6/Add_output_0.out10_76" + ], + "const_args": [ + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_5_gma" + ], + "out_args": [ + "/transformer_blocks.6/Add_3_output_0.out10_77" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_6", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.6/Add_3_output_0.out10_77", + "block_controlnet_hidden_states_6.out_35_1_10" + ], + "const_args": [], + "out_args": [ + "/Add_6_output_0.out_35_1_10" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_6_output_0.out_35_1_10" + ], + "const_args": [ + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_29" + ], + "out_args": [ + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_29_bfp.out15_29" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_29_bfp.out15_29" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_0_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_1" + ], + "out_args": [ + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_29_bfp.out1_81" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_28_bfp.out15_28" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_0_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_1" + ], + "out_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_28_bfp.out1_83" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_28_bfp.out1_83", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_29_bfp.out1_81" + ], + "const_args": [ + "onnx::MatMul_8966_onnx::MatMul_8951" + ], + "out_args": [ + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_28_bfp.out1_83", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_29_bfp.out1_81" + ], + "const_args": [ + "onnx::MatMul_8967_onnx::MatMul_8952" + ], + "out_args": [ + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_28_bfp.out1_83", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_29_bfp.out1_81" + ], + "const_args": [ + "onnx::MatMul_8968_onnx::MatMul_8953" + ], + "out_args": [ + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/MatMulmha_18_0_7", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21", + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22", + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7" + ], + "const_args": [ + "onnx::MatMul_8976_45_5_15" + ], + "out_args": [ + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15", + "/transformer_blocks.6/Add_7_output_0.out10_79" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_2_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_4_output_0.out10_82" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7" + ], + "const_args": [ + "onnx::MatMul_8977_45_5_14" + ], + "out_args": [ + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14", + "/Add_6_output_0.out_35_1_10" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_2_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_output_0.out10_80" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.7/Add_output_0.out10_80" + ], + "const_args": [ + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30" + ], + "out_args": [ + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_3_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_4" + ], + "out_args": [ + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_82" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_82" + ], + "const_args": [ + "onnx::MatMul_8978_45_1_29" + ], + "out_args": [ + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_33" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_33" + ], + "const_args": [ + "onnx::MatMul_8979_45_1_31" + ], + "out_args": [ + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.7/Add_4_output_0.out10_82" + ], + "const_args": [ + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31" + ], + "out_args": [ + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_3_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_4" + ], + "out_args": [ + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_84" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_84" + ], + "const_args": [ + "onnx::MatMul_8980_45_1_30" + ], + "out_args": [ + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_34" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_34" + ], + "const_args": [ + "onnx::MatMul_8981_45_1_32" + ], + "out_args": [ + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36", + "/transformer_blocks.7/Add_4_output_0.out10_82" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_5_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_7_output_0.out10_83" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.7/Add_7_output_0.out10_83" + ], + "const_args": [ + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_32" + ], + "out_args": [ + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_32_bfp.out15_32" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.7/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35", + "/transformer_blocks.7/Add_output_0.out10_80" + ], + "const_args": [ + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_5_gma" + ], + "out_args": [ + "/transformer_blocks.7/Add_3_output_0.out10_81" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_7", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.7/Add_3_output_0.out10_81", + "block_controlnet_hidden_states_7.out_35_1_11" + ], + "const_args": [], + "out_args": [ + "/Add_7_output_0.out_35_1_11" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_7_output_0.out_35_1_11" + ], + "const_args": [ + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_33" + ], + "out_args": [ + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_33_bfp.out15_33" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_33_bfp.out15_33" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_0_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_1" + ], + "out_args": [ + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_33_bfp.out1_85" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_32_bfp.out15_32" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_0_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_1" + ], + "out_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_32_bfp.out1_87" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_32_bfp.out1_87", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_33_bfp.out1_85" + ], + "const_args": [ + "onnx::MatMul_8997_onnx::MatMul_8982" + ], + "out_args": [ + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_32_bfp.out1_87", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_33_bfp.out1_85" + ], + "const_args": [ + "onnx::MatMul_8998_onnx::MatMul_8983" + ], + "out_args": [ + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_32_bfp.out1_87", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_33_bfp.out1_85" + ], + "const_args": [ + "onnx::MatMul_8999_onnx::MatMul_8984" + ], + "out_args": [ + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/MatMulmha_18_0_8", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24", + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25", + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8" + ], + "const_args": [ + "onnx::MatMul_9007_45_5_17" + ], + "out_args": [ + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17", + "/transformer_blocks.7/Add_7_output_0.out10_83" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_2_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_4_output_0.out10_86" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8" + ], + "const_args": [ + "onnx::MatMul_9008_45_5_16" + ], + "out_args": [ + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16", + "/Add_7_output_0.out_35_1_11" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_2_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_output_0.out10_84" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.8/Add_output_0.out10_84" + ], + "const_args": [ + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34" + ], + "out_args": [ + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_3_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_4" + ], + "out_args": [ + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_86" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_86" + ], + "const_args": [ + "onnx::MatMul_9009_45_1_33" + ], + "out_args": [ + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_37" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_37" + ], + "const_args": [ + "onnx::MatMul_9010_45_1_35" + ], + "out_args": [ + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.8/Add_4_output_0.out10_86" + ], + "const_args": [ + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35" + ], + "out_args": [ + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_3_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_4" + ], + "out_args": [ + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_88" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_88" + ], + "const_args": [ + "onnx::MatMul_9011_45_1_34" + ], + "out_args": [ + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_38" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_38" + ], + "const_args": [ + "onnx::MatMul_9012_45_1_36" + ], + "out_args": [ + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40", + "/transformer_blocks.8/Add_4_output_0.out10_86" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_5_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_7_output_0.out10_87" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.8/Add_7_output_0.out10_87" + ], + "const_args": [ + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_36" + ], + "out_args": [ + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_36_bfp.out15_36" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.8/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39", + "/transformer_blocks.8/Add_output_0.out10_84" + ], + "const_args": [ + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_5_gma" + ], + "out_args": [ + "/transformer_blocks.8/Add_3_output_0.out10_85" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_8", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.8/Add_3_output_0.out10_85", + "block_controlnet_hidden_states_8.out_35_1_12" + ], + "const_args": [], + "out_args": [ + "/Add_8_output_0.out_35_1_12" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_8_output_0.out_35_1_12" + ], + "const_args": [ + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_37" + ], + "out_args": [ + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_37_bfp.out15_37" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_37_bfp.out15_37" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_0_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_1" + ], + "out_args": [ + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_37_bfp.out1_89" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_36_bfp.out15_36" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_0_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_1" + ], + "out_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_36_bfp.out1_91" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_36_bfp.out1_91", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_37_bfp.out1_89" + ], + "const_args": [ + "onnx::MatMul_9028_onnx::MatMul_9013" + ], + "out_args": [ + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_36_bfp.out1_91", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_37_bfp.out1_89" + ], + "const_args": [ + "onnx::MatMul_9029_onnx::MatMul_9014" + ], + "out_args": [ + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_36_bfp.out1_91", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_37_bfp.out1_89" + ], + "const_args": [ + "onnx::MatMul_9030_onnx::MatMul_9015" + ], + "out_args": [ + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/MatMulmha_18_0_9", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27", + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28", + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9" + ], + "const_args": [ + "onnx::MatMul_9038_45_5_19" + ], + "out_args": [ + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19", + "/transformer_blocks.8/Add_7_output_0.out10_87" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_2_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_4_output_0.out10_90" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9" + ], + "const_args": [ + "onnx::MatMul_9039_45_5_18" + ], + "out_args": [ + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18", + "/Add_8_output_0.out_35_1_12" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_2_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_output_0.out10_88" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.9/Add_output_0.out10_88" + ], + "const_args": [ + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38" + ], + "out_args": [ + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_3_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_4" + ], + "out_args": [ + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_90" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_90" + ], + "const_args": [ + "onnx::MatMul_9040_45_1_37" + ], + "out_args": [ + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_41" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_41" + ], + "const_args": [ + "onnx::MatMul_9041_45_1_39" + ], + "out_args": [ + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43", + "/transformer_blocks.9/Add_output_0.out10_88" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_5_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_3_output_0.out10_89" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.9/Add_4_output_0.out10_90" + ], + "const_args": [ + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39" + ], + "out_args": [ + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_3_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_4" + ], + "out_args": [ + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_92" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_92" + ], + "const_args": [ + "onnx::MatMul_9042_45_1_38" + ], + "out_args": [ + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_42" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_42" + ], + "const_args": [ + "onnx::MatMul_9043_45_1_40" + ], + "out_args": [ + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.9/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44", + "/transformer_blocks.9/Add_4_output_0.out10_90" + ], + "const_args": [ + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_5_gma" + ], + "out_args": [ + "/transformer_blocks.9/Add_7_output_0.out10_91" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.9/Add_7_output_0.out10_91" + ], + "const_args": [ + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41" + ], + "out_args": [ + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_9", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.9/Add_3_output_0.out10_89", + "block_controlnet_hidden_states_9.out_35_1_13" + ], + "const_args": [], + "out_args": [ + "/Add_9_output_0.out_35_1_13" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_9_output_0.out_35_1_13" + ], + "const_args": [ + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40" + ], + "out_args": [ + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_0_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_1" + ], + "out_args": [ + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_0_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_1" + ], + "out_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "const_args": [ + "onnx::MatMul_9059_onnx::MatMul_9044" + ], + "out_args": [ + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "const_args": [ + "onnx::MatMul_9060_onnx::MatMul_9045" + ], + "out_args": [ + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6" + ], + "const_args": [ + "onnx::MatMul_9061_onnx::MatMul_9046" + ], + "out_args": [ + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/MatMulmha_18_0_10", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30", + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31", + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10" + ], + "const_args": [ + "onnx::MatMul_9070_45_5_20" + ], + "out_args": [ + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20", + "/Add_9_output_0.out_35_1_13" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_2_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_output_0.out10_6" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.10/Add_output_0.out10_6" + ], + "const_args": [ + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42" + ], + "out_args": [ + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_3_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_4" + ], + "out_args": [ + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7" + ], + "const_args": [ + "onnx::MatMul_9071_45_1_41" + ], + "out_args": [ + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_45" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_45" + ], + "const_args": [ + "onnx::MatMul_9072_45_1_43" + ], + "out_args": [ + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47", + "/transformer_blocks.10/Add_output_0.out10_6" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_5_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_3_output_0.out10_7" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10" + ], + "const_args": [ + "onnx::MatMul_9069_45_5_21" + ], + "out_args": [ + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21", + "/transformer_blocks.9/Add_7_output_0.out10_91" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_2_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_4_output_0.out10_8" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.10/Add_4_output_0.out10_8" + ], + "const_args": [ + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43" + ], + "out_args": [ + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_3_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_4" + ], + "out_args": [ + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9" + ], + "const_args": [ + "onnx::MatMul_9073_45_1_42" + ], + "out_args": [ + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_46" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_46" + ], + "const_args": [ + "onnx::MatMul_9074_45_1_44" + ], + "out_args": [ + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.10/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48", + "/transformer_blocks.10/Add_4_output_0.out10_8" + ], + "const_args": [ + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_5_gma" + ], + "out_args": [ + "/transformer_blocks.10/Add_7_output_0.out10_9" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.10/Add_7_output_0.out10_9" + ], + "const_args": [ + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_44" + ], + "out_args": [ + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_44_bfp.out15_44" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_10", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.10/Add_3_output_0.out10_7", + "block_controlnet_hidden_states_10.out_35_1_14" + ], + "const_args": [], + "out_args": [ + "/Add_10_output_0.out_35_1_14" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_10_output_0.out_35_1_14" + ], + "const_args": [ + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_45" + ], + "out_args": [ + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_45_bfp.out15_45" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_45_bfp.out15_45" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_0_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_1" + ], + "out_args": [ + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_45_bfp.out1_10" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_44_bfp.out15_44" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_0_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_1" + ], + "out_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_44_bfp.out1_12" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_44_bfp.out1_12", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_45_bfp.out1_10" + ], + "const_args": [ + "onnx::MatMul_9090_onnx::MatMul_9075" + ], + "out_args": [ + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_44_bfp.out1_12", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_45_bfp.out1_10" + ], + "const_args": [ + "onnx::MatMul_9091_onnx::MatMul_9076" + ], + "out_args": [ + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_44_bfp.out1_12", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_45_bfp.out1_10" + ], + "const_args": [ + "onnx::MatMul_9092_onnx::MatMul_9077" + ], + "out_args": [ + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/MatMulmha_18_0_11", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33", + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34", + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11" + ], + "const_args": [ + "onnx::MatMul_9101_45_5_22" + ], + "out_args": [ + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22", + "/Add_10_output_0.out_35_1_14" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_2_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_output_0.out10_10" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.11/Add_output_0.out10_10" + ], + "const_args": [ + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46" + ], + "out_args": [ + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_3_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_4" + ], + "out_args": [ + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11" + ], + "const_args": [ + "onnx::MatMul_9102_45_1_45" + ], + "out_args": [ + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_49" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_49" + ], + "const_args": [ + "onnx::MatMul_9103_45_1_47" + ], + "out_args": [ + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51", + "/transformer_blocks.11/Add_output_0.out10_10" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_5_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_3_output_0.out10_11" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11" + ], + "const_args": [ + "onnx::MatMul_9100_45_5_23" + ], + "out_args": [ + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23", + "/transformer_blocks.10/Add_7_output_0.out10_9" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_2_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_4_output_0.out10_12" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.11/Add_4_output_0.out10_12" + ], + "const_args": [ + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47", + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47" + ], + "out_args": [ + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_3_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_4" + ], + "out_args": [ + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13" + ], + "const_args": [ + "onnx::MatMul_9104_45_1_46" + ], + "out_args": [ + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_50" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_50" + ], + "const_args": [ + "onnx::MatMul_9105_45_1_48" + ], + "out_args": [ + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.11/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52", + "/transformer_blocks.11/Add_4_output_0.out10_12" + ], + "const_args": [ + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_5_gma" + ], + "out_args": [ + "/transformer_blocks.11/Add_7_output_0.out10_13" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.11/Add_7_output_0.out10_13" + ], + "const_args": [ + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_48", + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_48" + ], + "out_args": [ + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_48_bfp.out15_48" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_11", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.11/Add_3_output_0.out10_11", + "block_controlnet_hidden_states_11.out_35_1_15" + ], + "const_args": [], + "out_args": [ + "/Add_11_output_0.out_35_1_15" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_11_output_0.out_35_1_15" + ], + "const_args": [ + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_49", + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_49" + ], + "out_args": [ + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_49_bfp.out15_49" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_49_bfp.out15_49" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_0_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_1" + ], + "out_args": [ + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_49_bfp.out1_14" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_48_bfp.out15_48" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_0_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_1" + ], + "out_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_48_bfp.out1_16" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_48_bfp.out1_16", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_49_bfp.out1_14" + ], + "const_args": [ + "onnx::MatMul_9121_onnx::MatMul_9106" + ], + "out_args": [ + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_48_bfp.out1_16", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_49_bfp.out1_14" + ], + "const_args": [ + "onnx::MatMul_9122_onnx::MatMul_9107" + ], + "out_args": [ + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_48_bfp.out1_16", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_49_bfp.out1_14" + ], + "const_args": [ + "onnx::MatMul_9123_onnx::MatMul_9108" + ], + "out_args": [ + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/MatMulmha_18_0_12", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36", + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37", + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12" + ], + "const_args": [ + "onnx::MatMul_9132_45_5_24" + ], + "out_args": [ + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24", + "/Add_11_output_0.out_35_1_15" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_2_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_output_0.out10_14" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.12/Add_output_0.out10_14" + ], + "const_args": [ + "/transformer_blocks.12/norm2/Constant_output_0_14_50", + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50" + ], + "out_args": [ + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_3_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_4" + ], + "out_args": [ + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15" + ], + "const_args": [ + "onnx::MatMul_9133_45_1_49" + ], + "out_args": [ + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_53" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_53" + ], + "const_args": [ + "onnx::MatMul_9134_45_1_51" + ], + "out_args": [ + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55", + "/transformer_blocks.12/Add_output_0.out10_14" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_5_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_3_output_0.out10_15" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12" + ], + "const_args": [ + "onnx::MatMul_9131_45_5_25" + ], + "out_args": [ + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25", + "/transformer_blocks.11/Add_7_output_0.out10_13" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_2_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_4_output_0.out10_16" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.12/Add_4_output_0.out10_16" + ], + "const_args": [ + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51", + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51" + ], + "out_args": [ + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_3_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_4" + ], + "out_args": [ + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17" + ], + "const_args": [ + "onnx::MatMul_9135_45_1_50" + ], + "out_args": [ + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_54" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_54" + ], + "const_args": [ + "onnx::MatMul_9136_45_1_52" + ], + "out_args": [ + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.12/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56", + "/transformer_blocks.12/Add_4_output_0.out10_16" + ], + "const_args": [ + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_5_gma" + ], + "out_args": [ + "/transformer_blocks.12/Add_7_output_0.out10_17" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.12/Add_7_output_0.out10_17" + ], + "const_args": [ + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_52", + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_52" + ], + "out_args": [ + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_52_bfp.out15_52" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_12", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.12/Add_3_output_0.out10_15", + "block_controlnet_hidden_states_12.out_35_1_16" + ], + "const_args": [], + "out_args": [ + "/Add_12_output_0.out_35_1_16" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_12_output_0.out_35_1_16" + ], + "const_args": [ + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_53", + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_53" + ], + "out_args": [ + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_53_bfp.out15_53" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_53_bfp.out15_53" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_0_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_1" + ], + "out_args": [ + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_53_bfp.out1_18" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_52_bfp.out15_52" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_0_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_1" + ], + "out_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_52_bfp.out1_20" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_52_bfp.out1_20", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_53_bfp.out1_18" + ], + "const_args": [ + "onnx::MatMul_9152_onnx::MatMul_9137" + ], + "out_args": [ + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_52_bfp.out1_20", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_53_bfp.out1_18" + ], + "const_args": [ + "onnx::MatMul_9153_onnx::MatMul_9138" + ], + "out_args": [ + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_52_bfp.out1_20", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_53_bfp.out1_18" + ], + "const_args": [ + "onnx::MatMul_9154_onnx::MatMul_9139" + ], + "out_args": [ + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/MatMulmha_18_0_13", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39", + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40", + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13" + ], + "const_args": [ + "onnx::MatMul_9163_45_5_26" + ], + "out_args": [ + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26", + "/Add_12_output_0.out_35_1_16" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_2_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_output_0.out10_18" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.13/Add_output_0.out10_18" + ], + "const_args": [ + "/transformer_blocks.13/norm2/Constant_output_0_14_54", + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54" + ], + "out_args": [ + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_3_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_4" + ], + "out_args": [ + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19" + ], + "const_args": [ + "onnx::MatMul_9164_45_1_53" + ], + "out_args": [ + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_57" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_57" + ], + "const_args": [ + "onnx::MatMul_9165_45_1_55" + ], + "out_args": [ + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59", + "/transformer_blocks.13/Add_output_0.out10_18" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_5_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_3_output_0.out10_19" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13" + ], + "const_args": [ + "onnx::MatMul_9162_45_5_27" + ], + "out_args": [ + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27", + "/transformer_blocks.12/Add_7_output_0.out10_17" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_2_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_4_output_0.out10_20" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.13/Add_4_output_0.out10_20" + ], + "const_args": [ + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55", + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55" + ], + "out_args": [ + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_3_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_4" + ], + "out_args": [ + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21" + ], + "const_args": [ + "onnx::MatMul_9166_45_1_54" + ], + "out_args": [ + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_58" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_58" + ], + "const_args": [ + "onnx::MatMul_9167_45_1_56" + ], + "out_args": [ + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.13/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60", + "/transformer_blocks.13/Add_4_output_0.out10_20" + ], + "const_args": [ + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_5_gma" + ], + "out_args": [ + "/transformer_blocks.13/Add_7_output_0.out10_21" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.13/Add_7_output_0.out10_21" + ], + "const_args": [ + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_56", + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_56" + ], + "out_args": [ + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_56_bfp.out15_56" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_13", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.13/Add_3_output_0.out10_19", + "block_controlnet_hidden_states_13.out_35_1_17" + ], + "const_args": [], + "out_args": [ + "/Add_13_output_0.out_35_1_17" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_13_output_0.out_35_1_17" + ], + "const_args": [ + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_57", + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_57" + ], + "out_args": [ + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_57_bfp.out15_57" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_57_bfp.out15_57" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_0_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_1" + ], + "out_args": [ + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_57_bfp.out1_22" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_56_bfp.out15_56" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_0_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_1" + ], + "out_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_56_bfp.out1_24" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_56_bfp.out1_24", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_57_bfp.out1_22" + ], + "const_args": [ + "onnx::MatMul_9183_onnx::MatMul_9168" + ], + "out_args": [ + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_56_bfp.out1_24", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_57_bfp.out1_22" + ], + "const_args": [ + "onnx::MatMul_9184_onnx::MatMul_9169" + ], + "out_args": [ + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_56_bfp.out1_24", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_57_bfp.out1_22" + ], + "const_args": [ + "onnx::MatMul_9185_onnx::MatMul_9170" + ], + "out_args": [ + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/MatMulmha_18_0_14", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42", + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43", + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14" + ], + "const_args": [ + "onnx::MatMul_9194_45_5_28" + ], + "out_args": [ + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28", + "/Add_13_output_0.out_35_1_17" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_2_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_output_0.out10_22" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.14/Add_output_0.out10_22" + ], + "const_args": [ + "/transformer_blocks.14/norm2/Constant_output_0_14_58", + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58" + ], + "out_args": [ + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_3_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_4" + ], + "out_args": [ + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23" + ], + "const_args": [ + "onnx::MatMul_9195_45_1_57" + ], + "out_args": [ + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_61" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_61" + ], + "const_args": [ + "onnx::MatMul_9196_45_1_59" + ], + "out_args": [ + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63", + "/transformer_blocks.14/Add_output_0.out10_22" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_5_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_3_output_0.out10_23" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14" + ], + "const_args": [ + "onnx::MatMul_9193_45_5_29" + ], + "out_args": [ + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29", + "/transformer_blocks.13/Add_7_output_0.out10_21" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_2_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_4_output_0.out10_24" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.14/Add_4_output_0.out10_24" + ], + "const_args": [ + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59", + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59" + ], + "out_args": [ + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_3_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_4" + ], + "out_args": [ + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25" + ], + "const_args": [ + "onnx::MatMul_9197_45_1_58" + ], + "out_args": [ + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_62" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_62" + ], + "const_args": [ + "onnx::MatMul_9198_45_1_60" + ], + "out_args": [ + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.14/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64", + "/transformer_blocks.14/Add_4_output_0.out10_24" + ], + "const_args": [ + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_5_gma" + ], + "out_args": [ + "/transformer_blocks.14/Add_7_output_0.out10_25" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.14/Add_7_output_0.out10_25" + ], + "const_args": [ + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_60", + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_60" + ], + "out_args": [ + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_60_bfp.out15_60" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_14", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.14/Add_3_output_0.out10_23", + "block_controlnet_hidden_states_14.out_35_1_18" + ], + "const_args": [], + "out_args": [ + "/Add_14_output_0.out_35_1_18" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_14_output_0.out_35_1_18" + ], + "const_args": [ + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_61", + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_61" + ], + "out_args": [ + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_61_bfp.out15_61" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_61_bfp.out15_61" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_0_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_1" + ], + "out_args": [ + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_61_bfp.out1_26" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_60_bfp.out15_60" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_0_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_1" + ], + "out_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_60_bfp.out1_28" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_60_bfp.out1_28", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_61_bfp.out1_26" + ], + "const_args": [ + "onnx::MatMul_9214_onnx::MatMul_9199" + ], + "out_args": [ + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_60_bfp.out1_28", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_61_bfp.out1_26" + ], + "const_args": [ + "onnx::MatMul_9215_onnx::MatMul_9200" + ], + "out_args": [ + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_60_bfp.out1_28", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_61_bfp.out1_26" + ], + "const_args": [ + "onnx::MatMul_9216_onnx::MatMul_9201" + ], + "out_args": [ + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/MatMulmha_18_0_15", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45", + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46", + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15" + ], + "const_args": [ + "onnx::MatMul_9225_45_5_30" + ], + "out_args": [ + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30", + "/Add_14_output_0.out_35_1_18" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_2_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_output_0.out10_26" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.15/Add_output_0.out10_26" + ], + "const_args": [ + "/transformer_blocks.15/norm2/Constant_output_0_14_62", + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62" + ], + "out_args": [ + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_3_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_4" + ], + "out_args": [ + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27" + ], + "const_args": [ + "onnx::MatMul_9226_45_1_61" + ], + "out_args": [ + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_65" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_65" + ], + "const_args": [ + "onnx::MatMul_9227_45_1_63" + ], + "out_args": [ + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67", + "/transformer_blocks.15/Add_output_0.out10_26" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_5_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_3_output_0.out10_27" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15" + ], + "const_args": [ + "onnx::MatMul_9224_45_5_31" + ], + "out_args": [ + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31", + "/transformer_blocks.14/Add_7_output_0.out10_25" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_2_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_4_output_0.out10_28" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.15/Add_4_output_0.out10_28" + ], + "const_args": [ + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63", + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63" + ], + "out_args": [ + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_3_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_4" + ], + "out_args": [ + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29" + ], + "const_args": [ + "onnx::MatMul_9228_45_1_62" + ], + "out_args": [ + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_66" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_66" + ], + "const_args": [ + "onnx::MatMul_9229_45_1_64" + ], + "out_args": [ + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.15/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68", + "/transformer_blocks.15/Add_4_output_0.out10_28" + ], + "const_args": [ + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_5_gma" + ], + "out_args": [ + "/transformer_blocks.15/Add_7_output_0.out10_29" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.15/Add_7_output_0.out10_29" + ], + "const_args": [ + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_64", + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_64" + ], + "out_args": [ + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_64_bfp.out15_64" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_15", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.15/Add_3_output_0.out10_27", + "block_controlnet_hidden_states_15.out_35_1_19" + ], + "const_args": [], + "out_args": [ + "/Add_15_output_0.out_35_1_19" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_15_output_0.out_35_1_19" + ], + "const_args": [ + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_65", + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_65" + ], + "out_args": [ + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_65_bfp.out15_65" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_65_bfp.out15_65" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_0_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_1" + ], + "out_args": [ + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_65_bfp.out1_30" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_64_bfp.out15_64" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_0_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_1" + ], + "out_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_64_bfp.out1_32" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_64_bfp.out1_32", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_65_bfp.out1_30" + ], + "const_args": [ + "onnx::MatMul_9245_onnx::MatMul_9230" + ], + "out_args": [ + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_64_bfp.out1_32", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_65_bfp.out1_30" + ], + "const_args": [ + "onnx::MatMul_9246_onnx::MatMul_9231" + ], + "out_args": [ + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_64_bfp.out1_32", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_65_bfp.out1_30" + ], + "const_args": [ + "onnx::MatMul_9247_onnx::MatMul_9232" + ], + "out_args": [ + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/MatMulmha_18_0_16", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48", + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49", + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16" + ], + "const_args": [ + "onnx::MatMul_9256_45_5_32" + ], + "out_args": [ + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32", + "/Add_15_output_0.out_35_1_19" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_2_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_output_0.out10_30" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.16/Add_output_0.out10_30" + ], + "const_args": [ + "/transformer_blocks.16/norm2/Constant_output_0_14_66", + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66" + ], + "out_args": [ + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_3_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_4" + ], + "out_args": [ + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31" + ], + "const_args": [ + "onnx::MatMul_9257_45_1_65" + ], + "out_args": [ + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_69" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_69" + ], + "const_args": [ + "onnx::MatMul_9258_45_1_67" + ], + "out_args": [ + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71", + "/transformer_blocks.16/Add_output_0.out10_30" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_5_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_3_output_0.out10_31" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16" + ], + "const_args": [ + "onnx::MatMul_9255_45_5_33" + ], + "out_args": [ + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33", + "/transformer_blocks.15/Add_7_output_0.out10_29" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_2_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_4_output_0.out10_32" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.16/Add_4_output_0.out10_32" + ], + "const_args": [ + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67", + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67" + ], + "out_args": [ + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_3_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_4" + ], + "out_args": [ + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33" + ], + "const_args": [ + "onnx::MatMul_9259_45_1_66" + ], + "out_args": [ + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_70" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_70" + ], + "const_args": [ + "onnx::MatMul_9260_45_1_68" + ], + "out_args": [ + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.16/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72", + "/transformer_blocks.16/Add_4_output_0.out10_32" + ], + "const_args": [ + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_5_gma" + ], + "out_args": [ + "/transformer_blocks.16/Add_7_output_0.out10_33" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.16/Add_7_output_0.out10_33" + ], + "const_args": [ + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_68", + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_68" + ], + "out_args": [ + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_68_bfp.out15_68" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_16", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.16/Add_3_output_0.out10_31", + "block_controlnet_hidden_states_16.out_35_1_20" + ], + "const_args": [], + "out_args": [ + "/Add_16_output_0.out_35_1_20" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_16_output_0.out_35_1_20" + ], + "const_args": [ + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_69", + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_69" + ], + "out_args": [ + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_69_bfp.out15_69" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_69_bfp.out15_69" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_0_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_1" + ], + "out_args": [ + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_69_bfp.out1_34" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_68_bfp.out15_68" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_0_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_1" + ], + "out_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_68_bfp.out1_36" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_68_bfp.out1_36", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_69_bfp.out1_34" + ], + "const_args": [ + "onnx::MatMul_9276_onnx::MatMul_9261" + ], + "out_args": [ + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_68_bfp.out1_36", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_69_bfp.out1_34" + ], + "const_args": [ + "onnx::MatMul_9277_onnx::MatMul_9262" + ], + "out_args": [ + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_68_bfp.out1_36", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_69_bfp.out1_34" + ], + "const_args": [ + "onnx::MatMul_9278_onnx::MatMul_9263" + ], + "out_args": [ + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/MatMulmha_18_0_17", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51", + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52", + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17" + ], + "const_args": [ + "onnx::MatMul_9287_45_5_34" + ], + "out_args": [ + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34", + "/Add_16_output_0.out_35_1_20" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_2_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_output_0.out10_34" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.17/Add_output_0.out10_34" + ], + "const_args": [ + "/transformer_blocks.17/norm2/Constant_output_0_14_70", + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70" + ], + "out_args": [ + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_3_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_4" + ], + "out_args": [ + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35" + ], + "const_args": [ + "onnx::MatMul_9288_45_1_69" + ], + "out_args": [ + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_73" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_73" + ], + "const_args": [ + "onnx::MatMul_9289_45_1_71" + ], + "out_args": [ + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75", + "/transformer_blocks.17/Add_output_0.out10_34" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_5_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_3_output_0.out10_35" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17" + ], + "const_args": [ + "onnx::MatMul_9286_45_5_35" + ], + "out_args": [ + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35", + "/transformer_blocks.16/Add_7_output_0.out10_33" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_2_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_4_output_0.out10_36" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.17/Add_4_output_0.out10_36" + ], + "const_args": [ + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71", + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71" + ], + "out_args": [ + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_3_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_4" + ], + "out_args": [ + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37" + ], + "const_args": [ + "onnx::MatMul_9290_45_1_70" + ], + "out_args": [ + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_74" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_74" + ], + "const_args": [ + "onnx::MatMul_9291_45_1_72" + ], + "out_args": [ + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.17/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76", + "/transformer_blocks.17/Add_4_output_0.out10_36" + ], + "const_args": [ + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_5_gma" + ], + "out_args": [ + "/transformer_blocks.17/Add_7_output_0.out10_37" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.17/Add_7_output_0.out10_37" + ], + "const_args": [ + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_72", + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_72" + ], + "out_args": [ + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_72_bfp.out15_72" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_17", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.17/Add_3_output_0.out10_35", + "block_controlnet_hidden_states_17.out_35_1_21" + ], + "const_args": [], + "out_args": [ + "/Add_17_output_0.out_35_1_21" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_17_output_0.out_35_1_21" + ], + "const_args": [ + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_73", + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_73" + ], + "out_args": [ + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_73_bfp.out15_73" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_73_bfp.out15_73" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_0_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_1" + ], + "out_args": [ + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_73_bfp.out1_38" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_72_bfp.out15_72" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_0_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_1" + ], + "out_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_72_bfp.out1_40" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_72_bfp.out1_40", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_73_bfp.out1_38" + ], + "const_args": [ + "onnx::MatMul_9307_onnx::MatMul_9292" + ], + "out_args": [ + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_72_bfp.out1_40", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_73_bfp.out1_38" + ], + "const_args": [ + "onnx::MatMul_9308_onnx::MatMul_9293" + ], + "out_args": [ + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_72_bfp.out1_40", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_73_bfp.out1_38" + ], + "const_args": [ + "onnx::MatMul_9309_onnx::MatMul_9294" + ], + "out_args": [ + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/MatMulmha_18_0_18", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54", + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55", + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18" + ], + "const_args": [ + "onnx::MatMul_9318_45_5_36" + ], + "out_args": [ + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36", + "/Add_17_output_0.out_35_1_21" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_2_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_output_0.out10_38" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.18/Add_output_0.out10_38" + ], + "const_args": [ + "/transformer_blocks.18/norm2/Constant_output_0_14_74", + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74" + ], + "out_args": [ + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_3_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_4" + ], + "out_args": [ + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39" + ], + "const_args": [ + "onnx::MatMul_9319_45_1_73" + ], + "out_args": [ + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_77" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_77" + ], + "const_args": [ + "onnx::MatMul_9320_45_1_75" + ], + "out_args": [ + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79", + "/transformer_blocks.18/Add_output_0.out10_38" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_5_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_3_output_0.out10_39" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18" + ], + "const_args": [ + "onnx::MatMul_9317_45_5_37" + ], + "out_args": [ + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37", + "/transformer_blocks.17/Add_7_output_0.out10_37" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_2_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_4_output_0.out10_40" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.18/Add_4_output_0.out10_40" + ], + "const_args": [ + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75", + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75" + ], + "out_args": [ + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_3_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_4" + ], + "out_args": [ + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41" + ], + "const_args": [ + "onnx::MatMul_9321_45_1_74" + ], + "out_args": [ + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_78" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_78" + ], + "const_args": [ + "onnx::MatMul_9322_45_1_76" + ], + "out_args": [ + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.18/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80", + "/transformer_blocks.18/Add_4_output_0.out10_40" + ], + "const_args": [ + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_5_gma" + ], + "out_args": [ + "/transformer_blocks.18/Add_7_output_0.out10_41" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.18/Add_7_output_0.out10_41" + ], + "const_args": [ + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_76", + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_76" + ], + "out_args": [ + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_76_bfp.out15_76" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_18", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.18/Add_3_output_0.out10_39", + "block_controlnet_hidden_states_18.out_35_1_22" + ], + "const_args": [], + "out_args": [ + "/Add_18_output_0.out_35_1_22" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_18_output_0.out_35_1_22" + ], + "const_args": [ + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_77", + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_77" + ], + "out_args": [ + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_77_bfp.out15_77" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_77_bfp.out15_77" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_0_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_1" + ], + "out_args": [ + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_77_bfp.out1_42" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_76_bfp.out15_76" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_0_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_1" + ], + "out_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_76_bfp.out1_44" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_76_bfp.out1_44", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_77_bfp.out1_42" + ], + "const_args": [ + "onnx::MatMul_9338_onnx::MatMul_9323" + ], + "out_args": [ + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_76_bfp.out1_44", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_77_bfp.out1_42" + ], + "const_args": [ + "onnx::MatMul_9339_onnx::MatMul_9324" + ], + "out_args": [ + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_76_bfp.out1_44", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_77_bfp.out1_42" + ], + "const_args": [ + "onnx::MatMul_9340_onnx::MatMul_9325" + ], + "out_args": [ + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/MatMulmha_18_0_19", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57", + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58", + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19" + ], + "const_args": [ + "onnx::MatMul_9349_45_5_38" + ], + "out_args": [ + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38", + "/Add_18_output_0.out_35_1_22" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_2_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_output_0.out10_42" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.19/Add_output_0.out10_42" + ], + "const_args": [ + "/transformer_blocks.19/norm2/Constant_output_0_14_78", + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78" + ], + "out_args": [ + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_3_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_4" + ], + "out_args": [ + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43" + ], + "const_args": [ + "onnx::MatMul_9350_45_1_77" + ], + "out_args": [ + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_81" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_81" + ], + "const_args": [ + "onnx::MatMul_9351_45_1_79" + ], + "out_args": [ + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83", + "/transformer_blocks.19/Add_output_0.out10_42" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_5_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_3_output_0.out10_43" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19" + ], + "const_args": [ + "onnx::MatMul_9348_45_5_39" + ], + "out_args": [ + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39", + "/transformer_blocks.18/Add_7_output_0.out10_41" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_2_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_4_output_0.out10_44" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.19/Add_4_output_0.out10_44" + ], + "const_args": [ + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79", + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79" + ], + "out_args": [ + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_3_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_4" + ], + "out_args": [ + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45" + ], + "const_args": [ + "onnx::MatMul_9352_45_1_78" + ], + "out_args": [ + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_82" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_82" + ], + "const_args": [ + "onnx::MatMul_9353_45_1_80" + ], + "out_args": [ + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.19/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84", + "/transformer_blocks.19/Add_4_output_0.out10_44" + ], + "const_args": [ + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_5_gma" + ], + "out_args": [ + "/transformer_blocks.19/Add_7_output_0.out10_45" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.19/Add_7_output_0.out10_45" + ], + "const_args": [ + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_80", + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_80" + ], + "out_args": [ + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_80_bfp.out15_80" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_19", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.19/Add_3_output_0.out10_43", + "block_controlnet_hidden_states_19.out_35_1_23" + ], + "const_args": [], + "out_args": [ + "/Add_19_output_0.out_35_1_23" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_19_output_0.out_35_1_23" + ], + "const_args": [ + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_81", + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_81" + ], + "out_args": [ + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_81_bfp.out15_81" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_81_bfp.out15_81" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_0_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_1" + ], + "out_args": [ + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_81_bfp.out1_50" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_80_bfp.out15_80" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_0_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_1" + ], + "out_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_80_bfp.out1_52" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_80_bfp.out1_52", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_81_bfp.out1_50" + ], + "const_args": [ + "onnx::MatMul_9369_onnx::MatMul_9354" + ], + "out_args": [ + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_80_bfp.out1_52", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_81_bfp.out1_50" + ], + "const_args": [ + "onnx::MatMul_9370_onnx::MatMul_9355" + ], + "out_args": [ + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_80_bfp.out1_52", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_81_bfp.out1_50" + ], + "const_args": [ + "onnx::MatMul_9371_onnx::MatMul_9356" + ], + "out_args": [ + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/MatMulmha_18_0_20", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60", + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61", + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20" + ], + "const_args": [ + "onnx::MatMul_9380_45_5_40" + ], + "out_args": [ + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40", + "/Add_19_output_0.out_35_1_23" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_2_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_output_0.out10_50" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.20/Add_output_0.out10_50" + ], + "const_args": [ + "/transformer_blocks.20/norm2/Constant_output_0_14_82", + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82" + ], + "out_args": [ + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_3_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_4" + ], + "out_args": [ + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51" + ], + "const_args": [ + "onnx::MatMul_9381_45_1_81" + ], + "out_args": [ + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_85" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_85" + ], + "const_args": [ + "onnx::MatMul_9382_45_1_83" + ], + "out_args": [ + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87", + "/transformer_blocks.20/Add_output_0.out10_50" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_5_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_3_output_0.out10_51" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20" + ], + "const_args": [ + "onnx::MatMul_9379_45_5_41" + ], + "out_args": [ + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41", + "/transformer_blocks.19/Add_7_output_0.out10_45" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_2_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_4_output_0.out10_52" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.20/Add_4_output_0.out10_52" + ], + "const_args": [ + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83", + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83" + ], + "out_args": [ + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_3_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_4" + ], + "out_args": [ + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53" + ], + "const_args": [ + "onnx::MatMul_9383_45_1_82" + ], + "out_args": [ + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_86" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_86" + ], + "const_args": [ + "onnx::MatMul_9384_45_1_84" + ], + "out_args": [ + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.20/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88", + "/transformer_blocks.20/Add_4_output_0.out10_52" + ], + "const_args": [ + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_5_gma" + ], + "out_args": [ + "/transformer_blocks.20/Add_7_output_0.out10_53" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.20/Add_7_output_0.out10_53" + ], + "const_args": [ + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_84", + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_84" + ], + "out_args": [ + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_84_bfp.out15_84" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_20", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.20/Add_3_output_0.out10_51", + "block_controlnet_hidden_states_20.out_35_1_24" + ], + "const_args": [], + "out_args": [ + "/Add_20_output_0.out_35_1_24" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_20_output_0.out_35_1_24" + ], + "const_args": [ + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_85", + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_85" + ], + "out_args": [ + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_85_bfp.out15_85" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_85_bfp.out15_85" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_0_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_1" + ], + "out_args": [ + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_85_bfp.out1_54" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_84_bfp.out15_84" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_0_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_1" + ], + "out_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_84_bfp.out1_56" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_84_bfp.out1_56", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_85_bfp.out1_54" + ], + "const_args": [ + "onnx::MatMul_9400_onnx::MatMul_9385" + ], + "out_args": [ + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_84_bfp.out1_56", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_85_bfp.out1_54" + ], + "const_args": [ + "onnx::MatMul_9401_onnx::MatMul_9386" + ], + "out_args": [ + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_84_bfp.out1_56", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_85_bfp.out1_54" + ], + "const_args": [ + "onnx::MatMul_9402_onnx::MatMul_9387" + ], + "out_args": [ + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/MatMulmha_18_0_21", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63", + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64", + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21" + ], + "const_args": [ + "onnx::MatMul_9411_45_5_42" + ], + "out_args": [ + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42", + "/Add_20_output_0.out_35_1_24" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_2_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_output_0.out10_54" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.21/Add_output_0.out10_54" + ], + "const_args": [ + "/transformer_blocks.21/norm2/Constant_output_0_14_86", + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86" + ], + "out_args": [ + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_3_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_4" + ], + "out_args": [ + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55" + ], + "const_args": [ + "onnx::MatMul_9412_45_1_85" + ], + "out_args": [ + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_89" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_89" + ], + "const_args": [ + "onnx::MatMul_9413_45_1_87" + ], + "out_args": [ + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91", + "/transformer_blocks.21/Add_output_0.out10_54" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_5_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_3_output_0.out10_55" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21" + ], + "const_args": [ + "onnx::MatMul_9410_45_5_43" + ], + "out_args": [ + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43", + "/transformer_blocks.20/Add_7_output_0.out10_53" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_2_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_4_output_0.out10_56" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.21/Add_4_output_0.out10_56" + ], + "const_args": [ + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87", + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87" + ], + "out_args": [ + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_3_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_4" + ], + "out_args": [ + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57" + ], + "const_args": [ + "onnx::MatMul_9414_45_1_86" + ], + "out_args": [ + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_90" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_90" + ], + "const_args": [ + "onnx::MatMul_9415_45_1_88" + ], + "out_args": [ + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.21/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92", + "/transformer_blocks.21/Add_4_output_0.out10_56" + ], + "const_args": [ + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_5_gma" + ], + "out_args": [ + "/transformer_blocks.21/Add_7_output_0.out10_57" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.21/Add_7_output_0.out10_57" + ], + "const_args": [ + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_88", + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_88" + ], + "out_args": [ + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_88_bfp.out15_88" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_21", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.21/Add_3_output_0.out10_55", + "block_controlnet_hidden_states_21.out_35_1_25" + ], + "const_args": [], + "out_args": [ + "/Add_21_output_0.out_35_1_25" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_21_output_0.out_35_1_25" + ], + "const_args": [ + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_89", + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_89" + ], + "out_args": [ + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_89_bfp.out15_89" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_89_bfp.out15_89" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_0_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_1" + ], + "out_args": [ + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_89_bfp.out1_58" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_88_bfp.out15_88" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_0_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_1" + ], + "out_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_88_bfp.out1_60" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_88_bfp.out1_60", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_89_bfp.out1_58" + ], + "const_args": [ + "onnx::MatMul_9431_onnx::MatMul_9416" + ], + "out_args": [ + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_88_bfp.out1_60", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_89_bfp.out1_58" + ], + "const_args": [ + "onnx::MatMul_9432_onnx::MatMul_9417" + ], + "out_args": [ + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_88_bfp.out1_60", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_89_bfp.out1_58" + ], + "const_args": [ + "onnx::MatMul_9433_onnx::MatMul_9418" + ], + "out_args": [ + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/MatMulmha_18_0_22", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66", + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67", + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22" + ], + "const_args": [ + "onnx::MatMul_9442_45_5_44" + ], + "out_args": [ + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44", + "/Add_21_output_0.out_35_1_25" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_2_gma" + ], + "out_args": [ + "/transformer_blocks.22/Add_output_0.out10_58" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.22/Add_output_0.out10_58" + ], + "const_args": [ + "/transformer_blocks.22/norm2/Constant_output_0_14_90", + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90" + ], + "out_args": [ + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_3_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_4" + ], + "out_args": [ + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59" + ], + "const_args": [ + "onnx::MatMul_9443_45_1_89" + ], + "out_args": [ + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_93" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_93" + ], + "const_args": [ + "onnx::MatMul_9444_45_1_91" + ], + "out_args": [ + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95", + "/transformer_blocks.22/Add_output_0.out10_58" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_5_gma" + ], + "out_args": [ + "/transformer_blocks.22/Add_3_output_0.out10_59" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/attn/to_add_out/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22" + ], + "const_args": [ + "onnx::MatMul_9441_45_5_45" + ], + "out_args": [ + "/transformer_blocks.22/attn/to_add_out/Add_output_0.out6_1_45_bfp.out7_45" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "end": { + "type": "int", + "value": [ + "2147483647" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add_4", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/attn/to_add_out/Add_output_0.out6_1_45_bfp.out7_45", + "/transformer_blocks.21/Add_7_output_0.out10_57" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_2_gma" + ], + "out_args": [ + "/transformer_blocks.22/Add_4_output_0.out10_60" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/norm2_context/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.22/Add_4_output_0.out10_60" + ], + "const_args": [ + "/transformer_blocks.22/norm2_context/Constant_output_0_14_91", + "/transformer_blocks.22/norm2_context/Constant_1_output_0_14_91" + ], + "out_args": [ + "/transformer_blocks.22/norm2_context/LayerNormalization_output_0.out14_91_bfp.out15_91" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add_6", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/norm2_context/LayerNormalization_output_0.out14_91_bfp.out15_91" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_3_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_4" + ], + "out_args": [ + "/transformer_blocks.22/Add_6_output_0.out0_0_91_bfp.out1_61" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/ff_context/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.22/Add_6_output_0.out0_0_91_bfp.out1_61" + ], + "const_args": [ + "onnx::MatMul_9445_45_1_90" + ], + "out_args": [ + "/transformer_blocks.22/ff_context/net.0/Mul_5_output_0.out17_2_45_bfp.out25_94" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/ff_context/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.22/ff_context/net.0/Mul_5_output_0.out17_2_45_bfp.out25_94" + ], + "const_args": [ + "onnx::MatMul_9446_45_1_92" + ], + "out_args": [ + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.22/Add_7", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96", + "/transformer_blocks.22/Add_4_output_0.out10_60" + ], + "const_args": [ + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_5_gma" + ], + "out_args": [ + "/transformer_blocks.22/Add_7_output_0.out10_61" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/norm1_context/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.22/Add_7_output_0.out10_61" + ], + "const_args": [ + "/transformer_blocks.23/norm1_context/norm/Constant_output_0_14_92", + "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0_14_92" + ], + "out_args": [ + "/transformer_blocks.23/norm1_context/norm/LayerNormalization_output_0.out14_92_bfp.out15_92" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/Add_22", + "type": "SDAdd", + "in_args": [ + "/transformer_blocks.22/Add_3_output_0.out10_59", + "block_controlnet_hidden_states_22.out_35_1_26" + ], + "const_args": [], + "out_args": [ + "/Add_22_output_0.out_35_1_26" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/Add_22_output_0.out_35_1_26" + ], + "const_args": [ + "/transformer_blocks.23/norm1/norm/Constant_output_0_14_93", + "/transformer_blocks.23/norm1/norm/Constant_1_output_0_14_93" + ], + "out_args": [ + "/transformer_blocks.23/norm1/norm/LayerNormalization_output_0.out14_93_bfp.out15_93" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.23/norm1/norm/LayerNormalization_output_0.out14_93_bfp.out15_93" + ], + "const_args": [ + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_0_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_1" + ], + "out_args": [ + "/transformer_blocks.23/norm1/Add_2_output_0.out0_0_93_bfp.out1_62" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/norm1_context/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.23/norm1_context/norm/LayerNormalization_output_0.out14_92_bfp.out15_92" + ], + "const_args": [ + "existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_0_existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_1" + ], + "out_args": [ + "/transformer_blocks.23/norm1_context/Add_2_output_0.out0_0_92_bfp.out1_64" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/attn/Concat_6_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.23/norm1_context/Add_2_output_0.out0_0_92_bfp.out1_64", + "/transformer_blocks.23/norm1/Add_2_output_0.out0_0_93_bfp.out1_62" + ], + "const_args": [ + "onnx::MatMul_9462_onnx::MatMul_9447" + ], + "out_args": [ + "/transformer_blocks.23/attn/Concat_6_3d.out22_23_bfp.out23_69" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/attn/Concat_7_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.23/norm1_context/Add_2_output_0.out0_0_92_bfp.out1_64", + "/transformer_blocks.23/norm1/Add_2_output_0.out0_0_93_bfp.out1_62" + ], + "const_args": [ + "onnx::MatMul_9463_onnx::MatMul_9448" + ], + "out_args": [ + "/transformer_blocks.23/attn/Concat_7_3d.out22_23_bfp.out23_70" + ], + "attrs": { + "trans_head": { + "type": "int", + "value": [ + "1" + ] + }, + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/attn/Concat_8_3d", + "type": "SDGemmConcat_bfp", + "in_args": [ + "/transformer_blocks.23/norm1_context/Add_2_output_0.out0_0_92_bfp.out1_64", + "/transformer_blocks.23/norm1/Add_2_output_0.out0_0_93_bfp.out1_62" + ], + "const_args": [ + "onnx::MatMul_9464_onnx::MatMul_9449" + ], + "out_args": [ + "/transformer_blocks.23/attn/Concat_8_3d.out22_23_bfp.out23_71" + ], + "attrs": { + "head_num": { + "type": "int", + "value": [ + "24" + ] + }, + "concat_axis": { + "type": "int", + "value": [ + "-2" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape_0": { + "type": "str", + "value": [ + "batch_size", + "max_length", + "1536" + ] + }, + "output_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "trans_head": { + "type": "int", + "value": [ + "3" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/attn/MatMulmha_18_0_23", + "type": "SDMHA_bfp", + "in_args": [ + "/transformer_blocks.23/attn/Concat_6_3d.out22_23_bfp.out23_69", + "/transformer_blocks.23/attn/Concat_7_3d.out22_23_bfp.out23_70", + "/transformer_blocks.23/attn/Concat_8_3d.out22_23_bfp.out23_71" + ], + "const_args": [], + "out_args": [ + "/transformer_blocks.23/attn/Reshape_6_output_0.out22_23_bfp.out27_0_23" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "24" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1+max_length", + "64", + "state_dim1+max_length" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v2" + ] + }, + "is_flash_mha": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/attn/to_out.0/MatMul", + "type": "SDSliceGemm_bfp", + "in_args": [ + "/transformer_blocks.23/attn/Reshape_6_output_0.out22_23_bfp.out27_0_23" + ], + "const_args": [ + "onnx::MatMul_9469_45_5_46" + ], + "out_args": [ + "/transformer_blocks.23/attn/to_out.0/Add_output_0.out6_1_46_bfp.out7_46" + ], + "attrs": { + "axes": { + "type": "int", + "value": [ + "1" + ] + }, + "slice_shape": { + "type": "str", + "value": [ + "batch_size", + "max_length + state_dim1", + "1536" + ] + }, + "start": { + "type": "int", + "value": [ + "0" + ] + }, + "end": { + "type": "str", + "value": [ + "state_dim1" + ] + }, + "step": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/Add", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.23/attn/to_out.0/Add_output_0.out6_1_46_bfp.out7_46", + "/Add_22_output_0.out_35_1_26" + ], + "const_args": [ + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_2_gma" + ], + "out_args": [ + "/transformer_blocks.23/Add_output_0.out10_62" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/norm2/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.23/Add_output_0.out10_62" + ], + "const_args": [ + "/transformer_blocks.23/norm2/Constant_output_0_14_94", + "/transformer_blocks.23/norm2/Constant_1_output_0_14_94" + ], + "out_args": [ + "/transformer_blocks.23/norm2/LayerNormalization_output_0.out14_94_bfp.out15_94" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.23/norm2/LayerNormalization_output_0.out14_94_bfp.out15_94" + ], + "const_args": [ + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_3_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_4" + ], + "out_args": [ + "/transformer_blocks.23/Add_2_output_0.out0_0_94_bfp.out1_63" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/ff/net.0/proj/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.23/Add_2_output_0.out0_0_94_bfp.out1_63" + ], + "const_args": [ + "onnx::MatMul_9470_45_1_93" + ], + "out_args": [ + "/transformer_blocks.23/ff/net.0/Mul_5_output_0.out17_2_46_bfp.out25_97" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "6144" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/ff/net.2/MatMul", + "type": "SDGemm_bfp", + "in_args": [ + "/transformer_blocks.23/ff/net.0/Mul_5_output_0.out17_2_46_bfp.out25_97" + ], + "const_args": [ + "onnx::MatMul_9471_45_1_94" + ], + "out_args": [ + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "6144" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "6144", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.23/Add_3", + "type": "SDGemmMulAdd_bfpbfbf", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98", + "/transformer_blocks.23/Add_output_0.out10_62" + ], + "const_args": [ + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_5_gma" + ], + "out_args": [ + "/transformer_blocks.23/Add_3_output_0.out10_63" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "input_shape_2": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/norm_out/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/transformer_blocks.23/Add_3_output_0.out10_63" + ], + "const_args": [ + "/norm_out/norm/Constant_output_0_14_95", + "/norm_out/norm/Constant_1_output_0_14_95" + ], + "out_args": [ + "/norm_out/norm/LayerNormalization_output_0.out14_95_bfp.out15_95" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/norm_out/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "/norm_out/norm/LayerNormalization_output_0.out14_95_bfp.out15_95" + ], + "const_args": [ + "existing_model.norm_out.linear.weight_5_1_52_27_48_0_existing_model.norm_out.linear.weight_5_1_52_27_48_1" + ], + "out_args": [ + "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_size", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93_SDCastBfp2Bf", + "type": "SDCastBfp2Bf", + "in_args": [ + "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93" + ], + "const_args": [ + "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93_bfp.wts" + ], + "out_args": [ + "/norm_out/Add_2_output_0.out0_0_95" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/proj_out/MatMul", + "type": "SDGemm", + "in_args": [ + "/norm_out/Add_2_output_0.out0_0_95" + ], + "const_args": [ + "onnx::MatMul_9472_45_0_4" + ], + "out_args": [ + "/Reshape_output_0.out17_0_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_size", + "state_dim1", + "64" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "64" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 86560, + "xrt_arg_id": 0, + "packed_tensors": [ + "hidden_states_nhwc.out5_0_0", + "/pos_embed/Reshape_1_output_0.out_35_1_2", + "/time_text_embed/Cast_output_0.out17_4_3", + "pooled_projections.out17_4_1", + "encoder_hidden_states.out17_4_0", + "block_controlnet_hidden_states_0.out_35_1_4", + "block_controlnet_hidden_states_1.out_35_1_5", + "block_controlnet_hidden_states_2.out_35_1_6", + "block_controlnet_hidden_states_3.out_35_1_7", + "block_controlnet_hidden_states_4.out_35_1_8", + "block_controlnet_hidden_states_5.out_35_1_9", + "block_controlnet_hidden_states_6.out_35_1_10", + "block_controlnet_hidden_states_7.out_35_1_11", + "block_controlnet_hidden_states_8.out_35_1_12", + "block_controlnet_hidden_states_9.out_35_1_13", + "block_controlnet_hidden_states_10.out_35_1_14", + "block_controlnet_hidden_states_11.out_35_1_15", + "block_controlnet_hidden_states_12.out_35_1_16", + "block_controlnet_hidden_states_13.out_35_1_17", + "block_controlnet_hidden_states_14.out_35_1_18", + "block_controlnet_hidden_states_15.out_35_1_19", + "block_controlnet_hidden_states_16.out_35_1_20", + "block_controlnet_hidden_states_17.out_35_1_21", + "block_controlnet_hidden_states_18.out_35_1_22", + "block_controlnet_hidden_states_19.out_35_1_23", + "block_controlnet_hidden_states_20.out_35_1_24", + "block_controlnet_hidden_states_21.out_35_1_25", + "block_controlnet_hidden_states_22.out_35_1_26" + ] + }, + "out": { + "buffer_size": 128, + "xrt_arg_id": 1, + "packed_tensors": [ + "/Reshape_output_0.out17_0_0" + ] + }, + "scratch": { + "buffer_size": 1277440, + "xrt_arg_id": 2, + "packed_tensors": [ + "/pos_embed/Transpose_output_0.out5_0_0", + "/pos_embed/Add_2_output_0.out_35_1_2", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3", + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1", + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2", + "/time_text_embed/Add_output_0.out_35_1_3", + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", + "encoder_hidden_states.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0", + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0", + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1", + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_94", + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0", + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1", + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2", + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0", + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", + "/transformer_blocks.0/Add_4_output_0.out10_0", + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3", + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1", + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", + "/transformer_blocks.0/Add_7_output_0.out10_1", + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", + "/transformer_blocks.0/Add_output_0.out10_92", + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2", + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_95", + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", + "/transformer_blocks.0/Add_3_output_0.out10_93", + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_4_bfp.out15_4", + "/Add_output_0.out_35_1_4", + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_5_bfp.out15_5", + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_5_bfp.out1_2", + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_4_bfp.out1_4", + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3", + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4", + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5", + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1", + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2", + "/transformer_blocks.1/Add_output_0.out10_2", + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6", + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3", + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_9", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11", + "/transformer_blocks.1/Add_3_output_0.out10_3", + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", + "/transformer_blocks.1/Add_4_output_0.out10_4", + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7", + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5", + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_10", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12", + "/transformer_blocks.1/Add_7_output_0.out10_5", + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_8_bfp.out15_8", + "/Add_1_output_0.out_35_1_5", + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_9_bfp.out15_9", + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_9_bfp.out1_46", + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_8_bfp.out1_48", + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6", + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7", + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8", + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2", + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5", + "/transformer_blocks.2/Add_4_output_0.out10_48", + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4", + "/transformer_blocks.2/Add_output_0.out10_46", + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10", + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47", + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_13", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15", + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11", + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49", + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_14", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16", + "/transformer_blocks.2/Add_7_output_0.out10_49", + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_12_bfp.out15_12", + "/transformer_blocks.2/Add_3_output_0.out10_47", + "/Add_2_output_0.out_35_1_6", + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_13_bfp.out15_13", + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_13_bfp.out1_65", + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_12_bfp.out1_67", + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9", + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10", + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11", + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3", + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7", + "/transformer_blocks.3/Add_4_output_0.out10_66", + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6", + "/transformer_blocks.3/Add_output_0.out10_64", + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14", + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_66", + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_17", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19", + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15", + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_68", + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_18", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20", + "/transformer_blocks.3/Add_7_output_0.out10_67", + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_16_bfp.out15_16", + "/transformer_blocks.3/Add_3_output_0.out10_65", + "/Add_3_output_0.out_35_1_7", + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_17_bfp.out15_17", + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_17_bfp.out1_69", + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_16_bfp.out1_71", + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12", + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13", + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14", + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4", + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9", + "/transformer_blocks.4/Add_4_output_0.out10_70", + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8", + "/transformer_blocks.4/Add_output_0.out10_68", + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18", + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_70", + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_21", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23", + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19", + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_72", + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_22", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24", + "/transformer_blocks.4/Add_7_output_0.out10_71", + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_20_bfp.out15_20", + "/transformer_blocks.4/Add_3_output_0.out10_69", + "/Add_4_output_0.out_35_1_8", + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_21_bfp.out15_21", + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_21_bfp.out1_73", + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_20_bfp.out1_75", + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15", + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16", + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17", + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5", + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11", + "/transformer_blocks.5/Add_4_output_0.out10_74", + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10", + "/transformer_blocks.5/Add_output_0.out10_72", + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22", + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_74", + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_25", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27", + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23", + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_76", + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_26", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28", + "/transformer_blocks.5/Add_7_output_0.out10_75", + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_24_bfp.out15_24", + "/transformer_blocks.5/Add_3_output_0.out10_73", + "/Add_5_output_0.out_35_1_9", + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_25_bfp.out15_25", + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_25_bfp.out1_77", + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_24_bfp.out1_79", + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18", + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19", + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20", + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6", + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13", + "/transformer_blocks.6/Add_4_output_0.out10_78", + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12", + "/transformer_blocks.6/Add_output_0.out10_76", + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26", + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_78", + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_29", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31", + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27", + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_80", + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_30", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32", + "/transformer_blocks.6/Add_7_output_0.out10_79", + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_28_bfp.out15_28", + "/transformer_blocks.6/Add_3_output_0.out10_77", + "/Add_6_output_0.out_35_1_10", + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_29_bfp.out15_29", + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_29_bfp.out1_81", + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_28_bfp.out1_83", + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21", + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22", + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23", + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7", + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15", + "/transformer_blocks.7/Add_4_output_0.out10_82", + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14", + "/transformer_blocks.7/Add_output_0.out10_80", + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30", + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_82", + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_33", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35", + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31", + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_84", + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_34", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36", + "/transformer_blocks.7/Add_7_output_0.out10_83", + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_32_bfp.out15_32", + "/transformer_blocks.7/Add_3_output_0.out10_81", + "/Add_7_output_0.out_35_1_11", + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_33_bfp.out15_33", + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_33_bfp.out1_85", + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_32_bfp.out1_87", + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24", + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25", + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26", + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8", + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17", + "/transformer_blocks.8/Add_4_output_0.out10_86", + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16", + "/transformer_blocks.8/Add_output_0.out10_84", + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34", + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_86", + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_37", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39", + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35", + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_88", + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_38", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40", + "/transformer_blocks.8/Add_7_output_0.out10_87", + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_36_bfp.out15_36", + "/transformer_blocks.8/Add_3_output_0.out10_85", + "/Add_8_output_0.out_35_1_12", + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_37_bfp.out15_37", + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_37_bfp.out1_89", + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_36_bfp.out1_91", + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27", + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28", + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29", + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9", + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19", + "/transformer_blocks.9/Add_4_output_0.out10_90", + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18", + "/transformer_blocks.9/Add_output_0.out10_88", + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38", + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_90", + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_41", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43", + "/transformer_blocks.9/Add_3_output_0.out10_89", + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39", + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_92", + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_42", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44", + "/transformer_blocks.9/Add_7_output_0.out10_91", + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41", + "/Add_9_output_0.out_35_1_13", + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40", + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6", + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30", + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31", + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32", + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10", + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20", + "/transformer_blocks.10/Add_output_0.out10_6", + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42", + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7", + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_45", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47", + "/transformer_blocks.10/Add_3_output_0.out10_7", + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21", + "/transformer_blocks.10/Add_4_output_0.out10_8", + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43", + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9", + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_46", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48", + "/transformer_blocks.10/Add_7_output_0.out10_9", + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_44_bfp.out15_44", + "/Add_10_output_0.out_35_1_14", + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_45_bfp.out15_45", + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_45_bfp.out1_10", + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_44_bfp.out1_12", + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33", + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34", + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35", + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11", + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22", + "/transformer_blocks.11/Add_output_0.out10_10", + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46", + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11", + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_49", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51", + "/transformer_blocks.11/Add_3_output_0.out10_11", + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23", + "/transformer_blocks.11/Add_4_output_0.out10_12", + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47", + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13", + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_50", + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52", + "/transformer_blocks.11/Add_7_output_0.out10_13", + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_48_bfp.out15_48", + "/Add_11_output_0.out_35_1_15", + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_49_bfp.out15_49", + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_49_bfp.out1_14", + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_48_bfp.out1_16", + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36", + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37", + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38", + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12", + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24", + "/transformer_blocks.12/Add_output_0.out10_14", + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50", + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15", + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_53", + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55", + "/transformer_blocks.12/Add_3_output_0.out10_15", + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25", + "/transformer_blocks.12/Add_4_output_0.out10_16", + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51", + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17", + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_54", + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56", + "/transformer_blocks.12/Add_7_output_0.out10_17", + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_52_bfp.out15_52", + "/Add_12_output_0.out_35_1_16", + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_53_bfp.out15_53", + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_53_bfp.out1_18", + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_52_bfp.out1_20", + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39", + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40", + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41", + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13", + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26", + "/transformer_blocks.13/Add_output_0.out10_18", + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54", + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19", + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_57", + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59", + "/transformer_blocks.13/Add_3_output_0.out10_19", + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27", + "/transformer_blocks.13/Add_4_output_0.out10_20", + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55", + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21", + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_58", + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60", + "/transformer_blocks.13/Add_7_output_0.out10_21", + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_56_bfp.out15_56", + "/Add_13_output_0.out_35_1_17", + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_57_bfp.out15_57", + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_57_bfp.out1_22", + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_56_bfp.out1_24", + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42", + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43", + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44", + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14", + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28", + "/transformer_blocks.14/Add_output_0.out10_22", + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58", + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23", + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_61", + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63", + "/transformer_blocks.14/Add_3_output_0.out10_23", + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29", + "/transformer_blocks.14/Add_4_output_0.out10_24", + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59", + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25", + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_62", + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64", + "/transformer_blocks.14/Add_7_output_0.out10_25", + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_60_bfp.out15_60", + "/Add_14_output_0.out_35_1_18", + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_61_bfp.out15_61", + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_61_bfp.out1_26", + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_60_bfp.out1_28", + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45", + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46", + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47", + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15", + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30", + "/transformer_blocks.15/Add_output_0.out10_26", + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62", + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27", + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_65", + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67", + "/transformer_blocks.15/Add_3_output_0.out10_27", + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31", + "/transformer_blocks.15/Add_4_output_0.out10_28", + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63", + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29", + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_66", + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68", + "/transformer_blocks.15/Add_7_output_0.out10_29", + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_64_bfp.out15_64", + "/Add_15_output_0.out_35_1_19", + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_65_bfp.out15_65", + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_65_bfp.out1_30", + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_64_bfp.out1_32", + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48", + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49", + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50", + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16", + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32", + "/transformer_blocks.16/Add_output_0.out10_30", + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66", + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31", + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_69", + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71", + "/transformer_blocks.16/Add_3_output_0.out10_31", + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33", + "/transformer_blocks.16/Add_4_output_0.out10_32", + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67", + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33", + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_70", + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72", + "/transformer_blocks.16/Add_7_output_0.out10_33", + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_68_bfp.out15_68", + "/Add_16_output_0.out_35_1_20", + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_69_bfp.out15_69", + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_69_bfp.out1_34", + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_68_bfp.out1_36", + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51", + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52", + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53", + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17", + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34", + "/transformer_blocks.17/Add_output_0.out10_34", + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70", + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35", + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_73", + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75", + "/transformer_blocks.17/Add_3_output_0.out10_35", + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35", + "/transformer_blocks.17/Add_4_output_0.out10_36", + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71", + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37", + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_74", + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76", + "/transformer_blocks.17/Add_7_output_0.out10_37", + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_72_bfp.out15_72", + "/Add_17_output_0.out_35_1_21", + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_73_bfp.out15_73", + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_73_bfp.out1_38", + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_72_bfp.out1_40", + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54", + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55", + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56", + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18", + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36", + "/transformer_blocks.18/Add_output_0.out10_38", + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74", + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39", + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_77", + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79", + "/transformer_blocks.18/Add_3_output_0.out10_39", + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37", + "/transformer_blocks.18/Add_4_output_0.out10_40", + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75", + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41", + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_78", + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80", + "/transformer_blocks.18/Add_7_output_0.out10_41", + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_76_bfp.out15_76", + "/Add_18_output_0.out_35_1_22", + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_77_bfp.out15_77", + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_77_bfp.out1_42", + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_76_bfp.out1_44", + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57", + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58", + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59", + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19", + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38", + "/transformer_blocks.19/Add_output_0.out10_42", + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78", + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43", + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_81", + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83", + "/transformer_blocks.19/Add_3_output_0.out10_43", + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39", + "/transformer_blocks.19/Add_4_output_0.out10_44", + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79", + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45", + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_82", + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84", + "/transformer_blocks.19/Add_7_output_0.out10_45", + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_80_bfp.out15_80", + "/Add_19_output_0.out_35_1_23", + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_81_bfp.out15_81", + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_81_bfp.out1_50", + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_80_bfp.out1_52", + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60", + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61", + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62", + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20", + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40", + "/transformer_blocks.20/Add_output_0.out10_50", + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82", + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51", + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_85", + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87", + "/transformer_blocks.20/Add_3_output_0.out10_51", + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41", + "/transformer_blocks.20/Add_4_output_0.out10_52", + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83", + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53", + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_86", + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88", + "/transformer_blocks.20/Add_7_output_0.out10_53", + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_84_bfp.out15_84", + "/Add_20_output_0.out_35_1_24", + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_85_bfp.out15_85", + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_85_bfp.out1_54", + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_84_bfp.out1_56", + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63", + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64", + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65", + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21", + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42", + "/transformer_blocks.21/Add_output_0.out10_54", + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86", + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55", + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_89", + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91", + "/transformer_blocks.21/Add_3_output_0.out10_55", + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43", + "/transformer_blocks.21/Add_4_output_0.out10_56", + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87", + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57", + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_90", + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92", + "/transformer_blocks.21/Add_7_output_0.out10_57", + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_88_bfp.out15_88", + "/Add_21_output_0.out_35_1_25", + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_89_bfp.out15_89", + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_89_bfp.out1_58", + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_88_bfp.out1_60", + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66", + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67", + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68", + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22", + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44", + "/transformer_blocks.22/Add_output_0.out10_58", + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90", + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59", + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_93", + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95", + "/transformer_blocks.22/Add_3_output_0.out10_59", + "/transformer_blocks.22/attn/to_add_out/Add_output_0.out6_1_45_bfp.out7_45", + "/transformer_blocks.22/Add_4_output_0.out10_60", + "/transformer_blocks.22/norm2_context/LayerNormalization_output_0.out14_91_bfp.out15_91", + "/transformer_blocks.22/Add_6_output_0.out0_0_91_bfp.out1_61", + "/transformer_blocks.22/ff_context/net.0/Mul_5_output_0.out17_2_45_bfp.out25_94", + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96", + "/transformer_blocks.22/Add_7_output_0.out10_61", + "/transformer_blocks.23/norm1_context/norm/LayerNormalization_output_0.out14_92_bfp.out15_92", + "/Add_22_output_0.out_35_1_26", + "/transformer_blocks.23/norm1/norm/LayerNormalization_output_0.out14_93_bfp.out15_93", + "/transformer_blocks.23/norm1/Add_2_output_0.out0_0_93_bfp.out1_62", + "/transformer_blocks.23/norm1_context/Add_2_output_0.out0_0_92_bfp.out1_64", + "/transformer_blocks.23/attn/Concat_6_3d.out22_23_bfp.out23_69", + "/transformer_blocks.23/attn/Concat_7_3d.out22_23_bfp.out23_70", + "/transformer_blocks.23/attn/Concat_8_3d.out22_23_bfp.out23_71", + "/transformer_blocks.23/attn/Reshape_6_output_0.out22_23_bfp.out27_0_23", + "/transformer_blocks.23/attn/to_out.0/Add_output_0.out6_1_46_bfp.out7_46", + "/transformer_blocks.23/Add_output_0.out10_62", + "/transformer_blocks.23/norm2/LayerNormalization_output_0.out14_94_bfp.out15_94", + "/transformer_blocks.23/Add_2_output_0.out0_0_94_bfp.out1_63", + "/transformer_blocks.23/ff/net.0/Mul_5_output_0.out17_2_46_bfp.out25_97", + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98", + "/transformer_blocks.23/Add_3_output_0.out10_63", + "/norm_out/norm/LayerNormalization_output_0.out14_95_bfp.out15_95", + "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93", + "/norm_out/Add_2_output_0.out0_0_95" + ] + }, + "const": { + "buffer_size": 2313015936, + "xrt_arg_id": 3, + "packed_tensors": [ + "existing_model.pos_embed.proj.weight", + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_45_0_2", + "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1", + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_45_0_3", + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_45_0_0", + "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0", + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_45_0_1", + "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2", + "encoder_hidden_states.out17_4_0_bfp.wts", + "onnx::MatMul_8733_45_1_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts", + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0", + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1", + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1", + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1", + "onnx::MatMul_8749_onnx::MatMul_8734", + "onnx::MatMul_8750_onnx::MatMul_8735", + "onnx::MatMul_8751_onnx::MatMul_8736", + "onnx::MatMul_8759_45_5_1", + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2_gma", + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3", + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4", + "onnx::MatMul_8763_45_1_2", + "onnx::MatMul_8764_45_1_4", + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma", + "onnx::MatMul_8760_45_5_0", + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma", + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2", + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4", + "onnx::MatMul_8761_45_1_1", + "onnx::MatMul_8762_45_1_3", + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma", + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_5", + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1", + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1", + "onnx::MatMul_8780_onnx::MatMul_8765", + "onnx::MatMul_8781_onnx::MatMul_8766", + "onnx::MatMul_8782_onnx::MatMul_8767", + "onnx::MatMul_8791_45_5_2", + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma", + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6", + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4", + "onnx::MatMul_8792_45_1_5", + "onnx::MatMul_8793_45_1_7", + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma", + "onnx::MatMul_8790_45_5_3", + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma", + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7", + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4", + "onnx::MatMul_8794_45_1_6", + "onnx::MatMul_8795_45_1_8", + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma", + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_9", + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1", + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1", + "onnx::MatMul_8811_onnx::MatMul_8796", + "onnx::MatMul_8812_onnx::MatMul_8797", + "onnx::MatMul_8813_onnx::MatMul_8798", + "onnx::MatMul_8821_45_5_5", + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma", + "onnx::MatMul_8822_45_5_4", + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma", + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10", + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4", + "onnx::MatMul_8823_45_1_9", + "onnx::MatMul_8824_45_1_11", + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11", + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4", + "onnx::MatMul_8825_45_1_10", + "onnx::MatMul_8826_45_1_12", + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma", + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_12", + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma", + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_13", + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1", + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1", + "onnx::MatMul_8842_onnx::MatMul_8827", + "onnx::MatMul_8843_onnx::MatMul_8828", + "onnx::MatMul_8844_onnx::MatMul_8829", + "onnx::MatMul_8852_45_5_7", + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma", + "onnx::MatMul_8853_45_5_6", + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma", + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14", + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4", + "onnx::MatMul_8854_45_1_13", + "onnx::MatMul_8855_45_1_15", + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15", + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4", + "onnx::MatMul_8856_45_1_14", + "onnx::MatMul_8857_45_1_16", + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma", + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_16", + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma", + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_17", + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1", + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1", + "onnx::MatMul_8873_onnx::MatMul_8858", + "onnx::MatMul_8874_onnx::MatMul_8859", + "onnx::MatMul_8875_onnx::MatMul_8860", + "onnx::MatMul_8883_45_5_9", + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma", + "onnx::MatMul_8884_45_5_8", + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma", + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18", + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4", + "onnx::MatMul_8885_45_1_17", + "onnx::MatMul_8886_45_1_19", + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19", + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4", + "onnx::MatMul_8887_45_1_18", + "onnx::MatMul_8888_45_1_20", + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma", + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_20", + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma", + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_21", + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1", + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1", + "onnx::MatMul_8904_onnx::MatMul_8889", + "onnx::MatMul_8905_onnx::MatMul_8890", + "onnx::MatMul_8906_onnx::MatMul_8891", + "onnx::MatMul_8914_45_5_11", + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_2_gma", + "onnx::MatMul_8915_45_5_10", + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma", + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22", + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4", + "onnx::MatMul_8916_45_1_21", + "onnx::MatMul_8917_45_1_23", + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23", + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_3_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_4", + "onnx::MatMul_8918_45_1_22", + "onnx::MatMul_8919_45_1_24", + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_5_gma", + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_24", + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma", + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_25", + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_0_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_1", + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_0_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_1", + "onnx::MatMul_8935_onnx::MatMul_8920", + "onnx::MatMul_8936_onnx::MatMul_8921", + "onnx::MatMul_8937_onnx::MatMul_8922", + "onnx::MatMul_8945_45_5_13", + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_2_gma", + "onnx::MatMul_8946_45_5_12", + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_2_gma", + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26", + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_3_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_4", + "onnx::MatMul_8947_45_1_25", + "onnx::MatMul_8948_45_1_27", + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27", + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_3_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_4", + "onnx::MatMul_8949_45_1_26", + "onnx::MatMul_8950_45_1_28", + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_5_gma", + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_28", + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_5_gma", + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_29", + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_0_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_1", + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_0_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_1", + "onnx::MatMul_8966_onnx::MatMul_8951", + "onnx::MatMul_8967_onnx::MatMul_8952", + "onnx::MatMul_8968_onnx::MatMul_8953", + "onnx::MatMul_8976_45_5_15", + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_2_gma", + "onnx::MatMul_8977_45_5_14", + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_2_gma", + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30", + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_3_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_4", + "onnx::MatMul_8978_45_1_29", + "onnx::MatMul_8979_45_1_31", + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31", + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_3_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_4", + "onnx::MatMul_8980_45_1_30", + "onnx::MatMul_8981_45_1_32", + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_5_gma", + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_32", + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_5_gma", + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_33", + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_0_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_1", + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_0_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_1", + "onnx::MatMul_8997_onnx::MatMul_8982", + "onnx::MatMul_8998_onnx::MatMul_8983", + "onnx::MatMul_8999_onnx::MatMul_8984", + "onnx::MatMul_9007_45_5_17", + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_2_gma", + "onnx::MatMul_9008_45_5_16", + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_2_gma", + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34", + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_3_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_4", + "onnx::MatMul_9009_45_1_33", + "onnx::MatMul_9010_45_1_35", + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35", + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_3_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_4", + "onnx::MatMul_9011_45_1_34", + "onnx::MatMul_9012_45_1_36", + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_5_gma", + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_36", + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_5_gma", + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_37", + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_0_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_1", + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_0_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_1", + "onnx::MatMul_9028_onnx::MatMul_9013", + "onnx::MatMul_9029_onnx::MatMul_9014", + "onnx::MatMul_9030_onnx::MatMul_9015", + "onnx::MatMul_9038_45_5_19", + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_2_gma", + "onnx::MatMul_9039_45_5_18", + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_2_gma", + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38", + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_3_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_4", + "onnx::MatMul_9040_45_1_37", + "onnx::MatMul_9041_45_1_39", + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_5_gma", + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39", + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_3_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_4", + "onnx::MatMul_9042_45_1_38", + "onnx::MatMul_9043_45_1_40", + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_5_gma", + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41", + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40", + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_0_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_1", + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_0_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_1", + "onnx::MatMul_9059_onnx::MatMul_9044", + "onnx::MatMul_9060_onnx::MatMul_9045", + "onnx::MatMul_9061_onnx::MatMul_9046", + "onnx::MatMul_9070_45_5_20", + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_2_gma", + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42", + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_3_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_4", + "onnx::MatMul_9071_45_1_41", + "onnx::MatMul_9072_45_1_43", + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_5_gma", + "onnx::MatMul_9069_45_5_21", + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_2_gma", + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43", + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_3_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_4", + "onnx::MatMul_9073_45_1_42", + "onnx::MatMul_9074_45_1_44", + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_5_gma", + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_44", + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_45", + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_0_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_1", + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_0_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_1", + "onnx::MatMul_9090_onnx::MatMul_9075", + "onnx::MatMul_9091_onnx::MatMul_9076", + "onnx::MatMul_9092_onnx::MatMul_9077", + "onnx::MatMul_9101_45_5_22", + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_2_gma", + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46", + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_3_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_4", + "onnx::MatMul_9102_45_1_45", + "onnx::MatMul_9103_45_1_47", + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_5_gma", + "onnx::MatMul_9100_45_5_23", + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_2_gma", + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47", + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47", + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_3_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_4", + "onnx::MatMul_9104_45_1_46", + "onnx::MatMul_9105_45_1_48", + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_5_gma", + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_48", + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_48", + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_49", + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_49", + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_0_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_1", + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_0_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_1", + "onnx::MatMul_9121_onnx::MatMul_9106", + "onnx::MatMul_9122_onnx::MatMul_9107", + "onnx::MatMul_9123_onnx::MatMul_9108", + "onnx::MatMul_9132_45_5_24", + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_2_gma", + "/transformer_blocks.12/norm2/Constant_output_0_14_50", + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50", + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_3_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_4", + "onnx::MatMul_9133_45_1_49", + "onnx::MatMul_9134_45_1_51", + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_5_gma", + "onnx::MatMul_9131_45_5_25", + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_2_gma", + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51", + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51", + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_3_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_4", + "onnx::MatMul_9135_45_1_50", + "onnx::MatMul_9136_45_1_52", + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_5_gma", + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_52", + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_52", + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_53", + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_53", + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_0_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_1", + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_0_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_1", + "onnx::MatMul_9152_onnx::MatMul_9137", + "onnx::MatMul_9153_onnx::MatMul_9138", + "onnx::MatMul_9154_onnx::MatMul_9139", + "onnx::MatMul_9163_45_5_26", + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_2_gma", + "/transformer_blocks.13/norm2/Constant_output_0_14_54", + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54", + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_3_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_4", + "onnx::MatMul_9164_45_1_53", + "onnx::MatMul_9165_45_1_55", + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_5_gma", + "onnx::MatMul_9162_45_5_27", + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_2_gma", + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55", + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55", + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_3_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_4", + "onnx::MatMul_9166_45_1_54", + "onnx::MatMul_9167_45_1_56", + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_5_gma", + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_56", + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_56", + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_57", + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_57", + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_0_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_1", + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_0_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_1", + "onnx::MatMul_9183_onnx::MatMul_9168", + "onnx::MatMul_9184_onnx::MatMul_9169", + "onnx::MatMul_9185_onnx::MatMul_9170", + "onnx::MatMul_9194_45_5_28", + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_2_gma", + "/transformer_blocks.14/norm2/Constant_output_0_14_58", + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58", + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_3_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_4", + "onnx::MatMul_9195_45_1_57", + "onnx::MatMul_9196_45_1_59", + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_5_gma", + "onnx::MatMul_9193_45_5_29", + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_2_gma", + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59", + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59", + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_3_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_4", + "onnx::MatMul_9197_45_1_58", + "onnx::MatMul_9198_45_1_60", + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_5_gma", + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_60", + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_60", + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_61", + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_61", + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_0_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_1", + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_0_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_1", + "onnx::MatMul_9214_onnx::MatMul_9199", + "onnx::MatMul_9215_onnx::MatMul_9200", + "onnx::MatMul_9216_onnx::MatMul_9201", + "onnx::MatMul_9225_45_5_30", + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_2_gma", + "/transformer_blocks.15/norm2/Constant_output_0_14_62", + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62", + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_3_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_4", + "onnx::MatMul_9226_45_1_61", + "onnx::MatMul_9227_45_1_63", + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_5_gma", + "onnx::MatMul_9224_45_5_31", + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_2_gma", + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63", + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63", + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_3_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_4", + "onnx::MatMul_9228_45_1_62", + "onnx::MatMul_9229_45_1_64", + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_5_gma", + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_64", + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_64", + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_65", + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_65", + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_0_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_1", + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_0_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_1", + "onnx::MatMul_9245_onnx::MatMul_9230", + "onnx::MatMul_9246_onnx::MatMul_9231", + "onnx::MatMul_9247_onnx::MatMul_9232", + "onnx::MatMul_9256_45_5_32", + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_2_gma", + "/transformer_blocks.16/norm2/Constant_output_0_14_66", + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66", + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_3_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_4", + "onnx::MatMul_9257_45_1_65", + "onnx::MatMul_9258_45_1_67", + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_5_gma", + "onnx::MatMul_9255_45_5_33", + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_2_gma", + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67", + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67", + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_3_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_4", + "onnx::MatMul_9259_45_1_66", + "onnx::MatMul_9260_45_1_68", + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_5_gma", + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_68", + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_68", + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_69", + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_69", + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_0_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_1", + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_0_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_1", + "onnx::MatMul_9276_onnx::MatMul_9261", + "onnx::MatMul_9277_onnx::MatMul_9262", + "onnx::MatMul_9278_onnx::MatMul_9263", + "onnx::MatMul_9287_45_5_34", + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_2_gma", + "/transformer_blocks.17/norm2/Constant_output_0_14_70", + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70", + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_3_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_4", + "onnx::MatMul_9288_45_1_69", + "onnx::MatMul_9289_45_1_71", + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_5_gma", + "onnx::MatMul_9286_45_5_35", + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_2_gma", + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71", + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71", + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_3_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_4", + "onnx::MatMul_9290_45_1_70", + "onnx::MatMul_9291_45_1_72", + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_5_gma", + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_72", + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_72", + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_73", + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_73", + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_0_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_1", + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_0_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_1", + "onnx::MatMul_9307_onnx::MatMul_9292", + "onnx::MatMul_9308_onnx::MatMul_9293", + "onnx::MatMul_9309_onnx::MatMul_9294", + "onnx::MatMul_9318_45_5_36", + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_2_gma", + "/transformer_blocks.18/norm2/Constant_output_0_14_74", + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74", + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_3_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_4", + "onnx::MatMul_9319_45_1_73", + "onnx::MatMul_9320_45_1_75", + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_5_gma", + "onnx::MatMul_9317_45_5_37", + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_2_gma", + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75", + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75", + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_3_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_4", + "onnx::MatMul_9321_45_1_74", + "onnx::MatMul_9322_45_1_76", + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_5_gma", + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_76", + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_76", + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_77", + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_77", + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_0_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_1", + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_0_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_1", + "onnx::MatMul_9338_onnx::MatMul_9323", + "onnx::MatMul_9339_onnx::MatMul_9324", + "onnx::MatMul_9340_onnx::MatMul_9325", + "onnx::MatMul_9349_45_5_38", + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_2_gma", + "/transformer_blocks.19/norm2/Constant_output_0_14_78", + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78", + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_3_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_4", + "onnx::MatMul_9350_45_1_77", + "onnx::MatMul_9351_45_1_79", + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_5_gma", + "onnx::MatMul_9348_45_5_39", + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_2_gma", + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79", + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79", + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_3_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_4", + "onnx::MatMul_9352_45_1_78", + "onnx::MatMul_9353_45_1_80", + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_5_gma", + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_80", + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_80", + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_81", + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_81", + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_0_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_1", + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_0_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_1", + "onnx::MatMul_9369_onnx::MatMul_9354", + "onnx::MatMul_9370_onnx::MatMul_9355", + "onnx::MatMul_9371_onnx::MatMul_9356", + "onnx::MatMul_9380_45_5_40", + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_2_gma", + "/transformer_blocks.20/norm2/Constant_output_0_14_82", + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82", + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_3_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_4", + "onnx::MatMul_9381_45_1_81", + "onnx::MatMul_9382_45_1_83", + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_5_gma", + "onnx::MatMul_9379_45_5_41", + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_2_gma", + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83", + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83", + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_3_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_4", + "onnx::MatMul_9383_45_1_82", + "onnx::MatMul_9384_45_1_84", + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_5_gma", + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_84", + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_84", + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_85", + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_85", + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_0_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_1", + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_0_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_1", + "onnx::MatMul_9400_onnx::MatMul_9385", + "onnx::MatMul_9401_onnx::MatMul_9386", + "onnx::MatMul_9402_onnx::MatMul_9387", + "onnx::MatMul_9411_45_5_42", + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_2_gma", + "/transformer_blocks.21/norm2/Constant_output_0_14_86", + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86", + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_3_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_4", + "onnx::MatMul_9412_45_1_85", + "onnx::MatMul_9413_45_1_87", + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_5_gma", + "onnx::MatMul_9410_45_5_43", + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_2_gma", + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87", + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87", + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_3_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_4", + "onnx::MatMul_9414_45_1_86", + "onnx::MatMul_9415_45_1_88", + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_5_gma", + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_88", + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_88", + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_89", + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_89", + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_0_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_1", + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_0_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_1", + "onnx::MatMul_9431_onnx::MatMul_9416", + "onnx::MatMul_9432_onnx::MatMul_9417", + "onnx::MatMul_9433_onnx::MatMul_9418", + "onnx::MatMul_9442_45_5_44", + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_2_gma", + "/transformer_blocks.22/norm2/Constant_output_0_14_90", + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90", + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_3_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_4", + "onnx::MatMul_9443_45_1_89", + "onnx::MatMul_9444_45_1_91", + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_5_gma", + "onnx::MatMul_9441_45_5_45", + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_2_gma", + "/transformer_blocks.22/norm2_context/Constant_output_0_14_91", + "/transformer_blocks.22/norm2_context/Constant_1_output_0_14_91", + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_3_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_4", + "onnx::MatMul_9445_45_1_90", + "onnx::MatMul_9446_45_1_92", + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_5_gma", + "/transformer_blocks.23/norm1_context/norm/Constant_output_0_14_92", + "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0_14_92", + "/transformer_blocks.23/norm1/norm/Constant_output_0_14_93", + "/transformer_blocks.23/norm1/norm/Constant_1_output_0_14_93", + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_0_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_1", + "existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_0_existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_1", + "onnx::MatMul_9462_onnx::MatMul_9447", + "onnx::MatMul_9463_onnx::MatMul_9448", + "onnx::MatMul_9464_onnx::MatMul_9449", + "onnx::MatMul_9469_45_5_46", + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_2_gma", + "/transformer_blocks.23/norm2/Constant_output_0_14_94", + "/transformer_blocks.23/norm2/Constant_1_output_0_14_94", + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_3_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_4", + "onnx::MatMul_9470_45_1_93", + "onnx::MatMul_9471_45_1_94", + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_5_gma", + "/norm_out/norm/Constant_output_0_14_95", + "/norm_out/norm/Constant_1_output_0_14_95", + "existing_model.norm_out.linear.weight_5_1_52_27_48_0_existing_model.norm_out.linear.weight_5_1_52_27_48_1", + "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93_bfp.wts", + "onnx::MatMul_9472_45_0_4" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "hidden_states_nhwc.out5_0_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 16 + ], + "size_in_bytes": 32, + "op_tensor_size": 32, + "dynamic_shapes": [ + "batch_size", + "w", + "h", + "False" + ], + "offset": 0 + }, + "/pos_embed/Reshape_1_output_0.out_35_1_2": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "False", + "state_dim1", + "False" + ], + "offset": 32 + }, + "/time_text_embed/Cast_output_0.out17_4_3": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 3104 + }, + "pooled_projections.out17_4_1": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 2048 + ], + "size_in_bytes": 4096, + "op_tensor_size": 4096, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 3616 + }, + "encoder_hidden_states.out17_4_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 4096 + ], + "size_in_bytes": 8192, + "op_tensor_size": 8192, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 7712 + }, + "block_controlnet_hidden_states_0.out_35_1_4": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 15904 + }, + "block_controlnet_hidden_states_1.out_35_1_5": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 18976 + }, + "block_controlnet_hidden_states_2.out_35_1_6": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 22048 + }, + "block_controlnet_hidden_states_3.out_35_1_7": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 25120 + }, + "block_controlnet_hidden_states_4.out_35_1_8": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 28192 + }, + "block_controlnet_hidden_states_5.out_35_1_9": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 31264 + }, + "block_controlnet_hidden_states_6.out_35_1_10": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 34336 + }, + "block_controlnet_hidden_states_7.out_35_1_11": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 37408 + }, + "block_controlnet_hidden_states_8.out_35_1_12": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 40480 + }, + "block_controlnet_hidden_states_9.out_35_1_13": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 43552 + }, + "block_controlnet_hidden_states_10.out_35_1_14": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 46624 + }, + "block_controlnet_hidden_states_11.out_35_1_15": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 49696 + }, + "block_controlnet_hidden_states_12.out_35_1_16": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 52768 + }, + "block_controlnet_hidden_states_13.out_35_1_17": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 55840 + }, + "block_controlnet_hidden_states_14.out_35_1_18": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 58912 + }, + "block_controlnet_hidden_states_15.out_35_1_19": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 61984 + }, + "block_controlnet_hidden_states_16.out_35_1_20": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 65056 + }, + "block_controlnet_hidden_states_17.out_35_1_21": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 68128 + }, + "block_controlnet_hidden_states_18.out_35_1_22": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 71200 + }, + "block_controlnet_hidden_states_19.out_35_1_23": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 74272 + }, + "block_controlnet_hidden_states_20.out_35_1_24": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 77344 + }, + "block_controlnet_hidden_states_21.out_35_1_25": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 80416 + }, + "block_controlnet_hidden_states_22.out_35_1_26": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 83488 + }, + "/Reshape_output_0.out17_0_0": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 2, + 2, + 16 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "dynamic_shapes": [ + "batch_size", + "floor(w/2)", + "floor(h/2)", + "False", + "False", + "False" + ], + "offset": 0 + }, + "/pos_embed/Transpose_output_0.out5_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 0 + }, + "/pos_embed/Add_2_output_0.out_35_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 3072 + }, + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 6144 + }, + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 9216 + }, + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 12288 + }, + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 15360 + }, + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 18432 + }, + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 21504 + }, + "/time_text_embed/Add_output_0.out_35_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 24576 + }, + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "False" + ], + "offset": 27648 + }, + "encoder_hidden_states.out17_4_0_bfp.out25_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 4096 + ], + "size_in_bytes": 4096, + "op_tensor_size": 4096, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 30720 + }, + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 34816 + }, + "/context_embedder/Add_output_0.out17_4_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 36352 + }, + "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 39424 + }, + "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 40960 + }, + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 42496 + }, + "/transformer_blocks.0/norm1/Add_2_output_0.out0_0_1_bfp.out1_94": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 44032 + }, + "/transformer_blocks.0/attn/Concat_6_3d.out22_0_bfp.out23_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 45568 + }, + "/transformer_blocks.0/attn/Concat_7_3d.out22_0_bfp.out23_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 47104 + }, + "/transformer_blocks.0/attn/Concat_8_3d.out22_0_bfp.out23_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 48640 + }, + "/transformer_blocks.0/attn/Reshape_6_output_0.out22_0_bfp.out27_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 50176 + }, + "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 51712 + }, + "/transformer_blocks.0/Add_4_output_0.out10_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 53248 + }, + "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 56320 + }, + "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 57856 + }, + "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 59392 + }, + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 65536 + }, + "/transformer_blocks.0/Add_7_output_0.out10_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 67072 + }, + "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 70144 + }, + "/transformer_blocks.0/Add_output_0.out10_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 71680 + }, + "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 74752 + }, + "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 76288 + }, + "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 77824 + }, + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 83968 + }, + "/transformer_blocks.0/Add_3_output_0.out10_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 85504 + }, + "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_4_bfp.out15_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 88576 + }, + "/Add_output_0.out_35_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 90112 + }, + "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_5_bfp.out15_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 93184 + }, + "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_5_bfp.out1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 94720 + }, + "/transformer_blocks.1/norm1_context/Add_2_output_0.out0_0_4_bfp.out1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 96256 + }, + "/transformer_blocks.1/attn/Concat_6_3d.out22_1_bfp.out23_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 97792 + }, + "/transformer_blocks.1/attn/Concat_7_3d.out22_1_bfp.out23_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 99328 + }, + "/transformer_blocks.1/attn/Concat_8_3d.out22_1_bfp.out23_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 100864 + }, + "/transformer_blocks.1/attn/Reshape_6_output_0.out22_1_bfp.out27_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 102400 + }, + "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 103936 + }, + "/transformer_blocks.1/Add_output_0.out10_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 105472 + }, + "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 108544 + }, + "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 110080 + }, + "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 111616 + }, + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 117760 + }, + "/transformer_blocks.1/Add_3_output_0.out10_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 119296 + }, + "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 122368 + }, + "/transformer_blocks.1/Add_4_output_0.out10_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 123904 + }, + "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 126976 + }, + "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 128512 + }, + "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 130048 + }, + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 136192 + }, + "/transformer_blocks.1/Add_7_output_0.out10_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 137728 + }, + "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_8_bfp.out15_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 140800 + }, + "/Add_1_output_0.out_35_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 142336 + }, + "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_9_bfp.out15_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 145408 + }, + "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_9_bfp.out1_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 146944 + }, + "/transformer_blocks.2/norm1_context/Add_2_output_0.out0_0_8_bfp.out1_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 148480 + }, + "/transformer_blocks.2/attn/Concat_6_3d.out22_2_bfp.out23_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 150016 + }, + "/transformer_blocks.2/attn/Concat_7_3d.out22_2_bfp.out23_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 151552 + }, + "/transformer_blocks.2/attn/Concat_8_3d.out22_2_bfp.out23_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 153088 + }, + "/transformer_blocks.2/attn/Reshape_6_output_0.out22_2_bfp.out27_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 154624 + }, + "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 156160 + }, + "/transformer_blocks.2/Add_4_output_0.out10_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 157696 + }, + "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 160768 + }, + "/transformer_blocks.2/Add_output_0.out10_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 162304 + }, + "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 165376 + }, + "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 166912 + }, + "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 168448 + }, + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 174592 + }, + "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 176128 + }, + "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 177664 + }, + "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 179200 + }, + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 185344 + }, + "/transformer_blocks.2/Add_7_output_0.out10_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 186880 + }, + "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_12_bfp.out15_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 189952 + }, + "/transformer_blocks.2/Add_3_output_0.out10_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 191488 + }, + "/Add_2_output_0.out_35_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 194560 + }, + "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_13_bfp.out15_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 197632 + }, + "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_13_bfp.out1_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 199168 + }, + "/transformer_blocks.3/norm1_context/Add_2_output_0.out0_0_12_bfp.out1_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 200704 + }, + "/transformer_blocks.3/attn/Concat_6_3d.out22_3_bfp.out23_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 202240 + }, + "/transformer_blocks.3/attn/Concat_7_3d.out22_3_bfp.out23_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 203776 + }, + "/transformer_blocks.3/attn/Concat_8_3d.out22_3_bfp.out23_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 205312 + }, + "/transformer_blocks.3/attn/Reshape_6_output_0.out22_3_bfp.out27_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 206848 + }, + "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 208384 + }, + "/transformer_blocks.3/Add_4_output_0.out10_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 209920 + }, + "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 212992 + }, + "/transformer_blocks.3/Add_output_0.out10_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 214528 + }, + "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 217600 + }, + "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 219136 + }, + "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 220672 + }, + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 226816 + }, + "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 228352 + }, + "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 229888 + }, + "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 231424 + }, + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 237568 + }, + "/transformer_blocks.3/Add_7_output_0.out10_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 239104 + }, + "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_16_bfp.out15_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 242176 + }, + "/transformer_blocks.3/Add_3_output_0.out10_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 243712 + }, + "/Add_3_output_0.out_35_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 246784 + }, + "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_17_bfp.out15_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 249856 + }, + "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_17_bfp.out1_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 251392 + }, + "/transformer_blocks.4/norm1_context/Add_2_output_0.out0_0_16_bfp.out1_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 252928 + }, + "/transformer_blocks.4/attn/Concat_6_3d.out22_4_bfp.out23_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 254464 + }, + "/transformer_blocks.4/attn/Concat_7_3d.out22_4_bfp.out23_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 256000 + }, + "/transformer_blocks.4/attn/Concat_8_3d.out22_4_bfp.out23_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 257536 + }, + "/transformer_blocks.4/attn/Reshape_6_output_0.out22_4_bfp.out27_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 259072 + }, + "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 260608 + }, + "/transformer_blocks.4/Add_4_output_0.out10_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 262144 + }, + "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 265216 + }, + "/transformer_blocks.4/Add_output_0.out10_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 266752 + }, + "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 269824 + }, + "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 271360 + }, + "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 272896 + }, + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 279040 + }, + "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 280576 + }, + "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 282112 + }, + "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 283648 + }, + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 289792 + }, + "/transformer_blocks.4/Add_7_output_0.out10_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 291328 + }, + "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_20_bfp.out15_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 294400 + }, + "/transformer_blocks.4/Add_3_output_0.out10_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 295936 + }, + "/Add_4_output_0.out_35_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 299008 + }, + "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_21_bfp.out15_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 302080 + }, + "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_21_bfp.out1_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 303616 + }, + "/transformer_blocks.5/norm1_context/Add_2_output_0.out0_0_20_bfp.out1_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 305152 + }, + "/transformer_blocks.5/attn/Concat_6_3d.out22_5_bfp.out23_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 306688 + }, + "/transformer_blocks.5/attn/Concat_7_3d.out22_5_bfp.out23_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 308224 + }, + "/transformer_blocks.5/attn/Concat_8_3d.out22_5_bfp.out23_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 309760 + }, + "/transformer_blocks.5/attn/Reshape_6_output_0.out22_5_bfp.out27_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 311296 + }, + "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 312832 + }, + "/transformer_blocks.5/Add_4_output_0.out10_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 314368 + }, + "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 317440 + }, + "/transformer_blocks.5/Add_output_0.out10_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 318976 + }, + "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 322048 + }, + "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 323584 + }, + "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 325120 + }, + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 331264 + }, + "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 332800 + }, + "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 334336 + }, + "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 335872 + }, + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 342016 + }, + "/transformer_blocks.5/Add_7_output_0.out10_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 343552 + }, + "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_24_bfp.out15_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 346624 + }, + "/transformer_blocks.5/Add_3_output_0.out10_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 348160 + }, + "/Add_5_output_0.out_35_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 351232 + }, + "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_25_bfp.out15_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 354304 + }, + "/transformer_blocks.6/norm1/Add_2_output_0.out0_0_25_bfp.out1_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 355840 + }, + "/transformer_blocks.6/norm1_context/Add_2_output_0.out0_0_24_bfp.out1_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 357376 + }, + "/transformer_blocks.6/attn/Concat_6_3d.out22_6_bfp.out23_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 358912 + }, + "/transformer_blocks.6/attn/Concat_7_3d.out22_6_bfp.out23_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 360448 + }, + "/transformer_blocks.6/attn/Concat_8_3d.out22_6_bfp.out23_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 361984 + }, + "/transformer_blocks.6/attn/Reshape_6_output_0.out22_6_bfp.out27_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 363520 + }, + "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 365056 + }, + "/transformer_blocks.6/Add_4_output_0.out10_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 366592 + }, + "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 369664 + }, + "/transformer_blocks.6/Add_output_0.out10_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 371200 + }, + "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 374272 + }, + "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 375808 + }, + "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 377344 + }, + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 383488 + }, + "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 385024 + }, + "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 386560 + }, + "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 388096 + }, + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 394240 + }, + "/transformer_blocks.6/Add_7_output_0.out10_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 395776 + }, + "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_28_bfp.out15_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 398848 + }, + "/transformer_blocks.6/Add_3_output_0.out10_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 400384 + }, + "/Add_6_output_0.out_35_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 403456 + }, + "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_29_bfp.out15_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 406528 + }, + "/transformer_blocks.7/norm1/Add_2_output_0.out0_0_29_bfp.out1_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 408064 + }, + "/transformer_blocks.7/norm1_context/Add_2_output_0.out0_0_28_bfp.out1_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 409600 + }, + "/transformer_blocks.7/attn/Concat_6_3d.out22_7_bfp.out23_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 411136 + }, + "/transformer_blocks.7/attn/Concat_7_3d.out22_7_bfp.out23_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 412672 + }, + "/transformer_blocks.7/attn/Concat_8_3d.out22_7_bfp.out23_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 414208 + }, + "/transformer_blocks.7/attn/Reshape_6_output_0.out22_7_bfp.out27_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 415744 + }, + "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 417280 + }, + "/transformer_blocks.7/Add_4_output_0.out10_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 418816 + }, + "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 421888 + }, + "/transformer_blocks.7/Add_output_0.out10_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 423424 + }, + "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 426496 + }, + "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 428032 + }, + "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 429568 + }, + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 435712 + }, + "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 437248 + }, + "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 438784 + }, + "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 440320 + }, + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 446464 + }, + "/transformer_blocks.7/Add_7_output_0.out10_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 448000 + }, + "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_32_bfp.out15_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 451072 + }, + "/transformer_blocks.7/Add_3_output_0.out10_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 452608 + }, + "/Add_7_output_0.out_35_1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 455680 + }, + "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_33_bfp.out15_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 458752 + }, + "/transformer_blocks.8/norm1/Add_2_output_0.out0_0_33_bfp.out1_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 460288 + }, + "/transformer_blocks.8/norm1_context/Add_2_output_0.out0_0_32_bfp.out1_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 461824 + }, + "/transformer_blocks.8/attn/Concat_6_3d.out22_8_bfp.out23_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 463360 + }, + "/transformer_blocks.8/attn/Concat_7_3d.out22_8_bfp.out23_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 464896 + }, + "/transformer_blocks.8/attn/Concat_8_3d.out22_8_bfp.out23_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 466432 + }, + "/transformer_blocks.8/attn/Reshape_6_output_0.out22_8_bfp.out27_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 467968 + }, + "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 469504 + }, + "/transformer_blocks.8/Add_4_output_0.out10_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 471040 + }, + "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 474112 + }, + "/transformer_blocks.8/Add_output_0.out10_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 475648 + }, + "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 478720 + }, + "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 480256 + }, + "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 481792 + }, + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 487936 + }, + "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 489472 + }, + "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 491008 + }, + "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 492544 + }, + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 498688 + }, + "/transformer_blocks.8/Add_7_output_0.out10_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 500224 + }, + "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_36_bfp.out15_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 503296 + }, + "/transformer_blocks.8/Add_3_output_0.out10_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 504832 + }, + "/Add_8_output_0.out_35_1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 507904 + }, + "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_37_bfp.out15_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 510976 + }, + "/transformer_blocks.9/norm1/Add_2_output_0.out0_0_37_bfp.out1_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 512512 + }, + "/transformer_blocks.9/norm1_context/Add_2_output_0.out0_0_36_bfp.out1_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 514048 + }, + "/transformer_blocks.9/attn/Concat_6_3d.out22_9_bfp.out23_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 515584 + }, + "/transformer_blocks.9/attn/Concat_7_3d.out22_9_bfp.out23_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 517120 + }, + "/transformer_blocks.9/attn/Concat_8_3d.out22_9_bfp.out23_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 518656 + }, + "/transformer_blocks.9/attn/Reshape_6_output_0.out22_9_bfp.out27_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 520192 + }, + "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 521728 + }, + "/transformer_blocks.9/Add_4_output_0.out10_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 523264 + }, + "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 526336 + }, + "/transformer_blocks.9/Add_output_0.out10_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 527872 + }, + "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 530944 + }, + "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 532480 + }, + "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 534016 + }, + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 540160 + }, + "/transformer_blocks.9/Add_3_output_0.out10_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 541696 + }, + "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 544768 + }, + "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 546304 + }, + "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 547840 + }, + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 553984 + }, + "/transformer_blocks.9/Add_7_output_0.out10_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 555520 + }, + "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 558592 + }, + "/Add_9_output_0.out_35_1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 560128 + }, + "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 563200 + }, + "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 564736 + }, + "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 566272 + }, + "/transformer_blocks.10/attn/Concat_6_3d.out22_10_bfp.out23_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 567808 + }, + "/transformer_blocks.10/attn/Concat_7_3d.out22_10_bfp.out23_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 569344 + }, + "/transformer_blocks.10/attn/Concat_8_3d.out22_10_bfp.out23_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 570880 + }, + "/transformer_blocks.10/attn/Reshape_6_output_0.out22_10_bfp.out27_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 572416 + }, + "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 573952 + }, + "/transformer_blocks.10/Add_output_0.out10_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 575488 + }, + "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 578560 + }, + "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 580096 + }, + "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 581632 + }, + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 587776 + }, + "/transformer_blocks.10/Add_3_output_0.out10_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 589312 + }, + "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 592384 + }, + "/transformer_blocks.10/Add_4_output_0.out10_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 593920 + }, + "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 596992 + }, + "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 598528 + }, + "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 600064 + }, + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 606208 + }, + "/transformer_blocks.10/Add_7_output_0.out10_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 607744 + }, + "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_44_bfp.out15_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 610816 + }, + "/Add_10_output_0.out_35_1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 612352 + }, + "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_45_bfp.out15_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 615424 + }, + "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_45_bfp.out1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 616960 + }, + "/transformer_blocks.11/norm1_context/Add_2_output_0.out0_0_44_bfp.out1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 618496 + }, + "/transformer_blocks.11/attn/Concat_6_3d.out22_11_bfp.out23_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 620032 + }, + "/transformer_blocks.11/attn/Concat_7_3d.out22_11_bfp.out23_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 621568 + }, + "/transformer_blocks.11/attn/Concat_8_3d.out22_11_bfp.out23_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 623104 + }, + "/transformer_blocks.11/attn/Reshape_6_output_0.out22_11_bfp.out27_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 624640 + }, + "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 626176 + }, + "/transformer_blocks.11/Add_output_0.out10_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 627712 + }, + "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 630784 + }, + "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 632320 + }, + "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 633856 + }, + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 640000 + }, + "/transformer_blocks.11/Add_3_output_0.out10_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 641536 + }, + "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 644608 + }, + "/transformer_blocks.11/Add_4_output_0.out10_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 646144 + }, + "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 649216 + }, + "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 650752 + }, + "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 652288 + }, + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 658432 + }, + "/transformer_blocks.11/Add_7_output_0.out10_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 659968 + }, + "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_48_bfp.out15_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 663040 + }, + "/Add_11_output_0.out_35_1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 664576 + }, + "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_49_bfp.out15_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 667648 + }, + "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_49_bfp.out1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 669184 + }, + "/transformer_blocks.12/norm1_context/Add_2_output_0.out0_0_48_bfp.out1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 670720 + }, + "/transformer_blocks.12/attn/Concat_6_3d.out22_12_bfp.out23_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 672256 + }, + "/transformer_blocks.12/attn/Concat_7_3d.out22_12_bfp.out23_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 673792 + }, + "/transformer_blocks.12/attn/Concat_8_3d.out22_12_bfp.out23_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 675328 + }, + "/transformer_blocks.12/attn/Reshape_6_output_0.out22_12_bfp.out27_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 676864 + }, + "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 678400 + }, + "/transformer_blocks.12/Add_output_0.out10_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 679936 + }, + "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 683008 + }, + "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 684544 + }, + "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 686080 + }, + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 692224 + }, + "/transformer_blocks.12/Add_3_output_0.out10_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 693760 + }, + "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 696832 + }, + "/transformer_blocks.12/Add_4_output_0.out10_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 698368 + }, + "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 701440 + }, + "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 702976 + }, + "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 704512 + }, + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 710656 + }, + "/transformer_blocks.12/Add_7_output_0.out10_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 712192 + }, + "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_52_bfp.out15_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 715264 + }, + "/Add_12_output_0.out_35_1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 716800 + }, + "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_53_bfp.out15_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 719872 + }, + "/transformer_blocks.13/norm1/Add_2_output_0.out0_0_53_bfp.out1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 721408 + }, + "/transformer_blocks.13/norm1_context/Add_2_output_0.out0_0_52_bfp.out1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 722944 + }, + "/transformer_blocks.13/attn/Concat_6_3d.out22_13_bfp.out23_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 724480 + }, + "/transformer_blocks.13/attn/Concat_7_3d.out22_13_bfp.out23_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 726016 + }, + "/transformer_blocks.13/attn/Concat_8_3d.out22_13_bfp.out23_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 727552 + }, + "/transformer_blocks.13/attn/Reshape_6_output_0.out22_13_bfp.out27_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 729088 + }, + "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 730624 + }, + "/transformer_blocks.13/Add_output_0.out10_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 732160 + }, + "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 735232 + }, + "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 736768 + }, + "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 738304 + }, + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 744448 + }, + "/transformer_blocks.13/Add_3_output_0.out10_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 745984 + }, + "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 749056 + }, + "/transformer_blocks.13/Add_4_output_0.out10_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 750592 + }, + "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 753664 + }, + "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 755200 + }, + "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 756736 + }, + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 762880 + }, + "/transformer_blocks.13/Add_7_output_0.out10_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 764416 + }, + "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_56_bfp.out15_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 767488 + }, + "/Add_13_output_0.out_35_1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 769024 + }, + "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_57_bfp.out15_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 772096 + }, + "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_57_bfp.out1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 773632 + }, + "/transformer_blocks.14/norm1_context/Add_2_output_0.out0_0_56_bfp.out1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 775168 + }, + "/transformer_blocks.14/attn/Concat_6_3d.out22_14_bfp.out23_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 776704 + }, + "/transformer_blocks.14/attn/Concat_7_3d.out22_14_bfp.out23_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 778240 + }, + "/transformer_blocks.14/attn/Concat_8_3d.out22_14_bfp.out23_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 779776 + }, + "/transformer_blocks.14/attn/Reshape_6_output_0.out22_14_bfp.out27_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 781312 + }, + "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 782848 + }, + "/transformer_blocks.14/Add_output_0.out10_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 784384 + }, + "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 787456 + }, + "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 788992 + }, + "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 790528 + }, + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 796672 + }, + "/transformer_blocks.14/Add_3_output_0.out10_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 798208 + }, + "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 801280 + }, + "/transformer_blocks.14/Add_4_output_0.out10_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 802816 + }, + "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 805888 + }, + "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 807424 + }, + "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 808960 + }, + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 815104 + }, + "/transformer_blocks.14/Add_7_output_0.out10_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 816640 + }, + "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_60_bfp.out15_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 819712 + }, + "/Add_14_output_0.out_35_1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 821248 + }, + "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_61_bfp.out15_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 824320 + }, + "/transformer_blocks.15/norm1/Add_2_output_0.out0_0_61_bfp.out1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 825856 + }, + "/transformer_blocks.15/norm1_context/Add_2_output_0.out0_0_60_bfp.out1_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 827392 + }, + "/transformer_blocks.15/attn/Concat_6_3d.out22_15_bfp.out23_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 828928 + }, + "/transformer_blocks.15/attn/Concat_7_3d.out22_15_bfp.out23_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 830464 + }, + "/transformer_blocks.15/attn/Concat_8_3d.out22_15_bfp.out23_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 832000 + }, + "/transformer_blocks.15/attn/Reshape_6_output_0.out22_15_bfp.out27_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 833536 + }, + "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 835072 + }, + "/transformer_blocks.15/Add_output_0.out10_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 836608 + }, + "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 839680 + }, + "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 841216 + }, + "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 842752 + }, + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 848896 + }, + "/transformer_blocks.15/Add_3_output_0.out10_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 850432 + }, + "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 853504 + }, + "/transformer_blocks.15/Add_4_output_0.out10_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 855040 + }, + "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 858112 + }, + "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 859648 + }, + "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 861184 + }, + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 867328 + }, + "/transformer_blocks.15/Add_7_output_0.out10_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 868864 + }, + "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_64_bfp.out15_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 871936 + }, + "/Add_15_output_0.out_35_1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 873472 + }, + "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_65_bfp.out15_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 876544 + }, + "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_65_bfp.out1_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 878080 + }, + "/transformer_blocks.16/norm1_context/Add_2_output_0.out0_0_64_bfp.out1_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 879616 + }, + "/transformer_blocks.16/attn/Concat_6_3d.out22_16_bfp.out23_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 881152 + }, + "/transformer_blocks.16/attn/Concat_7_3d.out22_16_bfp.out23_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 882688 + }, + "/transformer_blocks.16/attn/Concat_8_3d.out22_16_bfp.out23_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 884224 + }, + "/transformer_blocks.16/attn/Reshape_6_output_0.out22_16_bfp.out27_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 885760 + }, + "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 887296 + }, + "/transformer_blocks.16/Add_output_0.out10_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 888832 + }, + "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 891904 + }, + "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 893440 + }, + "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 894976 + }, + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 901120 + }, + "/transformer_blocks.16/Add_3_output_0.out10_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 902656 + }, + "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 905728 + }, + "/transformer_blocks.16/Add_4_output_0.out10_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 907264 + }, + "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 910336 + }, + "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 911872 + }, + "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 913408 + }, + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 919552 + }, + "/transformer_blocks.16/Add_7_output_0.out10_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 921088 + }, + "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_68_bfp.out15_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 924160 + }, + "/Add_16_output_0.out_35_1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 925696 + }, + "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_69_bfp.out15_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 928768 + }, + "/transformer_blocks.17/norm1/Add_2_output_0.out0_0_69_bfp.out1_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 930304 + }, + "/transformer_blocks.17/norm1_context/Add_2_output_0.out0_0_68_bfp.out1_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 931840 + }, + "/transformer_blocks.17/attn/Concat_6_3d.out22_17_bfp.out23_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 933376 + }, + "/transformer_blocks.17/attn/Concat_7_3d.out22_17_bfp.out23_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 934912 + }, + "/transformer_blocks.17/attn/Concat_8_3d.out22_17_bfp.out23_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 936448 + }, + "/transformer_blocks.17/attn/Reshape_6_output_0.out22_17_bfp.out27_0_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 937984 + }, + "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 939520 + }, + "/transformer_blocks.17/Add_output_0.out10_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 941056 + }, + "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 944128 + }, + "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 945664 + }, + "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 947200 + }, + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 953344 + }, + "/transformer_blocks.17/Add_3_output_0.out10_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 954880 + }, + "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 957952 + }, + "/transformer_blocks.17/Add_4_output_0.out10_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 959488 + }, + "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 962560 + }, + "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 964096 + }, + "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 965632 + }, + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 971776 + }, + "/transformer_blocks.17/Add_7_output_0.out10_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 973312 + }, + "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_72_bfp.out15_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 976384 + }, + "/Add_17_output_0.out_35_1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 977920 + }, + "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_73_bfp.out15_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 980992 + }, + "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_73_bfp.out1_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 982528 + }, + "/transformer_blocks.18/norm1_context/Add_2_output_0.out0_0_72_bfp.out1_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 984064 + }, + "/transformer_blocks.18/attn/Concat_6_3d.out22_18_bfp.out23_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 985600 + }, + "/transformer_blocks.18/attn/Concat_7_3d.out22_18_bfp.out23_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 987136 + }, + "/transformer_blocks.18/attn/Concat_8_3d.out22_18_bfp.out23_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 988672 + }, + "/transformer_blocks.18/attn/Reshape_6_output_0.out22_18_bfp.out27_0_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 990208 + }, + "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 991744 + }, + "/transformer_blocks.18/Add_output_0.out10_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 993280 + }, + "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 996352 + }, + "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 997888 + }, + "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 999424 + }, + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1005568 + }, + "/transformer_blocks.18/Add_3_output_0.out10_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1007104 + }, + "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1010176 + }, + "/transformer_blocks.18/Add_4_output_0.out10_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1011712 + }, + "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1014784 + }, + "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1016320 + }, + "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1017856 + }, + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1024000 + }, + "/transformer_blocks.18/Add_7_output_0.out10_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1025536 + }, + "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_76_bfp.out15_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1028608 + }, + "/Add_18_output_0.out_35_1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1030144 + }, + "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_77_bfp.out15_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1033216 + }, + "/transformer_blocks.19/norm1/Add_2_output_0.out0_0_77_bfp.out1_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1034752 + }, + "/transformer_blocks.19/norm1_context/Add_2_output_0.out0_0_76_bfp.out1_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1036288 + }, + "/transformer_blocks.19/attn/Concat_6_3d.out22_19_bfp.out23_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1037824 + }, + "/transformer_blocks.19/attn/Concat_7_3d.out22_19_bfp.out23_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1039360 + }, + "/transformer_blocks.19/attn/Concat_8_3d.out22_19_bfp.out23_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1040896 + }, + "/transformer_blocks.19/attn/Reshape_6_output_0.out22_19_bfp.out27_0_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 1042432 + }, + "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1043968 + }, + "/transformer_blocks.19/Add_output_0.out10_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1045504 + }, + "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1048576 + }, + "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1050112 + }, + "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1051648 + }, + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1057792 + }, + "/transformer_blocks.19/Add_3_output_0.out10_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1059328 + }, + "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1062400 + }, + "/transformer_blocks.19/Add_4_output_0.out10_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1063936 + }, + "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1067008 + }, + "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1068544 + }, + "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1070080 + }, + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1076224 + }, + "/transformer_blocks.19/Add_7_output_0.out10_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1077760 + }, + "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_80_bfp.out15_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1080832 + }, + "/Add_19_output_0.out_35_1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1082368 + }, + "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_81_bfp.out15_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1085440 + }, + "/transformer_blocks.20/norm1/Add_2_output_0.out0_0_81_bfp.out1_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1086976 + }, + "/transformer_blocks.20/norm1_context/Add_2_output_0.out0_0_80_bfp.out1_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1088512 + }, + "/transformer_blocks.20/attn/Concat_6_3d.out22_20_bfp.out23_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1090048 + }, + "/transformer_blocks.20/attn/Concat_7_3d.out22_20_bfp.out23_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1091584 + }, + "/transformer_blocks.20/attn/Concat_8_3d.out22_20_bfp.out23_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1093120 + }, + "/transformer_blocks.20/attn/Reshape_6_output_0.out22_20_bfp.out27_0_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 1094656 + }, + "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1096192 + }, + "/transformer_blocks.20/Add_output_0.out10_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1097728 + }, + "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1100800 + }, + "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1102336 + }, + "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1103872 + }, + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1110016 + }, + "/transformer_blocks.20/Add_3_output_0.out10_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1111552 + }, + "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1114624 + }, + "/transformer_blocks.20/Add_4_output_0.out10_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1116160 + }, + "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1119232 + }, + "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1120768 + }, + "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1122304 + }, + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1128448 + }, + "/transformer_blocks.20/Add_7_output_0.out10_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1129984 + }, + "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_84_bfp.out15_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1133056 + }, + "/Add_20_output_0.out_35_1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1134592 + }, + "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_85_bfp.out15_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1137664 + }, + "/transformer_blocks.21/norm1/Add_2_output_0.out0_0_85_bfp.out1_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1139200 + }, + "/transformer_blocks.21/norm1_context/Add_2_output_0.out0_0_84_bfp.out1_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1140736 + }, + "/transformer_blocks.21/attn/Concat_6_3d.out22_21_bfp.out23_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1142272 + }, + "/transformer_blocks.21/attn/Concat_7_3d.out22_21_bfp.out23_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1143808 + }, + "/transformer_blocks.21/attn/Concat_8_3d.out22_21_bfp.out23_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1145344 + }, + "/transformer_blocks.21/attn/Reshape_6_output_0.out22_21_bfp.out27_0_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 1146880 + }, + "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1148416 + }, + "/transformer_blocks.21/Add_output_0.out10_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1149952 + }, + "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1153024 + }, + "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1154560 + }, + "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1156096 + }, + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1162240 + }, + "/transformer_blocks.21/Add_3_output_0.out10_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1163776 + }, + "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1166848 + }, + "/transformer_blocks.21/Add_4_output_0.out10_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1168384 + }, + "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1171456 + }, + "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1172992 + }, + "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1174528 + }, + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1180672 + }, + "/transformer_blocks.21/Add_7_output_0.out10_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1182208 + }, + "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_88_bfp.out15_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1185280 + }, + "/Add_21_output_0.out_35_1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1186816 + }, + "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_89_bfp.out15_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1189888 + }, + "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_89_bfp.out1_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1191424 + }, + "/transformer_blocks.22/norm1_context/Add_2_output_0.out0_0_88_bfp.out1_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1192960 + }, + "/transformer_blocks.22/attn/Concat_6_3d.out22_22_bfp.out23_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1194496 + }, + "/transformer_blocks.22/attn/Concat_7_3d.out22_22_bfp.out23_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1196032 + }, + "/transformer_blocks.22/attn/Concat_8_3d.out22_22_bfp.out23_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1197568 + }, + "/transformer_blocks.22/attn/Reshape_6_output_0.out22_22_bfp.out27_0_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 1199104 + }, + "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1200640 + }, + "/transformer_blocks.22/Add_output_0.out10_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1202176 + }, + "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1205248 + }, + "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1206784 + }, + "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1208320 + }, + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1214464 + }, + "/transformer_blocks.22/Add_3_output_0.out10_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1216000 + }, + "/transformer_blocks.22/attn/to_add_out/Add_output_0.out6_1_45_bfp.out7_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1219072 + }, + "/transformer_blocks.22/Add_4_output_0.out10_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1220608 + }, + "/transformer_blocks.22/norm2_context/LayerNormalization_output_0.out14_91_bfp.out15_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1223680 + }, + "/transformer_blocks.22/Add_6_output_0.out0_0_91_bfp.out1_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1225216 + }, + "/transformer_blocks.22/ff_context/net.0/Mul_5_output_0.out17_2_45_bfp.out25_94": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1226752 + }, + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1232896 + }, + "/transformer_blocks.22/Add_7_output_0.out10_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1234432 + }, + "/transformer_blocks.23/norm1_context/norm/LayerNormalization_output_0.out14_92_bfp.out15_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1237504 + }, + "/Add_22_output_0.out_35_1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1239040 + }, + "/transformer_blocks.23/norm1/norm/LayerNormalization_output_0.out14_93_bfp.out15_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1242112 + }, + "/transformer_blocks.23/norm1/Add_2_output_0.out0_0_93_bfp.out1_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1243648 + }, + "/transformer_blocks.23/norm1_context/Add_2_output_0.out0_0_92_bfp.out1_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length", + "False" + ], + "offset": 1245184 + }, + "/transformer_blocks.23/attn/Concat_6_3d.out22_23_bfp.out23_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1246720 + }, + "/transformer_blocks.23/attn/Concat_7_3d.out22_23_bfp.out23_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1248256 + }, + "/transformer_blocks.23/attn/Concat_8_3d.out22_23_bfp.out23_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 24, + 1, + 64 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "False", + "state_dim1+max_length", + "False" + ], + "offset": 1249792 + }, + "/transformer_blocks.23/attn/Reshape_6_output_0.out22_23_bfp.out27_0_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "max_length + state_dim1", + "False" + ], + "offset": 1251328 + }, + "/transformer_blocks.23/attn/to_out.0/Add_output_0.out6_1_46_bfp.out7_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1252864 + }, + "/transformer_blocks.23/Add_output_0.out10_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1254400 + }, + "/transformer_blocks.23/norm2/LayerNormalization_output_0.out14_94_bfp.out15_94": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1257472 + }, + "/transformer_blocks.23/Add_2_output_0.out0_0_94_bfp.out1_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1259008 + }, + "/transformer_blocks.23/ff/net.0/Mul_5_output_0.out17_2_46_bfp.out25_97": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 6144 + ], + "size_in_bytes": 6144, + "op_tensor_size": 6144, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1260544 + }, + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1266688 + }, + "/transformer_blocks.23/Add_3_output_0.out10_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1268224 + }, + "/norm_out/norm/LayerNormalization_output_0.out14_95_bfp.out15_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1271296 + }, + "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1272832 + }, + "/norm_out/Add_2_output_0.out0_0_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_size", + "state_dim1", + "False" + ], + "offset": 1274368 + }, + "existing_model.pos_embed.proj.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135168 + ], + "size_in_bytes": 135168, + "op_tensor_size": 135168, + "offset": 0, + "file_name": "cache/pos_embedprojConv_0.const", + "file_size": 135168 + }, + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_45_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 448512 + ], + "size_in_bytes": 448512, + "op_tensor_size": 448512, + "offset": 135168, + "file_name": "cache/pos_embedprojConv_1.const", + "file_size": 448512 + }, + "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 583680, + "file_name": "cache/pos_embedprojConv_2.const", + "file_size": 256 + }, + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_45_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 583936, + "file_name": "cache/pos_embedprojConv_3.const", + "file_size": 2691072 + }, + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_45_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 3588096 + ], + "size_in_bytes": 3588096, + "op_tensor_size": 3588096, + "offset": 3275008, + "file_name": "cache/pos_embedprojConv_4.const", + "file_size": 3588096 + }, + "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 6863104, + "file_name": "cache/pos_embedprojConv_5.const", + "file_size": 256 + }, + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_45_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 6863360, + "file_name": "cache/pos_embedprojConv_6.const", + "file_size": 2691072 + }, + "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 9554432, + "file_name": "cache/pos_embedprojConv_7.const", + "file_size": 256 + }, + "encoder_hidden_states.out17_4_0_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 9554688, + "file_name": "cache/pos_embedprojConv_8.const", + "file_size": 128 + }, + "onnx::MatMul_8733_45_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7274496 + ], + "size_in_bytes": 7274496, + "op_tensor_size": 7274496, + "offset": 9554816, + "file_name": "cache/pos_embedprojConv_9.const", + "file_size": 7274496 + }, + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 16829312, + "file_name": "cache/pos_embedprojConv_10.const", + "file_size": 128 + }, + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 16829440, + "file_name": "cache/pos_embedprojConv_11.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 16832512, + "file_name": "cache/pos_embedprojConv_12.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 16835584, + "file_name": "cache/pos_embedprojConv_13.const", + "file_size": 5382144 + }, + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 22217728, + "file_name": "cache/pos_embedprojConv_14.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 22220800, + "file_name": "cache/pos_embedprojConv_15.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 22223872, + "file_name": "cache/pos_embedprojConv_16.const", + "file_size": 5382144 + }, + "onnx::MatMul_8749_onnx::MatMul_8734": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 27606016, + "file_name": "cache/pos_embedprojConv_17.const", + "file_size": 5382144 + }, + "onnx::MatMul_8750_onnx::MatMul_8735": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 32988160, + "file_name": "cache/pos_embedprojConv_18.const", + "file_size": 5382144 + }, + "onnx::MatMul_8751_onnx::MatMul_8736": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 38370304, + "file_name": "cache/pos_embedprojConv_19.const", + "file_size": 5382144 + }, + "onnx::MatMul_8759_45_5_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 43752448, + "file_name": "cache/pos_embedprojConv_20.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 46443520, + "file_name": "cache/pos_embedprojConv_21.const", + "file_size": 2691072 + }, + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 49134592, + "file_name": "cache/pos_embedprojConv_22.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 49137664, + "file_name": "cache/pos_embedprojConv_23.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 49140736, + "file_name": "cache/pos_embedprojConv_24.const", + "file_size": 5382144 + }, + "onnx::MatMul_8763_45_1_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 54522880, + "file_name": "cache/pos_embedprojConv_25.const", + "file_size": 10764288 + }, + "onnx::MatMul_8764_45_1_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 65287168, + "file_name": "cache/pos_embedprojConv_26.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 76051456, + "file_name": "cache/pos_embedprojConv_27.const", + "file_size": 2691072 + }, + "onnx::MatMul_8760_45_5_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 78742528, + "file_name": "cache/pos_embedprojConv_28.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 81433600, + "file_name": "cache/pos_embedprojConv_29.const", + "file_size": 2691072 + }, + "/transformer_blocks.0/norm2/Constant_output_0_14_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 84124672, + "file_name": "cache/pos_embedprojConv_30.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 84127744, + "file_name": "cache/pos_embedprojConv_31.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 84130816, + "file_name": "cache/pos_embedprojConv_32.const", + "file_size": 5382144 + }, + "onnx::MatMul_8761_45_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 89512960, + "file_name": "cache/pos_embedprojConv_33.const", + "file_size": 10764288 + }, + "onnx::MatMul_8762_45_1_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 100277248, + "file_name": "cache/pos_embedprojConv_34.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 111041536, + "file_name": "cache/pos_embedprojConv_35.const", + "file_size": 2691072 + }, + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 113732608, + "file_name": "cache/pos_embedprojConv_36.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 113735680, + "file_name": "cache/pos_embedprojConv_37.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 113738752, + "file_name": "cache/pos_embedprojConv_38.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 113741824, + "file_name": "cache/pos_embedprojConv_39.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 113744896, + "file_name": "cache/pos_embedprojConv_40.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 119127040, + "file_name": "cache/pos_embedprojConv_41.const", + "file_size": 5382144 + }, + "onnx::MatMul_8780_onnx::MatMul_8765": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 124509184, + "file_name": "cache/pos_embedprojConv_42.const", + "file_size": 5382144 + }, + "onnx::MatMul_8781_onnx::MatMul_8766": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 129891328, + "file_name": "cache/pos_embedprojConv_43.const", + "file_size": 5382144 + }, + "onnx::MatMul_8782_onnx::MatMul_8767": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 135273472, + "file_name": "cache/pos_embedprojConv_44.const", + "file_size": 5382144 + }, + "onnx::MatMul_8791_45_5_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 140655616, + "file_name": "cache/pos_embedprojConv_45.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 143346688, + "file_name": "cache/pos_embedprojConv_46.const", + "file_size": 2691072 + }, + "/transformer_blocks.1/norm2/Constant_output_0_14_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 146037760, + "file_name": "cache/pos_embedprojConv_47.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 146040832, + "file_name": "cache/pos_embedprojConv_48.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 146043904, + "file_name": "cache/pos_embedprojConv_49.const", + "file_size": 5382144 + }, + "onnx::MatMul_8792_45_1_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 151426048, + "file_name": "cache/pos_embedprojConv_50.const", + "file_size": 10764288 + }, + "onnx::MatMul_8793_45_1_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 162190336, + "file_name": "cache/pos_embedprojConv_51.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 172954624, + "file_name": "cache/pos_embedprojConv_52.const", + "file_size": 2691072 + }, + "onnx::MatMul_8790_45_5_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 175645696, + "file_name": "cache/pos_embedprojConv_53.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 178336768, + "file_name": "cache/pos_embedprojConv_54.const", + "file_size": 2691072 + }, + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 181027840, + "file_name": "cache/pos_embedprojConv_55.const", + "file_size": 3072 + }, + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 181030912, + "file_name": "cache/pos_embedprojConv_56.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 181033984, + "file_name": "cache/pos_embedprojConv_57.const", + "file_size": 5382144 + }, + "onnx::MatMul_8794_45_1_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 186416128, + "file_name": "cache/pos_embedprojConv_58.const", + "file_size": 10764288 + }, + "onnx::MatMul_8795_45_1_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 197180416, + "file_name": "cache/pos_embedprojConv_59.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 207944704, + "file_name": "cache/pos_embedprojConv_60.const", + "file_size": 2691072 + }, + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 210635776, + "file_name": "cache/pos_embedprojConv_61.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 210638848, + "file_name": "cache/pos_embedprojConv_62.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 210641920, + "file_name": "cache/pos_embedprojConv_63.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 210644992, + "file_name": "cache/pos_embedprojConv_64.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 210648064, + "file_name": "cache/pos_embedprojConv_65.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 216030208, + "file_name": "cache/pos_embedprojConv_66.const", + "file_size": 5382144 + }, + "onnx::MatMul_8811_onnx::MatMul_8796": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 221412352, + "file_name": "cache/pos_embedprojConv_67.const", + "file_size": 5382144 + }, + "onnx::MatMul_8812_onnx::MatMul_8797": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 226794496, + "file_name": "cache/pos_embedprojConv_68.const", + "file_size": 5382144 + }, + "onnx::MatMul_8813_onnx::MatMul_8798": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 232176640, + "file_name": "cache/pos_embedprojConv_69.const", + "file_size": 5382144 + }, + "onnx::MatMul_8821_45_5_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 237558784, + "file_name": "cache/pos_embedprojConv_70.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 240249856, + "file_name": "cache/pos_embedprojConv_71.const", + "file_size": 2691072 + }, + "onnx::MatMul_8822_45_5_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 242940928, + "file_name": "cache/pos_embedprojConv_72.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 245632000, + "file_name": "cache/pos_embedprojConv_73.const", + "file_size": 2691072 + }, + "/transformer_blocks.2/norm2/Constant_output_0_14_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 248323072, + "file_name": "cache/pos_embedprojConv_74.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 248326144, + "file_name": "cache/pos_embedprojConv_75.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 248329216, + "file_name": "cache/pos_embedprojConv_76.const", + "file_size": 5382144 + }, + "onnx::MatMul_8823_45_1_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 253711360, + "file_name": "cache/pos_embedprojConv_77.const", + "file_size": 10764288 + }, + "onnx::MatMul_8824_45_1_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 264475648, + "file_name": "cache/pos_embedprojConv_78.const", + "file_size": 10764288 + }, + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 275239936, + "file_name": "cache/pos_embedprojConv_79.const", + "file_size": 3072 + }, + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 275243008, + "file_name": "cache/pos_embedprojConv_80.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 275246080, + "file_name": "cache/pos_embedprojConv_81.const", + "file_size": 5382144 + }, + "onnx::MatMul_8825_45_1_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 280628224, + "file_name": "cache/pos_embedprojConv_82.const", + "file_size": 10764288 + }, + "onnx::MatMul_8826_45_1_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 291392512, + "file_name": "cache/pos_embedprojConv_83.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 302156800, + "file_name": "cache/pos_embedprojConv_84.const", + "file_size": 2691072 + }, + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 304847872, + "file_name": "cache/pos_embedprojConv_85.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 304850944, + "file_name": "cache/pos_embedprojConv_86.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 304854016, + "file_name": "cache/pos_embedprojConv_87.const", + "file_size": 2691072 + }, + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 307545088, + "file_name": "cache/pos_embedprojConv_88.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 307548160, + "file_name": "cache/pos_embedprojConv_89.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 307551232, + "file_name": "cache/pos_embedprojConv_90.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 312933376, + "file_name": "cache/pos_embedprojConv_91.const", + "file_size": 5382144 + }, + "onnx::MatMul_8842_onnx::MatMul_8827": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 318315520, + "file_name": "cache/pos_embedprojConv_92.const", + "file_size": 5382144 + }, + "onnx::MatMul_8843_onnx::MatMul_8828": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 323697664, + "file_name": "cache/pos_embedprojConv_93.const", + "file_size": 5382144 + }, + "onnx::MatMul_8844_onnx::MatMul_8829": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 329079808, + "file_name": "cache/pos_embedprojConv_94.const", + "file_size": 5382144 + }, + "onnx::MatMul_8852_45_5_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 334461952, + "file_name": "cache/pos_embedprojConv_95.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 337153024, + "file_name": "cache/pos_embedprojConv_96.const", + "file_size": 2691072 + }, + "onnx::MatMul_8853_45_5_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 339844096, + "file_name": "cache/pos_embedprojConv_97.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 342535168, + "file_name": "cache/pos_embedprojConv_98.const", + "file_size": 2691072 + }, + "/transformer_blocks.3/norm2/Constant_output_0_14_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 345226240, + "file_name": "cache/pos_embedprojConv_99.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 345229312, + "file_name": "cache/pos_embedprojConv_100.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 345232384, + "file_name": "cache/pos_embedprojConv_101.const", + "file_size": 5382144 + }, + "onnx::MatMul_8854_45_1_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 350614528, + "file_name": "cache/pos_embedprojConv_102.const", + "file_size": 10764288 + }, + "onnx::MatMul_8855_45_1_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 361378816, + "file_name": "cache/pos_embedprojConv_103.const", + "file_size": 10764288 + }, + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 372143104, + "file_name": "cache/pos_embedprojConv_104.const", + "file_size": 3072 + }, + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 372146176, + "file_name": "cache/pos_embedprojConv_105.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 372149248, + "file_name": "cache/pos_embedprojConv_106.const", + "file_size": 5382144 + }, + "onnx::MatMul_8856_45_1_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 377531392, + "file_name": "cache/pos_embedprojConv_107.const", + "file_size": 10764288 + }, + "onnx::MatMul_8857_45_1_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 388295680, + "file_name": "cache/pos_embedprojConv_108.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 399059968, + "file_name": "cache/pos_embedprojConv_109.const", + "file_size": 2691072 + }, + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 401751040, + "file_name": "cache/pos_embedprojConv_110.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 401754112, + "file_name": "cache/pos_embedprojConv_111.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 401757184, + "file_name": "cache/pos_embedprojConv_112.const", + "file_size": 2691072 + }, + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 404448256, + "file_name": "cache/pos_embedprojConv_113.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 404451328, + "file_name": "cache/pos_embedprojConv_114.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 404454400, + "file_name": "cache/pos_embedprojConv_115.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 409836544, + "file_name": "cache/pos_embedprojConv_116.const", + "file_size": 5382144 + }, + "onnx::MatMul_8873_onnx::MatMul_8858": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 415218688, + "file_name": "cache/pos_embedprojConv_117.const", + "file_size": 5382144 + }, + "onnx::MatMul_8874_onnx::MatMul_8859": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 420600832, + "file_name": "cache/pos_embedprojConv_118.const", + "file_size": 5382144 + }, + "onnx::MatMul_8875_onnx::MatMul_8860": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 425982976, + "file_name": "cache/pos_embedprojConv_119.const", + "file_size": 5382144 + }, + "onnx::MatMul_8883_45_5_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 431365120, + "file_name": "cache/pos_embedprojConv_120.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 434056192, + "file_name": "cache/pos_embedprojConv_121.const", + "file_size": 2691072 + }, + "onnx::MatMul_8884_45_5_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 436747264, + "file_name": "cache/pos_embedprojConv_122.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 439438336, + "file_name": "cache/pos_embedprojConv_123.const", + "file_size": 2691072 + }, + "/transformer_blocks.4/norm2/Constant_output_0_14_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 442129408, + "file_name": "cache/pos_embedprojConv_124.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 442132480, + "file_name": "cache/pos_embedprojConv_125.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 442135552, + "file_name": "cache/pos_embedprojConv_126.const", + "file_size": 5382144 + }, + "onnx::MatMul_8885_45_1_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 447517696, + "file_name": "cache/pos_embedprojConv_127.const", + "file_size": 10764288 + }, + "onnx::MatMul_8886_45_1_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 458281984, + "file_name": "cache/pos_embedprojConv_128.const", + "file_size": 10764288 + }, + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 469046272, + "file_name": "cache/pos_embedprojConv_129.const", + "file_size": 3072 + }, + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 469049344, + "file_name": "cache/pos_embedprojConv_130.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 469052416, + "file_name": "cache/pos_embedprojConv_131.const", + "file_size": 5382144 + }, + "onnx::MatMul_8887_45_1_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 474434560, + "file_name": "cache/pos_embedprojConv_132.const", + "file_size": 10764288 + }, + "onnx::MatMul_8888_45_1_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 485198848, + "file_name": "cache/pos_embedprojConv_133.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 495963136, + "file_name": "cache/pos_embedprojConv_134.const", + "file_size": 2691072 + }, + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 498654208, + "file_name": "cache/pos_embedprojConv_135.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 498657280, + "file_name": "cache/pos_embedprojConv_136.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 498660352, + "file_name": "cache/pos_embedprojConv_137.const", + "file_size": 2691072 + }, + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 501351424, + "file_name": "cache/pos_embedprojConv_138.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 501354496, + "file_name": "cache/pos_embedprojConv_139.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 501357568, + "file_name": "cache/pos_embedprojConv_140.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 506739712, + "file_name": "cache/pos_embedprojConv_141.const", + "file_size": 5382144 + }, + "onnx::MatMul_8904_onnx::MatMul_8889": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 512121856, + "file_name": "cache/pos_embedprojConv_142.const", + "file_size": 5382144 + }, + "onnx::MatMul_8905_onnx::MatMul_8890": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 517504000, + "file_name": "cache/pos_embedprojConv_143.const", + "file_size": 5382144 + }, + "onnx::MatMul_8906_onnx::MatMul_8891": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 522886144, + "file_name": "cache/pos_embedprojConv_144.const", + "file_size": 5382144 + }, + "onnx::MatMul_8914_45_5_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 528268288, + "file_name": "cache/pos_embedprojConv_145.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 530959360, + "file_name": "cache/pos_embedprojConv_146.const", + "file_size": 2691072 + }, + "onnx::MatMul_8915_45_5_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 533650432, + "file_name": "cache/pos_embedprojConv_147.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 536341504, + "file_name": "cache/pos_embedprojConv_148.const", + "file_size": 2691072 + }, + "/transformer_blocks.5/norm2/Constant_output_0_14_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 539032576, + "file_name": "cache/pos_embedprojConv_149.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 539035648, + "file_name": "cache/pos_embedprojConv_150.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 539038720, + "file_name": "cache/pos_embedprojConv_151.const", + "file_size": 5382144 + }, + "onnx::MatMul_8916_45_1_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 544420864, + "file_name": "cache/pos_embedprojConv_152.const", + "file_size": 10764288 + }, + "onnx::MatMul_8917_45_1_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 555185152, + "file_name": "cache/pos_embedprojConv_153.const", + "file_size": 10764288 + }, + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 565949440, + "file_name": "cache/pos_embedprojConv_154.const", + "file_size": 3072 + }, + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 565952512, + "file_name": "cache/pos_embedprojConv_155.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_3_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 565955584, + "file_name": "cache/pos_embedprojConv_156.const", + "file_size": 5382144 + }, + "onnx::MatMul_8918_45_1_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 571337728, + "file_name": "cache/pos_embedprojConv_157.const", + "file_size": 10764288 + }, + "onnx::MatMul_8919_45_1_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 582102016, + "file_name": "cache/pos_embedprojConv_158.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 592866304, + "file_name": "cache/pos_embedprojConv_159.const", + "file_size": 2691072 + }, + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 595557376, + "file_name": "cache/pos_embedprojConv_160.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 595560448, + "file_name": "cache/pos_embedprojConv_161.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 595563520, + "file_name": "cache/pos_embedprojConv_162.const", + "file_size": 2691072 + }, + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 598254592, + "file_name": "cache/pos_embedprojConv_163.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 598257664, + "file_name": "cache/pos_embedprojConv_164.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_0_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 598260736, + "file_name": "cache/pos_embedprojConv_165.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_0_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 603642880, + "file_name": "cache/pos_embedprojConv_166.const", + "file_size": 5382144 + }, + "onnx::MatMul_8935_onnx::MatMul_8920": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 609025024, + "file_name": "cache/pos_embedprojConv_167.const", + "file_size": 5382144 + }, + "onnx::MatMul_8936_onnx::MatMul_8921": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 614407168, + "file_name": "cache/pos_embedprojConv_168.const", + "file_size": 5382144 + }, + "onnx::MatMul_8937_onnx::MatMul_8922": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 619789312, + "file_name": "cache/pos_embedprojConv_169.const", + "file_size": 5382144 + }, + "onnx::MatMul_8945_45_5_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 625171456, + "file_name": "cache/pos_embedprojConv_170.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 627862528, + "file_name": "cache/pos_embedprojConv_171.const", + "file_size": 2691072 + }, + "onnx::MatMul_8946_45_5_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 630553600, + "file_name": "cache/pos_embedprojConv_172.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 633244672, + "file_name": "cache/pos_embedprojConv_173.const", + "file_size": 2691072 + }, + "/transformer_blocks.6/norm2/Constant_output_0_14_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 635935744, + "file_name": "cache/pos_embedprojConv_174.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 635938816, + "file_name": "cache/pos_embedprojConv_175.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_3_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 635941888, + "file_name": "cache/pos_embedprojConv_176.const", + "file_size": 5382144 + }, + "onnx::MatMul_8947_45_1_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 641324032, + "file_name": "cache/pos_embedprojConv_177.const", + "file_size": 10764288 + }, + "onnx::MatMul_8948_45_1_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 652088320, + "file_name": "cache/pos_embedprojConv_178.const", + "file_size": 10764288 + }, + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 662852608, + "file_name": "cache/pos_embedprojConv_179.const", + "file_size": 3072 + }, + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 662855680, + "file_name": "cache/pos_embedprojConv_180.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_3_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 662858752, + "file_name": "cache/pos_embedprojConv_181.const", + "file_size": 5382144 + }, + "onnx::MatMul_8949_45_1_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 668240896, + "file_name": "cache/pos_embedprojConv_182.const", + "file_size": 10764288 + }, + "onnx::MatMul_8950_45_1_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 679005184, + "file_name": "cache/pos_embedprojConv_183.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 689769472, + "file_name": "cache/pos_embedprojConv_184.const", + "file_size": 2691072 + }, + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 692460544, + "file_name": "cache/pos_embedprojConv_185.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 692463616, + "file_name": "cache/pos_embedprojConv_186.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 692466688, + "file_name": "cache/pos_embedprojConv_187.const", + "file_size": 2691072 + }, + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 695157760, + "file_name": "cache/pos_embedprojConv_188.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 695160832, + "file_name": "cache/pos_embedprojConv_189.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_0_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 695163904, + "file_name": "cache/pos_embedprojConv_190.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_0_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 700546048, + "file_name": "cache/pos_embedprojConv_191.const", + "file_size": 5382144 + }, + "onnx::MatMul_8966_onnx::MatMul_8951": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 705928192, + "file_name": "cache/pos_embedprojConv_192.const", + "file_size": 5382144 + }, + "onnx::MatMul_8967_onnx::MatMul_8952": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 711310336, + "file_name": "cache/pos_embedprojConv_193.const", + "file_size": 5382144 + }, + "onnx::MatMul_8968_onnx::MatMul_8953": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 716692480, + "file_name": "cache/pos_embedprojConv_194.const", + "file_size": 5382144 + }, + "onnx::MatMul_8976_45_5_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 722074624, + "file_name": "cache/pos_embedprojConv_195.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 724765696, + "file_name": "cache/pos_embedprojConv_196.const", + "file_size": 2691072 + }, + "onnx::MatMul_8977_45_5_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 727456768, + "file_name": "cache/pos_embedprojConv_197.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 730147840, + "file_name": "cache/pos_embedprojConv_198.const", + "file_size": 2691072 + }, + "/transformer_blocks.7/norm2/Constant_output_0_14_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 732838912, + "file_name": "cache/pos_embedprojConv_199.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 732841984, + "file_name": "cache/pos_embedprojConv_200.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_3_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 732845056, + "file_name": "cache/pos_embedprojConv_201.const", + "file_size": 5382144 + }, + "onnx::MatMul_8978_45_1_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 738227200, + "file_name": "cache/pos_embedprojConv_202.const", + "file_size": 10764288 + }, + "onnx::MatMul_8979_45_1_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 748991488, + "file_name": "cache/pos_embedprojConv_203.const", + "file_size": 10764288 + }, + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 759755776, + "file_name": "cache/pos_embedprojConv_204.const", + "file_size": 3072 + }, + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 759758848, + "file_name": "cache/pos_embedprojConv_205.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_3_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 759761920, + "file_name": "cache/pos_embedprojConv_206.const", + "file_size": 5382144 + }, + "onnx::MatMul_8980_45_1_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 765144064, + "file_name": "cache/pos_embedprojConv_207.const", + "file_size": 10764288 + }, + "onnx::MatMul_8981_45_1_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 775908352, + "file_name": "cache/pos_embedprojConv_208.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 786672640, + "file_name": "cache/pos_embedprojConv_209.const", + "file_size": 2691072 + }, + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 789363712, + "file_name": "cache/pos_embedprojConv_210.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 789366784, + "file_name": "cache/pos_embedprojConv_211.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 789369856, + "file_name": "cache/pos_embedprojConv_212.const", + "file_size": 2691072 + }, + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 792060928, + "file_name": "cache/pos_embedprojConv_213.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 792064000, + "file_name": "cache/pos_embedprojConv_214.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_0_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 792067072, + "file_name": "cache/pos_embedprojConv_215.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_0_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 797449216, + "file_name": "cache/pos_embedprojConv_216.const", + "file_size": 5382144 + }, + "onnx::MatMul_8997_onnx::MatMul_8982": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 802831360, + "file_name": "cache/pos_embedprojConv_217.const", + "file_size": 5382144 + }, + "onnx::MatMul_8998_onnx::MatMul_8983": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 808213504, + "file_name": "cache/pos_embedprojConv_218.const", + "file_size": 5382144 + }, + "onnx::MatMul_8999_onnx::MatMul_8984": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 813595648, + "file_name": "cache/pos_embedprojConv_219.const", + "file_size": 5382144 + }, + "onnx::MatMul_9007_45_5_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 818977792, + "file_name": "cache/pos_embedprojConv_220.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 821668864, + "file_name": "cache/pos_embedprojConv_221.const", + "file_size": 2691072 + }, + "onnx::MatMul_9008_45_5_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 824359936, + "file_name": "cache/pos_embedprojConv_222.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 827051008, + "file_name": "cache/pos_embedprojConv_223.const", + "file_size": 2691072 + }, + "/transformer_blocks.8/norm2/Constant_output_0_14_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 829742080, + "file_name": "cache/pos_embedprojConv_224.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 829745152, + "file_name": "cache/pos_embedprojConv_225.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_3_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 829748224, + "file_name": "cache/pos_embedprojConv_226.const", + "file_size": 5382144 + }, + "onnx::MatMul_9009_45_1_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 835130368, + "file_name": "cache/pos_embedprojConv_227.const", + "file_size": 10764288 + }, + "onnx::MatMul_9010_45_1_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 845894656, + "file_name": "cache/pos_embedprojConv_228.const", + "file_size": 10764288 + }, + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 856658944, + "file_name": "cache/pos_embedprojConv_229.const", + "file_size": 3072 + }, + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 856662016, + "file_name": "cache/pos_embedprojConv_230.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_3_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 856665088, + "file_name": "cache/pos_embedprojConv_231.const", + "file_size": 5382144 + }, + "onnx::MatMul_9011_45_1_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 862047232, + "file_name": "cache/pos_embedprojConv_232.const", + "file_size": 10764288 + }, + "onnx::MatMul_9012_45_1_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 872811520, + "file_name": "cache/pos_embedprojConv_233.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 883575808, + "file_name": "cache/pos_embedprojConv_234.const", + "file_size": 2691072 + }, + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 886266880, + "file_name": "cache/pos_embedprojConv_235.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 886269952, + "file_name": "cache/pos_embedprojConv_236.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 886273024, + "file_name": "cache/pos_embedprojConv_237.const", + "file_size": 2691072 + }, + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 888964096, + "file_name": "cache/pos_embedprojConv_238.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 888967168, + "file_name": "cache/pos_embedprojConv_239.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_0_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 888970240, + "file_name": "cache/pos_embedprojConv_240.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_0_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 894352384, + "file_name": "cache/pos_embedprojConv_241.const", + "file_size": 5382144 + }, + "onnx::MatMul_9028_onnx::MatMul_9013": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 899734528, + "file_name": "cache/pos_embedprojConv_242.const", + "file_size": 5382144 + }, + "onnx::MatMul_9029_onnx::MatMul_9014": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 905116672, + "file_name": "cache/pos_embedprojConv_243.const", + "file_size": 5382144 + }, + "onnx::MatMul_9030_onnx::MatMul_9015": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 910498816, + "file_name": "cache/pos_embedprojConv_244.const", + "file_size": 5382144 + }, + "onnx::MatMul_9038_45_5_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 915880960, + "file_name": "cache/pos_embedprojConv_245.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 918572032, + "file_name": "cache/pos_embedprojConv_246.const", + "file_size": 2691072 + }, + "onnx::MatMul_9039_45_5_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 921263104, + "file_name": "cache/pos_embedprojConv_247.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 923954176, + "file_name": "cache/pos_embedprojConv_248.const", + "file_size": 2691072 + }, + "/transformer_blocks.9/norm2/Constant_output_0_14_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 926645248, + "file_name": "cache/pos_embedprojConv_249.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 926648320, + "file_name": "cache/pos_embedprojConv_250.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_3_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 926651392, + "file_name": "cache/pos_embedprojConv_251.const", + "file_size": 5382144 + }, + "onnx::MatMul_9040_45_1_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 932033536, + "file_name": "cache/pos_embedprojConv_252.const", + "file_size": 10764288 + }, + "onnx::MatMul_9041_45_1_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 942797824, + "file_name": "cache/pos_embedprojConv_253.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 953562112, + "file_name": "cache/pos_embedprojConv_254.const", + "file_size": 2691072 + }, + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 956253184, + "file_name": "cache/pos_embedprojConv_255.const", + "file_size": 3072 + }, + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 956256256, + "file_name": "cache/pos_embedprojConv_256.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_3_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 956259328, + "file_name": "cache/pos_embedprojConv_257.const", + "file_size": 5382144 + }, + "onnx::MatMul_9042_45_1_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 961641472, + "file_name": "cache/pos_embedprojConv_258.const", + "file_size": 10764288 + }, + "onnx::MatMul_9043_45_1_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 972405760, + "file_name": "cache/pos_embedprojConv_259.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 983170048, + "file_name": "cache/pos_embedprojConv_260.const", + "file_size": 2691072 + }, + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 985861120, + "file_name": "cache/pos_embedprojConv_261.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 985864192, + "file_name": "cache/pos_embedprojConv_262.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 985867264, + "file_name": "cache/pos_embedprojConv_263.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 985870336, + "file_name": "cache/pos_embedprojConv_264.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_0_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 985873408, + "file_name": "cache/pos_embedprojConv_265.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_0_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 991255552, + "file_name": "cache/pos_embedprojConv_266.const", + "file_size": 5382144 + }, + "onnx::MatMul_9059_onnx::MatMul_9044": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 996637696, + "file_name": "cache/pos_embedprojConv_267.const", + "file_size": 5382144 + }, + "onnx::MatMul_9060_onnx::MatMul_9045": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1002019840, + "file_name": "cache/pos_embedprojConv_268.const", + "file_size": 5382144 + }, + "onnx::MatMul_9061_onnx::MatMul_9046": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1007401984, + "file_name": "cache/pos_embedprojConv_269.const", + "file_size": 5382144 + }, + "onnx::MatMul_9070_45_5_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1012784128, + "file_name": "cache/pos_embedprojConv_270.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1015475200, + "file_name": "cache/pos_embedprojConv_271.const", + "file_size": 2691072 + }, + "/transformer_blocks.10/norm2/Constant_output_0_14_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1018166272, + "file_name": "cache/pos_embedprojConv_272.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1018169344, + "file_name": "cache/pos_embedprojConv_273.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_3_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1018172416, + "file_name": "cache/pos_embedprojConv_274.const", + "file_size": 5382144 + }, + "onnx::MatMul_9071_45_1_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1023554560, + "file_name": "cache/pos_embedprojConv_275.const", + "file_size": 10764288 + }, + "onnx::MatMul_9072_45_1_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1034318848, + "file_name": "cache/pos_embedprojConv_276.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1045083136, + "file_name": "cache/pos_embedprojConv_277.const", + "file_size": 2691072 + }, + "onnx::MatMul_9069_45_5_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1047774208, + "file_name": "cache/pos_embedprojConv_278.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1050465280, + "file_name": "cache/pos_embedprojConv_279.const", + "file_size": 2691072 + }, + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1053156352, + "file_name": "cache/pos_embedprojConv_280.const", + "file_size": 3072 + }, + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1053159424, + "file_name": "cache/pos_embedprojConv_281.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_3_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1053162496, + "file_name": "cache/pos_embedprojConv_282.const", + "file_size": 5382144 + }, + "onnx::MatMul_9073_45_1_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1058544640, + "file_name": "cache/pos_embedprojConv_283.const", + "file_size": 10764288 + }, + "onnx::MatMul_9074_45_1_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1069308928, + "file_name": "cache/pos_embedprojConv_284.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1080073216, + "file_name": "cache/pos_embedprojConv_285.const", + "file_size": 2691072 + }, + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1082764288, + "file_name": "cache/pos_embedprojConv_286.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1082767360, + "file_name": "cache/pos_embedprojConv_287.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1082770432, + "file_name": "cache/pos_embedprojConv_288.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1082773504, + "file_name": "cache/pos_embedprojConv_289.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_0_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1082776576, + "file_name": "cache/pos_embedprojConv_290.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_0_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1088158720, + "file_name": "cache/pos_embedprojConv_291.const", + "file_size": 5382144 + }, + "onnx::MatMul_9090_onnx::MatMul_9075": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1093540864, + "file_name": "cache/pos_embedprojConv_292.const", + "file_size": 5382144 + }, + "onnx::MatMul_9091_onnx::MatMul_9076": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1098923008, + "file_name": "cache/pos_embedprojConv_293.const", + "file_size": 5382144 + }, + "onnx::MatMul_9092_onnx::MatMul_9077": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1104305152, + "file_name": "cache/pos_embedprojConv_294.const", + "file_size": 5382144 + }, + "onnx::MatMul_9101_45_5_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1109687296, + "file_name": "cache/pos_embedprojConv_295.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1112378368, + "file_name": "cache/pos_embedprojConv_296.const", + "file_size": 2691072 + }, + "/transformer_blocks.11/norm2/Constant_output_0_14_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1115069440, + "file_name": "cache/pos_embedprojConv_297.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1115072512, + "file_name": "cache/pos_embedprojConv_298.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_3_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1115075584, + "file_name": "cache/pos_embedprojConv_299.const", + "file_size": 5382144 + }, + "onnx::MatMul_9102_45_1_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1120457728, + "file_name": "cache/pos_embedprojConv_300.const", + "file_size": 10764288 + }, + "onnx::MatMul_9103_45_1_47": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1131222016, + "file_name": "cache/pos_embedprojConv_301.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1141986304, + "file_name": "cache/pos_embedprojConv_302.const", + "file_size": 2691072 + }, + "onnx::MatMul_9100_45_5_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1144677376, + "file_name": "cache/pos_embedprojConv_303.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1147368448, + "file_name": "cache/pos_embedprojConv_304.const", + "file_size": 2691072 + }, + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1150059520, + "file_name": "cache/pos_embedprojConv_305.const", + "file_size": 3072 + }, + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1150062592, + "file_name": "cache/pos_embedprojConv_306.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_3_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1150065664, + "file_name": "cache/pos_embedprojConv_307.const", + "file_size": 5382144 + }, + "onnx::MatMul_9104_45_1_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1155447808, + "file_name": "cache/pos_embedprojConv_308.const", + "file_size": 10764288 + }, + "onnx::MatMul_9105_45_1_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1166212096, + "file_name": "cache/pos_embedprojConv_309.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1176976384, + "file_name": "cache/pos_embedprojConv_310.const", + "file_size": 2691072 + }, + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1179667456, + "file_name": "cache/pos_embedprojConv_311.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1179670528, + "file_name": "cache/pos_embedprojConv_312.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1179673600, + "file_name": "cache/pos_embedprojConv_313.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1179676672, + "file_name": "cache/pos_embedprojConv_314.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_0_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1179679744, + "file_name": "cache/pos_embedprojConv_315.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_0_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1185061888, + "file_name": "cache/pos_embedprojConv_316.const", + "file_size": 5382144 + }, + "onnx::MatMul_9121_onnx::MatMul_9106": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1190444032, + "file_name": "cache/pos_embedprojConv_317.const", + "file_size": 5382144 + }, + "onnx::MatMul_9122_onnx::MatMul_9107": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1195826176, + "file_name": "cache/pos_embedprojConv_318.const", + "file_size": 5382144 + }, + "onnx::MatMul_9123_onnx::MatMul_9108": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1201208320, + "file_name": "cache/pos_embedprojConv_319.const", + "file_size": 5382144 + }, + "onnx::MatMul_9132_45_5_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1206590464, + "file_name": "cache/pos_embedprojConv_320.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1209281536, + "file_name": "cache/pos_embedprojConv_321.const", + "file_size": 2691072 + }, + "/transformer_blocks.12/norm2/Constant_output_0_14_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1211972608, + "file_name": "cache/pos_embedprojConv_322.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1211975680, + "file_name": "cache/pos_embedprojConv_323.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_3_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1211978752, + "file_name": "cache/pos_embedprojConv_324.const", + "file_size": 5382144 + }, + "onnx::MatMul_9133_45_1_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1217360896, + "file_name": "cache/pos_embedprojConv_325.const", + "file_size": 10764288 + }, + "onnx::MatMul_9134_45_1_51": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1228125184, + "file_name": "cache/pos_embedprojConv_326.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1238889472, + "file_name": "cache/pos_embedprojConv_327.const", + "file_size": 2691072 + }, + "onnx::MatMul_9131_45_5_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1241580544, + "file_name": "cache/pos_embedprojConv_328.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1244271616, + "file_name": "cache/pos_embedprojConv_329.const", + "file_size": 2691072 + }, + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1246962688, + "file_name": "cache/pos_embedprojConv_330.const", + "file_size": 3072 + }, + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1246965760, + "file_name": "cache/pos_embedprojConv_331.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_3_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1246968832, + "file_name": "cache/pos_embedprojConv_332.const", + "file_size": 5382144 + }, + "onnx::MatMul_9135_45_1_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1252350976, + "file_name": "cache/pos_embedprojConv_333.const", + "file_size": 10764288 + }, + "onnx::MatMul_9136_45_1_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1263115264, + "file_name": "cache/pos_embedprojConv_334.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1273879552, + "file_name": "cache/pos_embedprojConv_335.const", + "file_size": 2691072 + }, + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1276570624, + "file_name": "cache/pos_embedprojConv_336.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1276573696, + "file_name": "cache/pos_embedprojConv_337.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1276576768, + "file_name": "cache/pos_embedprojConv_338.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1276579840, + "file_name": "cache/pos_embedprojConv_339.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_0_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1276582912, + "file_name": "cache/pos_embedprojConv_340.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_0_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1281965056, + "file_name": "cache/pos_embedprojConv_341.const", + "file_size": 5382144 + }, + "onnx::MatMul_9152_onnx::MatMul_9137": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1287347200, + "file_name": "cache/pos_embedprojConv_342.const", + "file_size": 5382144 + }, + "onnx::MatMul_9153_onnx::MatMul_9138": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1292729344, + "file_name": "cache/pos_embedprojConv_343.const", + "file_size": 5382144 + }, + "onnx::MatMul_9154_onnx::MatMul_9139": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1298111488, + "file_name": "cache/pos_embedprojConv_344.const", + "file_size": 5382144 + }, + "onnx::MatMul_9163_45_5_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1303493632, + "file_name": "cache/pos_embedprojConv_345.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1306184704, + "file_name": "cache/pos_embedprojConv_346.const", + "file_size": 2691072 + }, + "/transformer_blocks.13/norm2/Constant_output_0_14_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1308875776, + "file_name": "cache/pos_embedprojConv_347.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1308878848, + "file_name": "cache/pos_embedprojConv_348.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_3_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1308881920, + "file_name": "cache/pos_embedprojConv_349.const", + "file_size": 5382144 + }, + "onnx::MatMul_9164_45_1_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1314264064, + "file_name": "cache/pos_embedprojConv_350.const", + "file_size": 10764288 + }, + "onnx::MatMul_9165_45_1_55": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1325028352, + "file_name": "cache/pos_embedprojConv_351.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1335792640, + "file_name": "cache/pos_embedprojConv_352.const", + "file_size": 2691072 + }, + "onnx::MatMul_9162_45_5_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1338483712, + "file_name": "cache/pos_embedprojConv_353.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1341174784, + "file_name": "cache/pos_embedprojConv_354.const", + "file_size": 2691072 + }, + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1343865856, + "file_name": "cache/pos_embedprojConv_355.const", + "file_size": 3072 + }, + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1343868928, + "file_name": "cache/pos_embedprojConv_356.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_3_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1343872000, + "file_name": "cache/pos_embedprojConv_357.const", + "file_size": 5382144 + }, + "onnx::MatMul_9166_45_1_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1349254144, + "file_name": "cache/pos_embedprojConv_358.const", + "file_size": 10764288 + }, + "onnx::MatMul_9167_45_1_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1360018432, + "file_name": "cache/pos_embedprojConv_359.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1370782720, + "file_name": "cache/pos_embedprojConv_360.const", + "file_size": 2691072 + }, + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1373473792, + "file_name": "cache/pos_embedprojConv_361.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1373476864, + "file_name": "cache/pos_embedprojConv_362.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1373479936, + "file_name": "cache/pos_embedprojConv_363.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1373483008, + "file_name": "cache/pos_embedprojConv_364.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_0_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1373486080, + "file_name": "cache/pos_embedprojConv_365.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_0_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1378868224, + "file_name": "cache/pos_embedprojConv_366.const", + "file_size": 5382144 + }, + "onnx::MatMul_9183_onnx::MatMul_9168": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1384250368, + "file_name": "cache/pos_embedprojConv_367.const", + "file_size": 5382144 + }, + "onnx::MatMul_9184_onnx::MatMul_9169": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1389632512, + "file_name": "cache/pos_embedprojConv_368.const", + "file_size": 5382144 + }, + "onnx::MatMul_9185_onnx::MatMul_9170": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1395014656, + "file_name": "cache/pos_embedprojConv_369.const", + "file_size": 5382144 + }, + "onnx::MatMul_9194_45_5_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1400396800, + "file_name": "cache/pos_embedprojConv_370.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1403087872, + "file_name": "cache/pos_embedprojConv_371.const", + "file_size": 2691072 + }, + "/transformer_blocks.14/norm2/Constant_output_0_14_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1405778944, + "file_name": "cache/pos_embedprojConv_372.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1405782016, + "file_name": "cache/pos_embedprojConv_373.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_3_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1405785088, + "file_name": "cache/pos_embedprojConv_374.const", + "file_size": 5382144 + }, + "onnx::MatMul_9195_45_1_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1411167232, + "file_name": "cache/pos_embedprojConv_375.const", + "file_size": 10764288 + }, + "onnx::MatMul_9196_45_1_59": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1421931520, + "file_name": "cache/pos_embedprojConv_376.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1432695808, + "file_name": "cache/pos_embedprojConv_377.const", + "file_size": 2691072 + }, + "onnx::MatMul_9193_45_5_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1435386880, + "file_name": "cache/pos_embedprojConv_378.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1438077952, + "file_name": "cache/pos_embedprojConv_379.const", + "file_size": 2691072 + }, + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1440769024, + "file_name": "cache/pos_embedprojConv_380.const", + "file_size": 3072 + }, + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1440772096, + "file_name": "cache/pos_embedprojConv_381.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_3_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1440775168, + "file_name": "cache/pos_embedprojConv_382.const", + "file_size": 5382144 + }, + "onnx::MatMul_9197_45_1_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1446157312, + "file_name": "cache/pos_embedprojConv_383.const", + "file_size": 10764288 + }, + "onnx::MatMul_9198_45_1_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1456921600, + "file_name": "cache/pos_embedprojConv_384.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1467685888, + "file_name": "cache/pos_embedprojConv_385.const", + "file_size": 2691072 + }, + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1470376960, + "file_name": "cache/pos_embedprojConv_386.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1470380032, + "file_name": "cache/pos_embedprojConv_387.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_61": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1470383104, + "file_name": "cache/pos_embedprojConv_388.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_61": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1470386176, + "file_name": "cache/pos_embedprojConv_389.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_0_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1470389248, + "file_name": "cache/pos_embedprojConv_390.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_0_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1475771392, + "file_name": "cache/pos_embedprojConv_391.const", + "file_size": 5382144 + }, + "onnx::MatMul_9214_onnx::MatMul_9199": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1481153536, + "file_name": "cache/pos_embedprojConv_392.const", + "file_size": 5382144 + }, + "onnx::MatMul_9215_onnx::MatMul_9200": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1486535680, + "file_name": "cache/pos_embedprojConv_393.const", + "file_size": 5382144 + }, + "onnx::MatMul_9216_onnx::MatMul_9201": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1491917824, + "file_name": "cache/pos_embedprojConv_394.const", + "file_size": 5382144 + }, + "onnx::MatMul_9225_45_5_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1497299968, + "file_name": "cache/pos_embedprojConv_395.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1499991040, + "file_name": "cache/pos_embedprojConv_396.const", + "file_size": 2691072 + }, + "/transformer_blocks.15/norm2/Constant_output_0_14_62": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1502682112, + "file_name": "cache/pos_embedprojConv_397.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1502685184, + "file_name": "cache/pos_embedprojConv_398.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_3_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1502688256, + "file_name": "cache/pos_embedprojConv_399.const", + "file_size": 5382144 + }, + "onnx::MatMul_9226_45_1_61": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1508070400, + "file_name": "cache/pos_embedprojConv_400.const", + "file_size": 10764288 + }, + "onnx::MatMul_9227_45_1_63": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1518834688, + "file_name": "cache/pos_embedprojConv_401.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1529598976, + "file_name": "cache/pos_embedprojConv_402.const", + "file_size": 2691072 + }, + "onnx::MatMul_9224_45_5_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1532290048, + "file_name": "cache/pos_embedprojConv_403.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1534981120, + "file_name": "cache/pos_embedprojConv_404.const", + "file_size": 2691072 + }, + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1537672192, + "file_name": "cache/pos_embedprojConv_405.const", + "file_size": 3072 + }, + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1537675264, + "file_name": "cache/pos_embedprojConv_406.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_3_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1537678336, + "file_name": "cache/pos_embedprojConv_407.const", + "file_size": 5382144 + }, + "onnx::MatMul_9228_45_1_62": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1543060480, + "file_name": "cache/pos_embedprojConv_408.const", + "file_size": 10764288 + }, + "onnx::MatMul_9229_45_1_64": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1553824768, + "file_name": "cache/pos_embedprojConv_409.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1564589056, + "file_name": "cache/pos_embedprojConv_410.const", + "file_size": 2691072 + }, + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_64": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1567280128, + "file_name": "cache/pos_embedprojConv_411.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_64": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1567283200, + "file_name": "cache/pos_embedprojConv_412.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_65": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1567286272, + "file_name": "cache/pos_embedprojConv_413.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_65": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1567289344, + "file_name": "cache/pos_embedprojConv_414.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_0_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1567292416, + "file_name": "cache/pos_embedprojConv_415.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_0_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1572674560, + "file_name": "cache/pos_embedprojConv_416.const", + "file_size": 5382144 + }, + "onnx::MatMul_9245_onnx::MatMul_9230": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1578056704, + "file_name": "cache/pos_embedprojConv_417.const", + "file_size": 5382144 + }, + "onnx::MatMul_9246_onnx::MatMul_9231": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1583438848, + "file_name": "cache/pos_embedprojConv_418.const", + "file_size": 5382144 + }, + "onnx::MatMul_9247_onnx::MatMul_9232": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1588820992, + "file_name": "cache/pos_embedprojConv_419.const", + "file_size": 5382144 + }, + "onnx::MatMul_9256_45_5_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1594203136, + "file_name": "cache/pos_embedprojConv_420.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1596894208, + "file_name": "cache/pos_embedprojConv_421.const", + "file_size": 2691072 + }, + "/transformer_blocks.16/norm2/Constant_output_0_14_66": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1599585280, + "file_name": "cache/pos_embedprojConv_422.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1599588352, + "file_name": "cache/pos_embedprojConv_423.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_3_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1599591424, + "file_name": "cache/pos_embedprojConv_424.const", + "file_size": 5382144 + }, + "onnx::MatMul_9257_45_1_65": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1604973568, + "file_name": "cache/pos_embedprojConv_425.const", + "file_size": 10764288 + }, + "onnx::MatMul_9258_45_1_67": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1615737856, + "file_name": "cache/pos_embedprojConv_426.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1626502144, + "file_name": "cache/pos_embedprojConv_427.const", + "file_size": 2691072 + }, + "onnx::MatMul_9255_45_5_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1629193216, + "file_name": "cache/pos_embedprojConv_428.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1631884288, + "file_name": "cache/pos_embedprojConv_429.const", + "file_size": 2691072 + }, + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1634575360, + "file_name": "cache/pos_embedprojConv_430.const", + "file_size": 3072 + }, + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1634578432, + "file_name": "cache/pos_embedprojConv_431.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_3_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1634581504, + "file_name": "cache/pos_embedprojConv_432.const", + "file_size": 5382144 + }, + "onnx::MatMul_9259_45_1_66": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1639963648, + "file_name": "cache/pos_embedprojConv_433.const", + "file_size": 10764288 + }, + "onnx::MatMul_9260_45_1_68": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1650727936, + "file_name": "cache/pos_embedprojConv_434.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1661492224, + "file_name": "cache/pos_embedprojConv_435.const", + "file_size": 2691072 + }, + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_68": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1664183296, + "file_name": "cache/pos_embedprojConv_436.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_68": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1664186368, + "file_name": "cache/pos_embedprojConv_437.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_69": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1664189440, + "file_name": "cache/pos_embedprojConv_438.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_69": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1664192512, + "file_name": "cache/pos_embedprojConv_439.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_0_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1664195584, + "file_name": "cache/pos_embedprojConv_440.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_0_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1669577728, + "file_name": "cache/pos_embedprojConv_441.const", + "file_size": 5382144 + }, + "onnx::MatMul_9276_onnx::MatMul_9261": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1674959872, + "file_name": "cache/pos_embedprojConv_442.const", + "file_size": 5382144 + }, + "onnx::MatMul_9277_onnx::MatMul_9262": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1680342016, + "file_name": "cache/pos_embedprojConv_443.const", + "file_size": 5382144 + }, + "onnx::MatMul_9278_onnx::MatMul_9263": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1685724160, + "file_name": "cache/pos_embedprojConv_444.const", + "file_size": 5382144 + }, + "onnx::MatMul_9287_45_5_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1691106304, + "file_name": "cache/pos_embedprojConv_445.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1693797376, + "file_name": "cache/pos_embedprojConv_446.const", + "file_size": 2691072 + }, + "/transformer_blocks.17/norm2/Constant_output_0_14_70": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1696488448, + "file_name": "cache/pos_embedprojConv_447.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1696491520, + "file_name": "cache/pos_embedprojConv_448.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_3_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1696494592, + "file_name": "cache/pos_embedprojConv_449.const", + "file_size": 5382144 + }, + "onnx::MatMul_9288_45_1_69": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1701876736, + "file_name": "cache/pos_embedprojConv_450.const", + "file_size": 10764288 + }, + "onnx::MatMul_9289_45_1_71": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1712641024, + "file_name": "cache/pos_embedprojConv_451.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1723405312, + "file_name": "cache/pos_embedprojConv_452.const", + "file_size": 2691072 + }, + "onnx::MatMul_9286_45_5_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1726096384, + "file_name": "cache/pos_embedprojConv_453.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1728787456, + "file_name": "cache/pos_embedprojConv_454.const", + "file_size": 2691072 + }, + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1731478528, + "file_name": "cache/pos_embedprojConv_455.const", + "file_size": 3072 + }, + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1731481600, + "file_name": "cache/pos_embedprojConv_456.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_3_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1731484672, + "file_name": "cache/pos_embedprojConv_457.const", + "file_size": 5382144 + }, + "onnx::MatMul_9290_45_1_70": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1736866816, + "file_name": "cache/pos_embedprojConv_458.const", + "file_size": 10764288 + }, + "onnx::MatMul_9291_45_1_72": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1747631104, + "file_name": "cache/pos_embedprojConv_459.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1758395392, + "file_name": "cache/pos_embedprojConv_460.const", + "file_size": 2691072 + }, + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_72": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1761086464, + "file_name": "cache/pos_embedprojConv_461.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_72": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1761089536, + "file_name": "cache/pos_embedprojConv_462.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_73": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1761092608, + "file_name": "cache/pos_embedprojConv_463.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_73": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1761095680, + "file_name": "cache/pos_embedprojConv_464.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_0_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1761098752, + "file_name": "cache/pos_embedprojConv_465.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_0_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1766480896, + "file_name": "cache/pos_embedprojConv_466.const", + "file_size": 5382144 + }, + "onnx::MatMul_9307_onnx::MatMul_9292": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1771863040, + "file_name": "cache/pos_embedprojConv_467.const", + "file_size": 5382144 + }, + "onnx::MatMul_9308_onnx::MatMul_9293": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1777245184, + "file_name": "cache/pos_embedprojConv_468.const", + "file_size": 5382144 + }, + "onnx::MatMul_9309_onnx::MatMul_9294": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1782627328, + "file_name": "cache/pos_embedprojConv_469.const", + "file_size": 5382144 + }, + "onnx::MatMul_9318_45_5_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1788009472, + "file_name": "cache/pos_embedprojConv_470.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1790700544, + "file_name": "cache/pos_embedprojConv_471.const", + "file_size": 2691072 + }, + "/transformer_blocks.18/norm2/Constant_output_0_14_74": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1793391616, + "file_name": "cache/pos_embedprojConv_472.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1793394688, + "file_name": "cache/pos_embedprojConv_473.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_3_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1793397760, + "file_name": "cache/pos_embedprojConv_474.const", + "file_size": 5382144 + }, + "onnx::MatMul_9319_45_1_73": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1798779904, + "file_name": "cache/pos_embedprojConv_475.const", + "file_size": 10764288 + }, + "onnx::MatMul_9320_45_1_75": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1809544192, + "file_name": "cache/pos_embedprojConv_476.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1820308480, + "file_name": "cache/pos_embedprojConv_477.const", + "file_size": 2691072 + }, + "onnx::MatMul_9317_45_5_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1822999552, + "file_name": "cache/pos_embedprojConv_478.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1825690624, + "file_name": "cache/pos_embedprojConv_479.const", + "file_size": 2691072 + }, + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1828381696, + "file_name": "cache/pos_embedprojConv_480.const", + "file_size": 3072 + }, + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1828384768, + "file_name": "cache/pos_embedprojConv_481.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_3_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1828387840, + "file_name": "cache/pos_embedprojConv_482.const", + "file_size": 5382144 + }, + "onnx::MatMul_9321_45_1_74": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1833769984, + "file_name": "cache/pos_embedprojConv_483.const", + "file_size": 10764288 + }, + "onnx::MatMul_9322_45_1_76": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1844534272, + "file_name": "cache/pos_embedprojConv_484.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1855298560, + "file_name": "cache/pos_embedprojConv_485.const", + "file_size": 2691072 + }, + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_76": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1857989632, + "file_name": "cache/pos_embedprojConv_486.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_76": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1857992704, + "file_name": "cache/pos_embedprojConv_487.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_77": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1857995776, + "file_name": "cache/pos_embedprojConv_488.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_77": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1857998848, + "file_name": "cache/pos_embedprojConv_489.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_0_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1858001920, + "file_name": "cache/pos_embedprojConv_490.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_0_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1863384064, + "file_name": "cache/pos_embedprojConv_491.const", + "file_size": 5382144 + }, + "onnx::MatMul_9338_onnx::MatMul_9323": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1868766208, + "file_name": "cache/pos_embedprojConv_492.const", + "file_size": 5382144 + }, + "onnx::MatMul_9339_onnx::MatMul_9324": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1874148352, + "file_name": "cache/pos_embedprojConv_493.const", + "file_size": 5382144 + }, + "onnx::MatMul_9340_onnx::MatMul_9325": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1879530496, + "file_name": "cache/pos_embedprojConv_494.const", + "file_size": 5382144 + }, + "onnx::MatMul_9349_45_5_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1884912640, + "file_name": "cache/pos_embedprojConv_495.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1887603712, + "file_name": "cache/pos_embedprojConv_496.const", + "file_size": 2691072 + }, + "/transformer_blocks.19/norm2/Constant_output_0_14_78": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1890294784, + "file_name": "cache/pos_embedprojConv_497.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1890297856, + "file_name": "cache/pos_embedprojConv_498.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_3_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1890300928, + "file_name": "cache/pos_embedprojConv_499.const", + "file_size": 5382144 + }, + "onnx::MatMul_9350_45_1_77": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1895683072, + "file_name": "cache/pos_embedprojConv_500.const", + "file_size": 10764288 + }, + "onnx::MatMul_9351_45_1_79": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1906447360, + "file_name": "cache/pos_embedprojConv_501.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1917211648, + "file_name": "cache/pos_embedprojConv_502.const", + "file_size": 2691072 + }, + "onnx::MatMul_9348_45_5_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1919902720, + "file_name": "cache/pos_embedprojConv_503.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1922593792, + "file_name": "cache/pos_embedprojConv_504.const", + "file_size": 2691072 + }, + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1925284864, + "file_name": "cache/pos_embedprojConv_505.const", + "file_size": 3072 + }, + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1925287936, + "file_name": "cache/pos_embedprojConv_506.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_3_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1925291008, + "file_name": "cache/pos_embedprojConv_507.const", + "file_size": 5382144 + }, + "onnx::MatMul_9352_45_1_78": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1930673152, + "file_name": "cache/pos_embedprojConv_508.const", + "file_size": 10764288 + }, + "onnx::MatMul_9353_45_1_80": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1941437440, + "file_name": "cache/pos_embedprojConv_509.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1952201728, + "file_name": "cache/pos_embedprojConv_510.const", + "file_size": 2691072 + }, + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_80": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1954892800, + "file_name": "cache/pos_embedprojConv_511.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_80": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1954895872, + "file_name": "cache/pos_embedprojConv_512.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_81": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1954898944, + "file_name": "cache/pos_embedprojConv_513.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_81": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1954902016, + "file_name": "cache/pos_embedprojConv_514.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_0_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1954905088, + "file_name": "cache/pos_embedprojConv_515.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_0_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1960287232, + "file_name": "cache/pos_embedprojConv_516.const", + "file_size": 5382144 + }, + "onnx::MatMul_9369_onnx::MatMul_9354": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1965669376, + "file_name": "cache/pos_embedprojConv_517.const", + "file_size": 5382144 + }, + "onnx::MatMul_9370_onnx::MatMul_9355": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1971051520, + "file_name": "cache/pos_embedprojConv_518.const", + "file_size": 5382144 + }, + "onnx::MatMul_9371_onnx::MatMul_9356": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1976433664, + "file_name": "cache/pos_embedprojConv_519.const", + "file_size": 5382144 + }, + "onnx::MatMul_9380_45_5_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1981815808, + "file_name": "cache/pos_embedprojConv_520.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 1984506880, + "file_name": "cache/pos_embedprojConv_521.const", + "file_size": 2691072 + }, + "/transformer_blocks.20/norm2/Constant_output_0_14_82": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1987197952, + "file_name": "cache/pos_embedprojConv_522.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 1987201024, + "file_name": "cache/pos_embedprojConv_523.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_3_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 1987204096, + "file_name": "cache/pos_embedprojConv_524.const", + "file_size": 5382144 + }, + "onnx::MatMul_9381_45_1_81": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 1992586240, + "file_name": "cache/pos_embedprojConv_525.const", + "file_size": 10764288 + }, + "onnx::MatMul_9382_45_1_83": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2003350528, + "file_name": "cache/pos_embedprojConv_526.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2014114816, + "file_name": "cache/pos_embedprojConv_527.const", + "file_size": 2691072 + }, + "onnx::MatMul_9379_45_5_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2016805888, + "file_name": "cache/pos_embedprojConv_528.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2019496960, + "file_name": "cache/pos_embedprojConv_529.const", + "file_size": 2691072 + }, + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2022188032, + "file_name": "cache/pos_embedprojConv_530.const", + "file_size": 3072 + }, + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2022191104, + "file_name": "cache/pos_embedprojConv_531.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_3_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2022194176, + "file_name": "cache/pos_embedprojConv_532.const", + "file_size": 5382144 + }, + "onnx::MatMul_9383_45_1_82": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2027576320, + "file_name": "cache/pos_embedprojConv_533.const", + "file_size": 10764288 + }, + "onnx::MatMul_9384_45_1_84": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2038340608, + "file_name": "cache/pos_embedprojConv_534.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2049104896, + "file_name": "cache/pos_embedprojConv_535.const", + "file_size": 2691072 + }, + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_84": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2051795968, + "file_name": "cache/pos_embedprojConv_536.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_84": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2051799040, + "file_name": "cache/pos_embedprojConv_537.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_85": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2051802112, + "file_name": "cache/pos_embedprojConv_538.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_85": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2051805184, + "file_name": "cache/pos_embedprojConv_539.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_0_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2051808256, + "file_name": "cache/pos_embedprojConv_540.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_0_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2057190400, + "file_name": "cache/pos_embedprojConv_541.const", + "file_size": 5382144 + }, + "onnx::MatMul_9400_onnx::MatMul_9385": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2062572544, + "file_name": "cache/pos_embedprojConv_542.const", + "file_size": 5382144 + }, + "onnx::MatMul_9401_onnx::MatMul_9386": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2067954688, + "file_name": "cache/pos_embedprojConv_543.const", + "file_size": 5382144 + }, + "onnx::MatMul_9402_onnx::MatMul_9387": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2073336832, + "file_name": "cache/pos_embedprojConv_544.const", + "file_size": 5382144 + }, + "onnx::MatMul_9411_45_5_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2078718976, + "file_name": "cache/pos_embedprojConv_545.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2081410048, + "file_name": "cache/pos_embedprojConv_546.const", + "file_size": 2691072 + }, + "/transformer_blocks.21/norm2/Constant_output_0_14_86": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2084101120, + "file_name": "cache/pos_embedprojConv_547.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2084104192, + "file_name": "cache/pos_embedprojConv_548.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_3_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2084107264, + "file_name": "cache/pos_embedprojConv_549.const", + "file_size": 5382144 + }, + "onnx::MatMul_9412_45_1_85": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2089489408, + "file_name": "cache/pos_embedprojConv_550.const", + "file_size": 10764288 + }, + "onnx::MatMul_9413_45_1_87": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2100253696, + "file_name": "cache/pos_embedprojConv_551.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2111017984, + "file_name": "cache/pos_embedprojConv_552.const", + "file_size": 2691072 + }, + "onnx::MatMul_9410_45_5_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2113709056, + "file_name": "cache/pos_embedprojConv_553.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2116400128, + "file_name": "cache/pos_embedprojConv_554.const", + "file_size": 2691072 + }, + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2119091200, + "file_name": "cache/pos_embedprojConv_555.const", + "file_size": 3072 + }, + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2119094272, + "file_name": "cache/pos_embedprojConv_556.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_3_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2119097344, + "file_name": "cache/pos_embedprojConv_557.const", + "file_size": 5382144 + }, + "onnx::MatMul_9414_45_1_86": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2124479488, + "file_name": "cache/pos_embedprojConv_558.const", + "file_size": 10764288 + }, + "onnx::MatMul_9415_45_1_88": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2135243776, + "file_name": "cache/pos_embedprojConv_559.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2146008064, + "file_name": "cache/pos_embedprojConv_560.const", + "file_size": 2691072 + }, + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_88": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2148699136, + "file_name": "cache/pos_embedprojConv_561.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_88": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2148702208, + "file_name": "cache/pos_embedprojConv_562.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_89": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2148705280, + "file_name": "cache/pos_embedprojConv_563.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_89": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2148708352, + "file_name": "cache/pos_embedprojConv_564.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_0_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2148711424, + "file_name": "cache/pos_embedprojConv_565.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_0_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2154093568, + "file_name": "cache/pos_embedprojConv_566.const", + "file_size": 5382144 + }, + "onnx::MatMul_9431_onnx::MatMul_9416": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2159475712, + "file_name": "cache/pos_embedprojConv_567.const", + "file_size": 5382144 + }, + "onnx::MatMul_9432_onnx::MatMul_9417": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2164857856, + "file_name": "cache/pos_embedprojConv_568.const", + "file_size": 5382144 + }, + "onnx::MatMul_9433_onnx::MatMul_9418": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2170240000, + "file_name": "cache/pos_embedprojConv_569.const", + "file_size": 5382144 + }, + "onnx::MatMul_9442_45_5_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2175622144, + "file_name": "cache/pos_embedprojConv_570.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2178313216, + "file_name": "cache/pos_embedprojConv_571.const", + "file_size": 2691072 + }, + "/transformer_blocks.22/norm2/Constant_output_0_14_90": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2181004288, + "file_name": "cache/pos_embedprojConv_572.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2181007360, + "file_name": "cache/pos_embedprojConv_573.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_3_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2181010432, + "file_name": "cache/pos_embedprojConv_574.const", + "file_size": 5382144 + }, + "onnx::MatMul_9443_45_1_89": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2186392576, + "file_name": "cache/pos_embedprojConv_575.const", + "file_size": 10764288 + }, + "onnx::MatMul_9444_45_1_91": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2197156864, + "file_name": "cache/pos_embedprojConv_576.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2207921152, + "file_name": "cache/pos_embedprojConv_577.const", + "file_size": 2691072 + }, + "onnx::MatMul_9441_45_5_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2210612224, + "file_name": "cache/pos_embedprojConv_578.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2213303296, + "file_name": "cache/pos_embedprojConv_579.const", + "file_size": 2691072 + }, + "/transformer_blocks.22/norm2_context/Constant_output_0_14_91": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2215994368, + "file_name": "cache/pos_embedprojConv_580.const", + "file_size": 3072 + }, + "/transformer_blocks.22/norm2_context/Constant_1_output_0_14_91": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2215997440, + "file_name": "cache/pos_embedprojConv_581.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_3_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2216000512, + "file_name": "cache/pos_embedprojConv_582.const", + "file_size": 5382144 + }, + "onnx::MatMul_9445_45_1_90": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2221382656, + "file_name": "cache/pos_embedprojConv_583.const", + "file_size": 10764288 + }, + "onnx::MatMul_9446_45_1_92": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2232146944, + "file_name": "cache/pos_embedprojConv_584.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2242911232, + "file_name": "cache/pos_embedprojConv_585.const", + "file_size": 2691072 + }, + "/transformer_blocks.23/norm1_context/norm/Constant_output_0_14_92": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2245602304, + "file_name": "cache/pos_embedprojConv_586.const", + "file_size": 3072 + }, + "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0_14_92": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2245605376, + "file_name": "cache/pos_embedprojConv_587.const", + "file_size": 3072 + }, + "/transformer_blocks.23/norm1/norm/Constant_output_0_14_93": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2245608448, + "file_name": "cache/pos_embedprojConv_588.const", + "file_size": 3072 + }, + "/transformer_blocks.23/norm1/norm/Constant_1_output_0_14_93": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2245611520, + "file_name": "cache/pos_embedprojConv_589.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_0_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2245614592, + "file_name": "cache/pos_embedprojConv_590.const", + "file_size": 5382144 + }, + "existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_0_existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2250996736, + "file_name": "cache/pos_embedprojConv_591.const", + "file_size": 5382144 + }, + "onnx::MatMul_9462_onnx::MatMul_9447": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2256378880, + "file_name": "cache/pos_embedprojConv_592.const", + "file_size": 5382144 + }, + "onnx::MatMul_9463_onnx::MatMul_9448": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2261761024, + "file_name": "cache/pos_embedprojConv_593.const", + "file_size": 5382144 + }, + "onnx::MatMul_9464_onnx::MatMul_9449": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2267143168, + "file_name": "cache/pos_embedprojConv_594.const", + "file_size": 5382144 + }, + "onnx::MatMul_9469_45_5_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2272525312, + "file_name": "cache/pos_embedprojConv_595.const", + "file_size": 2691072 + }, + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_2_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2275216384, + "file_name": "cache/pos_embedprojConv_596.const", + "file_size": 2691072 + }, + "/transformer_blocks.23/norm2/Constant_output_0_14_94": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2277907456, + "file_name": "cache/pos_embedprojConv_597.const", + "file_size": 3072 + }, + "/transformer_blocks.23/norm2/Constant_1_output_0_14_94": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2277910528, + "file_name": "cache/pos_embedprojConv_598.const", + "file_size": 3072 + }, + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_3_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2277913600, + "file_name": "cache/pos_embedprojConv_599.const", + "file_size": 5382144 + }, + "onnx::MatMul_9470_45_1_93": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2283295744, + "file_name": "cache/pos_embedprojConv_600.const", + "file_size": 10764288 + }, + "onnx::MatMul_9471_45_1_94": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 10764288 + ], + "size_in_bytes": 10764288, + "op_tensor_size": 10764288, + "offset": 2294060032, + "file_name": "cache/pos_embedprojConv_601.const", + "file_size": 10764288 + }, + "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_5_gma": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 2304824320, + "file_name": "cache/pos_embedprojConv_602.const", + "file_size": 2691072 + }, + "/norm_out/norm/Constant_output_0_14_95": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2307515392, + "file_name": "cache/pos_embedprojConv_603.const", + "file_size": 3072 + }, + "/norm_out/norm/Constant_1_output_0_14_95": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 2307518464, + "file_name": "cache/pos_embedprojConv_604.const", + "file_size": 3072 + }, + "existing_model.norm_out.linear.weight_5_1_52_27_48_0_existing_model.norm_out.linear.weight_5_1_52_27_48_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 2307521536, + "file_name": "cache/pos_embedprojConv_605.const", + "file_size": 5382144 + }, + "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 2312903680, + "file_name": "cache/pos_embedprojConv_606.const", + "file_size": 128 + }, + "onnx::MatMul_9472_45_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 112128 + ], + "size_in_bytes": 112128, + "op_tensor_size": 112128, + "offset": 2312903808, + "file_name": "cache/pos_embedprojConv_607.const", + "file_size": 112128 + } + }, + "dynamic_shape_subgraph": true, + "dynamic_shape_list": [ + { + "max_length + state_dim1": 4256, + "w": 128, + "state_dim1": 4096, + "max_length": 160, + "floor(w/2)": 64, + "state_dim1+max_length": 4256, + "h": 128, + "floor(h/2)": 64, + "batch_size": 2 + } + ], + "aux_info": {} +} \ No newline at end of file diff --git a/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..b04c8b67b8a3f040db2ba7b90e9c99ee6bcdd0d5 --- /dev/null +++ b/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51df5d443923db3813aa5a61a9e6faacb7ad7036820f29e5ee9a4a3ed4a8788c +size 203755381 diff --git a/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst b/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst new file mode 100644 index 0000000000000000000000000000000000000000..0eee1f1d7bdba87856e6bb68326967fc822b54b2 --- /dev/null +++ b/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97db33bb861b37d8e1a31062d7190fcfe187abf8e4a54b0c9a3c4e2d7192d49a +size 2313015552 diff --git a/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state new file mode 100644 index 0000000000000000000000000000000000000000..9d7081a707577ee22f85b9393c5cebce1759165c Binary files /dev/null and b/inpainting/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/inpainting/transformer/dynamic/dd/replaced.onnx b/inpainting/transformer/dynamic/dd/replaced.onnx new file mode 100644 index 0000000000000000000000000000000000000000..70cd30bbc7d4c98e313e0fe1423bdb96b10b449a --- /dev/null +++ b/inpainting/transformer/dynamic/dd/replaced.onnx @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67f9832bdfd94513288839e568d9a668a19f9fb24002a92460206cc44a3911df +size 1054774 diff --git a/inpainting/transformer/dynamic/dd/replaced.onnx.data b/inpainting/transformer/dynamic/dd/replaced.onnx.data new file mode 100644 index 0000000000000000000000000000000000000000..7530b7257c754e617c16e4cfcbfa7ee310fe7811 --- /dev/null +++ b/inpainting/transformer/dynamic/dd/replaced.onnx.data @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:850d66e980fbceb846959c026f2ad96fc5fda8c0402866170acf95dfa534be13 +size 113246208 diff --git a/normal/controlnet-canny/dynamic/dd/cache/pos_embedprojConv_meta.json b/normal/controlnet-canny/dynamic/dd/cache/pos_embedprojConv_meta.json index 87001485fba36b6e28057c6bc1705ccb5f812750..476f4008325d9262ee3793a2c98dab0c4a3afb10 100644 --- a/normal/controlnet-canny/dynamic/dd/cache/pos_embedprojConv_meta.json +++ b/normal/controlnet-canny/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -341,13 +341,13 @@ "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "/time_text_embed/Cast_output_0.out17_3_3" + "/time_text_embed/Cast_output_0.out17_4_3" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2" + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "attrs": { "input_shape": { @@ -405,7 +405,7 @@ "name": "/time_text_embed/timestep_embedder/act/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "const_args": [ "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1" @@ -461,10 +461,10 @@ "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3" + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4" ], "attrs": { "input_shape": { @@ -522,13 +522,13 @@ "name": "/time_text_embed/text_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "pooled_projections.out17_3_1" + "pooled_projections.out17_4_1" ], "const_args": [ - "time_text_embed.text_embedder.linear_1.weight_5_1_0" + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0" ], "out_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "attrs": { "input_shape": { @@ -586,7 +586,7 @@ "name": "/time_text_embed/text_embedder/act_1/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "const_args": [ "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0" @@ -642,10 +642,10 @@ "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" ], "const_args": [ - "time_text_embed.text_embedder.linear_2.weight_5_1_1" + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1" ], "out_args": [ - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "attrs": { "input_shape": { @@ -703,8 +703,8 @@ "name": "/time_text_embed/Add", "type": "SDAdd", "in_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "const_args": [], "out_args": [ @@ -813,16 +813,16 @@ } }, { - "name": "encoder_hidden_states.out17_3_0_SDCastBf2Bfp", + "name": "encoder_hidden_states.out17_4_0_SDCastBf2Bfp", "type": "SDCastBf2Bfp", "in_args": [ - "encoder_hidden_states.out17_3_0" + "encoder_hidden_states.out17_4_0" ], "const_args": [ - "encoder_hidden_states.out17_3_0_bfp.wts" + "encoder_hidden_states.out17_4_0_bfp.wts" ], "out_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -865,13 +865,13 @@ "name": "/context_embedder/MatMul", "type": "SDGemm_bfp", "in_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "const_args": [ - "onnx::MatMul_1943" + "onnx::MatMul_1943_45_1_0" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -926,16 +926,16 @@ } }, { - "name": "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_SDCastBfp2Bf", + "name": "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_SDCastBfp2Bf", "type": "SDCastBfp2Bf", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "const_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "attrs": { "input_shape": { @@ -978,11 +978,11 @@ "name": "/transformer_blocks.0/norm1_context/norm/LayerNormalization", "type": "SDLayerNorm_bfbfp", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0" ], "out_args": [ "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" @@ -1142,8 +1142,8 @@ "/Add_output_0.out_35_1_3" ], "const_args": [ - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1" ], "out_args": [ "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" @@ -1328,6 +1328,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -1360,12 +1366,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -1434,6 +1434,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -1466,12 +1472,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -1534,6 +1534,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -1566,12 +1572,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -1688,7 +1688,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_1964" + "onnx::MatMul_1964_45_5_1" ], "out_args": [ "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1" @@ -1783,7 +1783,7 @@ "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4_gma" @@ -1867,8 +1867,8 @@ "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0" + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3" ], "out_args": [ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" @@ -2028,7 +2028,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_1963" + "onnx::MatMul_1963_45_5_0" ], "out_args": [ "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0" @@ -2124,7 +2124,7 @@ "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" ], "const_args": [ - "onnx::MatMul_1967" + "onnx::MatMul_1967_45_1_2" ], "out_args": [ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" @@ -2194,10 +2194,10 @@ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" ], "const_args": [ - "onnx::MatMul_1968" + "onnx::MatMul_1968_45_1_4" ], "out_args": [ - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8" + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8" ], "attrs": { "input_shape": { @@ -2256,7 +2256,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ @@ -2424,8 +2424,8 @@ "/transformer_blocks.0/Add_output_0.out10_20" ], "const_args": [ - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0" + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2" ], "out_args": [ "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" @@ -2585,7 +2585,7 @@ "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_21" ], "const_args": [ - "onnx::MatMul_1965" + "onnx::MatMul_1965_45_1_1" ], "out_args": [ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" @@ -2655,10 +2655,10 @@ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" ], "const_args": [ - "onnx::MatMul_1966" + "onnx::MatMul_1966_45_1_3" ], "out_args": [ - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7" + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7" ], "attrs": { "input_shape": { @@ -2717,7 +2717,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_output_0.out10_20" ], "const_args": [ @@ -2802,8 +2802,8 @@ "/transformer_blocks.0/Add_3_output_0.out10_21" ], "const_args": [ - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4" ], "out_args": [ "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" @@ -2963,10 +2963,10 @@ "/transformer_blocks.0/Add_3_output_0.out10_21" ], "const_args": [ - "onnx::MatMul_2088" + "onnx::MatMul_2088_45_0_4" ], "out_args": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13" + "/controlnet_blocks.0/Add_output_0.out17_4_13" ], "attrs": { "input_shape": { @@ -3027,8 +3027,8 @@ "/transformer_blocks.0/Add_7_output_0.out10_1" ], "const_args": [ - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5" ], "out_args": [ "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" @@ -3213,6 +3213,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -3245,12 +3251,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -3319,6 +3319,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -3351,12 +3357,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -3419,6 +3419,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -3451,12 +3457,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -3573,7 +3573,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_1988" + "onnx::MatMul_1988_45_5_2" ], "out_args": [ "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2" @@ -3752,8 +3752,8 @@ "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0" + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6" ], "out_args": [ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" @@ -3913,7 +3913,7 @@ "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" ], "const_args": [ - "onnx::MatMul_1990" + "onnx::MatMul_1990_45_1_5" ], "out_args": [ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" @@ -3983,10 +3983,10 @@ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" ], "const_args": [ - "onnx::MatMul_1991" + "onnx::MatMul_1991_45_1_7" ], "out_args": [ - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12" + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12" ], "attrs": { "input_shape": { @@ -4045,7 +4045,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ @@ -4130,10 +4130,10 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "onnx::MatMul_2089" + "onnx::MatMul_2089_45_0_5" ], "out_args": [ - "/controlnet_blocks.1/Add_output_0.out17_3_22" + "/controlnet_blocks.1/Add_output_0.out17_4_22" ], "attrs": { "input_shape": { @@ -4194,7 +4194,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_1989" + "onnx::MatMul_1989_45_5_3" ], "out_args": [ "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3" @@ -4373,8 +4373,8 @@ "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0" + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7" ], "out_args": [ "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" @@ -4534,7 +4534,7 @@ "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" ], "const_args": [ - "onnx::MatMul_1992" + "onnx::MatMul_1992_45_1_6" ], "out_args": [ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" @@ -4604,10 +4604,10 @@ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" ], "const_args": [ - "onnx::MatMul_1993" + "onnx::MatMul_1993_45_1_8" ], "out_args": [ - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13" + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13" ], "attrs": { "input_shape": { @@ -4666,7 +4666,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ @@ -4751,8 +4751,8 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8" ], "out_args": [ "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" @@ -4912,8 +4912,8 @@ "/transformer_blocks.1/Add_7_output_0.out10_5" ], "const_args": [ - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9" ], "out_args": [ "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" @@ -5098,6 +5098,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -5130,12 +5136,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -5204,6 +5204,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -5236,12 +5242,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -5304,6 +5304,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -5336,12 +5342,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -5458,7 +5458,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_2013" + "onnx::MatMul_2013_45_5_4" ], "out_args": [ "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4" @@ -5637,8 +5637,8 @@ "/transformer_blocks.2/Add_output_0.out10_6" ], "const_args": [ - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0" + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10" ], "out_args": [ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" @@ -5798,7 +5798,7 @@ "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_7" ], "const_args": [ - "onnx::MatMul_2015" + "onnx::MatMul_2015_45_1_9" ], "out_args": [ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" @@ -5868,10 +5868,10 @@ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" ], "const_args": [ - "onnx::MatMul_2016" + "onnx::MatMul_2016_45_1_11" ], "out_args": [ - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17" + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17" ], "attrs": { "input_shape": { @@ -5930,7 +5930,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_output_0.out10_6" ], "const_args": [ @@ -6015,10 +6015,10 @@ "/transformer_blocks.2/Add_3_output_0.out10_7" ], "const_args": [ - "onnx::MatMul_2090" + "onnx::MatMul_2090_45_0_6" ], "out_args": [ - "/controlnet_blocks.2/Add_output_0.out17_3_31" + "/controlnet_blocks.2/Add_output_0.out17_4_31" ], "attrs": { "input_shape": { @@ -6079,7 +6079,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_2014" + "onnx::MatMul_2014_45_5_5" ], "out_args": [ "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5" @@ -6258,8 +6258,8 @@ "/transformer_blocks.2/Add_4_output_0.out10_8" ], "const_args": [ - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0" + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11" ], "out_args": [ "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" @@ -6419,7 +6419,7 @@ "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_9" ], "const_args": [ - "onnx::MatMul_2017" + "onnx::MatMul_2017_45_1_10" ], "out_args": [ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" @@ -6489,10 +6489,10 @@ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" ], "const_args": [ - "onnx::MatMul_2018" + "onnx::MatMul_2018_45_1_12" ], "out_args": [ - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18" + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18" ], "attrs": { "input_shape": { @@ -6551,7 +6551,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_4_output_0.out10_8" ], "const_args": [ @@ -6636,8 +6636,8 @@ "/transformer_blocks.2/Add_3_output_0.out10_7" ], "const_args": [ - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12" ], "out_args": [ "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" @@ -6797,8 +6797,8 @@ "/transformer_blocks.2/Add_7_output_0.out10_9" ], "const_args": [ - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13" ], "out_args": [ "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" @@ -6983,6 +6983,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -7015,12 +7021,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -7089,6 +7089,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -7121,12 +7127,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -7189,6 +7189,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -7221,12 +7227,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -7343,7 +7343,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_2038" + "onnx::MatMul_2038_45_5_6" ], "out_args": [ "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6" @@ -7522,8 +7522,8 @@ "/transformer_blocks.3/Add_output_0.out10_10" ], "const_args": [ - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0" + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14" ], "out_args": [ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" @@ -7683,7 +7683,7 @@ "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_11" ], "const_args": [ - "onnx::MatMul_2040" + "onnx::MatMul_2040_45_1_13" ], "out_args": [ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" @@ -7753,10 +7753,10 @@ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" ], "const_args": [ - "onnx::MatMul_2041" + "onnx::MatMul_2041_45_1_15" ], "out_args": [ - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22" + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22" ], "attrs": { "input_shape": { @@ -7815,7 +7815,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_output_0.out10_10" ], "const_args": [ @@ -7900,10 +7900,10 @@ "/transformer_blocks.3/Add_3_output_0.out10_11" ], "const_args": [ - "onnx::MatMul_2091" + "onnx::MatMul_2091_45_0_7" ], "out_args": [ - "/controlnet_blocks.3/Add_output_0.out17_3_40" + "/controlnet_blocks.3/Add_output_0.out17_4_40" ], "attrs": { "input_shape": { @@ -7964,7 +7964,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_2039" + "onnx::MatMul_2039_45_5_7" ], "out_args": [ "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7" @@ -8143,8 +8143,8 @@ "/transformer_blocks.3/Add_4_output_0.out10_12" ], "const_args": [ - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0" + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15" ], "out_args": [ "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" @@ -8304,7 +8304,7 @@ "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_13" ], "const_args": [ - "onnx::MatMul_2042" + "onnx::MatMul_2042_45_1_14" ], "out_args": [ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" @@ -8374,10 +8374,10 @@ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" ], "const_args": [ - "onnx::MatMul_2043" + "onnx::MatMul_2043_45_1_16" ], "out_args": [ - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23" + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23" ], "attrs": { "input_shape": { @@ -8436,7 +8436,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_4_output_0.out10_12" ], "const_args": [ @@ -8521,8 +8521,8 @@ "/transformer_blocks.3/Add_3_output_0.out10_11" ], "const_args": [ - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16" ], "out_args": [ "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" @@ -8682,8 +8682,8 @@ "/transformer_blocks.3/Add_7_output_0.out10_13" ], "const_args": [ - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17" ], "out_args": [ "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" @@ -8868,6 +8868,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -8900,12 +8906,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -8974,6 +8974,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -9006,12 +9012,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -9074,6 +9074,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -9106,12 +9112,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -9228,7 +9228,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_2063" + "onnx::MatMul_2063_45_5_8" ], "out_args": [ "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8" @@ -9407,8 +9407,8 @@ "/transformer_blocks.4/Add_output_0.out10_14" ], "const_args": [ - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0" + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18" ], "out_args": [ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" @@ -9568,7 +9568,7 @@ "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_15" ], "const_args": [ - "onnx::MatMul_2065" + "onnx::MatMul_2065_45_1_17" ], "out_args": [ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" @@ -9638,10 +9638,10 @@ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" ], "const_args": [ - "onnx::MatMul_2066" + "onnx::MatMul_2066_45_1_19" ], "out_args": [ - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27" + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27" ], "attrs": { "input_shape": { @@ -9700,7 +9700,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_output_0.out10_14" ], "const_args": [ @@ -9785,10 +9785,10 @@ "/transformer_blocks.4/Add_3_output_0.out10_15" ], "const_args": [ - "onnx::MatMul_2092" + "onnx::MatMul_2092_45_0_8" ], "out_args": [ - "/controlnet_blocks.4/Add_output_0.out17_3_49" + "/controlnet_blocks.4/Add_output_0.out17_4_49" ], "attrs": { "input_shape": { @@ -9849,7 +9849,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_2064" + "onnx::MatMul_2064_45_5_9" ], "out_args": [ "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9" @@ -10028,8 +10028,8 @@ "/transformer_blocks.4/Add_4_output_0.out10_16" ], "const_args": [ - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0" + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19" ], "out_args": [ "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" @@ -10189,7 +10189,7 @@ "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_17" ], "const_args": [ - "onnx::MatMul_2067" + "onnx::MatMul_2067_45_1_18" ], "out_args": [ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" @@ -10259,10 +10259,10 @@ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" ], "const_args": [ - "onnx::MatMul_2068" + "onnx::MatMul_2068_45_1_20" ], "out_args": [ - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28" + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28" ], "attrs": { "input_shape": { @@ -10321,7 +10321,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_4_output_0.out10_16" ], "const_args": [ @@ -10406,8 +10406,8 @@ "/transformer_blocks.4/Add_3_output_0.out10_15" ], "const_args": [ - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20" ], "out_args": [ "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" @@ -10567,8 +10567,8 @@ "/transformer_blocks.4/Add_7_output_0.out10_17" ], "const_args": [ - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21" ], "out_args": [ "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" @@ -10753,6 +10753,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -10785,12 +10791,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -10859,6 +10859,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -10891,12 +10897,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -10959,6 +10959,12 @@ "1" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "input_shape_0": { "type": "str", "value": [ @@ -10991,12 +10997,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "weight_shape_0": { "type": "int", "value": [ @@ -11113,7 +11113,7 @@ "/transformer_blocks.5/attn/Reshape_3_output_0.out22_5_bfp.out27_0_5" ], "const_args": [ - "onnx::MatMul_2085" + "onnx::MatMul_2085_45_5_10" ], "out_args": [ "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10" @@ -11292,8 +11292,8 @@ "/transformer_blocks.5/Add_output_0.out10_18" ], "const_args": [ - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0" + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22" ], "out_args": [ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" @@ -11453,7 +11453,7 @@ "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_19" ], "const_args": [ - "onnx::MatMul_2086" + "onnx::MatMul_2086_45_1_21" ], "out_args": [ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" @@ -11523,10 +11523,10 @@ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" ], "const_args": [ - "onnx::MatMul_2087" + "onnx::MatMul_2087_45_1_22" ], "out_args": [ - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31" + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31" ], "attrs": { "input_shape": { @@ -11585,7 +11585,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31", "/transformer_blocks.5/Add_output_0.out10_18" ], "const_args": [ @@ -11670,10 +11670,10 @@ "/transformer_blocks.5/Add_3_output_0.out10_19" ], "const_args": [ - "onnx::MatMul_2093" + "onnx::MatMul_2093_45_0_9" ], "out_args": [ - "/controlnet_blocks.5/Add_output_0.out17_3_57" + "/controlnet_blocks.5/Add_output_0.out17_4_57" ], "attrs": { "input_shape": { @@ -11736,21 +11736,21 @@ "hidden_states_nhwc.out5_0_0", "/pos_embed/Reshape_1_output_0.out_35_1_2", "controlnet_cond_nhwc.out5_0_1", - "/time_text_embed/Cast_output_0.out17_3_3", - "pooled_projections.out17_3_1", - "encoder_hidden_states.out17_3_0" + "/time_text_embed/Cast_output_0.out17_4_3", + "pooled_projections.out17_4_1", + "encoder_hidden_states.out17_4_0" ] }, "out": { "buffer_size": 18432, "xrt_arg_id": 1, "packed_tensors": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13", - "/controlnet_blocks.1/Add_output_0.out17_3_22", - "/controlnet_blocks.2/Add_output_0.out17_3_31", - "/controlnet_blocks.3/Add_output_0.out17_3_40", - "/controlnet_blocks.4/Add_output_0.out17_3_49", - "/controlnet_blocks.5/Add_output_0.out17_3_57" + "/controlnet_blocks.0/Add_output_0.out17_4_13", + "/controlnet_blocks.1/Add_output_0.out17_4_22", + "/controlnet_blocks.2/Add_output_0.out17_4_31", + "/controlnet_blocks.3/Add_output_0.out17_4_40", + "/controlnet_blocks.4/Add_output_0.out17_4_49", + "/controlnet_blocks.5/Add_output_0.out17_4_57" ] }, "scratch": { @@ -11761,17 +11761,17 @@ "/pos_embed/Add_2_output_0.out_35_1_2", "/pos_embed_input/Transpose_output_0.out5_0_1", "/Add_output_0.out_35_1_3", - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3", "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1", - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1", "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2", "/time_text_embed/Add_output_0.out_35_1_4", "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "encoder_hidden_states.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0", + "encoder_hidden_states.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0", "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0", "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1", @@ -11786,13 +11786,13 @@ "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1", "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_7_output_0.out10_1", "/transformer_blocks.0/Add_output_0.out10_20", "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2", "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_21", "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_3_output_0.out10_21", "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4", "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2", @@ -11807,14 +11807,14 @@ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6", "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3", "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_3_output_0.out10_3", "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", "/transformer_blocks.1/Add_4_output_0.out10_4", "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7", "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5", "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_7_output_0.out10_5", "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8", "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_6", @@ -11829,14 +11829,14 @@ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10", "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_7", "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_3_output_0.out10_7", "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5", "/transformer_blocks.2/Add_4_output_0.out10_8", "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11", "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_9", "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_7_output_0.out10_9", "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12", "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_10", @@ -11851,14 +11851,14 @@ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14", "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_11", "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_3_output_0.out10_11", "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7", "/transformer_blocks.3/Add_4_output_0.out10_12", "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15", "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_13", "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_7_output_0.out10_13", "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16", "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_14", @@ -11873,14 +11873,14 @@ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18", "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_15", "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_3_output_0.out10_15", "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9", "/transformer_blocks.4/Add_4_output_0.out10_16", "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19", "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_17", "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_7_output_0.out10_17", "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20", "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_18", @@ -11895,7 +11895,7 @@ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22", "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_19", "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31", "/transformer_blocks.5/Add_3_output_0.out10_19" ] }, @@ -11905,164 +11905,164 @@ "packed_tensors": [ "pos_embed.proj.weight", "pos_embed_input.proj.weight", - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2", + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2", "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1", - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3", - "time_text_embed.text_embedder.linear_1.weight_5_1_0", + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3", + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0", "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0", - "time_text_embed.text_embedder.linear_2.weight_5_1_1", + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1", "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2", - "encoder_hidden_states.out17_3_0_bfp.wts", - "onnx::MatMul_1943", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts", - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0", + "encoder_hidden_states.out17_4_0_bfp.wts", + "onnx::MatMul_1943_45_1_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts", + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0", - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0", + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3", "onnx::MatMul_1947_onnx::MatMul_1944", "onnx::MatMul_1948_onnx::MatMul_1945", "onnx::MatMul_1949_onnx::MatMul_1946", - "onnx::MatMul_1964", + "onnx::MatMul_1964_45_5_1", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4_gma", - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0", + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2", - "onnx::MatMul_1963", - "onnx::MatMul_1967", - "onnx::MatMul_1968", + "onnx::MatMul_1963_45_5_0", + "onnx::MatMul_1967_45_1_2", + "onnx::MatMul_1968_45_1_4", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma", - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0", + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0", - "onnx::MatMul_1965", - "onnx::MatMul_1966", + "onnx::MatMul_1965_45_1_1", + "onnx::MatMul_1966_45_1_3", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma", - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0", + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1", - "onnx::MatMul_2088", - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0", + "onnx::MatMul_2088_45_0_4", + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1", "onnx::MatMul_1972_onnx::MatMul_1969", "onnx::MatMul_1973_onnx::MatMul_1970", "onnx::MatMul_1974_onnx::MatMul_1971", - "onnx::MatMul_1988", + "onnx::MatMul_1988_45_5_2", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma", - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0", + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4", - "onnx::MatMul_1990", - "onnx::MatMul_1991", + "onnx::MatMul_1990_45_1_5", + "onnx::MatMul_1991_45_1_7", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma", - "onnx::MatMul_2089", - "onnx::MatMul_1989", + "onnx::MatMul_2089_45_0_5", + "onnx::MatMul_1989_45_5_3", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma", - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0", + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4", - "onnx::MatMul_1992", - "onnx::MatMul_1993", + "onnx::MatMul_1992_45_1_6", + "onnx::MatMul_1993_45_1_8", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma", - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1", - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1", "onnx::MatMul_1997_onnx::MatMul_1994", "onnx::MatMul_1998_onnx::MatMul_1995", "onnx::MatMul_1999_onnx::MatMul_1996", - "onnx::MatMul_2013", + "onnx::MatMul_2013_45_5_4", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma", - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0", + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4", - "onnx::MatMul_2015", - "onnx::MatMul_2016", + "onnx::MatMul_2015_45_1_9", + "onnx::MatMul_2016_45_1_11", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma", - "onnx::MatMul_2090", - "onnx::MatMul_2014", + "onnx::MatMul_2090_45_0_6", + "onnx::MatMul_2014_45_5_5", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma", - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0", + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4", - "onnx::MatMul_2017", - "onnx::MatMul_2018", + "onnx::MatMul_2017_45_1_10", + "onnx::MatMul_2018_45_1_12", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma", - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1", - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1", "onnx::MatMul_2022_onnx::MatMul_2019", "onnx::MatMul_2023_onnx::MatMul_2020", "onnx::MatMul_2024_onnx::MatMul_2021", - "onnx::MatMul_2038", + "onnx::MatMul_2038_45_5_6", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma", - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0", + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4", - "onnx::MatMul_2040", - "onnx::MatMul_2041", + "onnx::MatMul_2040_45_1_13", + "onnx::MatMul_2041_45_1_15", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma", - "onnx::MatMul_2091", - "onnx::MatMul_2039", + "onnx::MatMul_2091_45_0_7", + "onnx::MatMul_2039_45_5_7", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma", - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0", + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4", - "onnx::MatMul_2042", - "onnx::MatMul_2043", + "onnx::MatMul_2042_45_1_14", + "onnx::MatMul_2043_45_1_16", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma", - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1", - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1", "onnx::MatMul_2047_onnx::MatMul_2044", "onnx::MatMul_2048_onnx::MatMul_2045", "onnx::MatMul_2049_onnx::MatMul_2046", - "onnx::MatMul_2063", + "onnx::MatMul_2063_45_5_8", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma", - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0", + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4", - "onnx::MatMul_2065", - "onnx::MatMul_2066", + "onnx::MatMul_2065_45_1_17", + "onnx::MatMul_2066_45_1_19", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma", - "onnx::MatMul_2092", - "onnx::MatMul_2064", + "onnx::MatMul_2092_45_0_8", + "onnx::MatMul_2064_45_5_9", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma", - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0", + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4", - "onnx::MatMul_2067", - "onnx::MatMul_2068", + "onnx::MatMul_2067_45_1_18", + "onnx::MatMul_2068_45_1_20", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma", - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1", - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21", "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1", "onnx::MatMul_2072_onnx::MatMul_2069", "onnx::MatMul_2073_onnx::MatMul_2070", "onnx::MatMul_2074_onnx::MatMul_2071", - "onnx::MatMul_2085", + "onnx::MatMul_2085_45_5_10", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma", - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0", + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4", - "onnx::MatMul_2086", - "onnx::MatMul_2087", + "onnx::MatMul_2086_45_1_21", + "onnx::MatMul_2087_45_1_22", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma", - "onnx::MatMul_2093" + "onnx::MatMul_2093_45_0_9" ] }, "super_instr": { @@ -12130,7 +12130,7 @@ ], "offset": 3104 }, - "/time_text_embed/Cast_output_0.out17_3_3": { + "/time_text_embed/Cast_output_0.out17_4_3": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12146,7 +12146,7 @@ ], "offset": 3136 }, - "pooled_projections.out17_3_1": { + "pooled_projections.out17_4_1": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12162,7 +12162,7 @@ ], "offset": 3648 }, - "encoder_hidden_states.out17_3_0": { + "encoder_hidden_states.out17_4_0": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12180,7 +12180,7 @@ ], "offset": 7744 }, - "/controlnet_blocks.0/Add_output_0.out17_3_13": { + "/controlnet_blocks.0/Add_output_0.out17_4_13": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12198,7 +12198,7 @@ ], "offset": 0 }, - "/controlnet_blocks.1/Add_output_0.out17_3_22": { + "/controlnet_blocks.1/Add_output_0.out17_4_22": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12216,7 +12216,7 @@ ], "offset": 3072 }, - "/controlnet_blocks.2/Add_output_0.out17_3_31": { + "/controlnet_blocks.2/Add_output_0.out17_4_31": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12234,7 +12234,7 @@ ], "offset": 6144 }, - "/controlnet_blocks.3/Add_output_0.out17_3_40": { + "/controlnet_blocks.3/Add_output_0.out17_4_40": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12252,7 +12252,7 @@ ], "offset": 9216 }, - "/controlnet_blocks.4/Add_output_0.out17_3_49": { + "/controlnet_blocks.4/Add_output_0.out17_4_49": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12270,7 +12270,7 @@ ], "offset": 12288 }, - "/controlnet_blocks.5/Add_output_0.out17_3_57": { + "/controlnet_blocks.5/Add_output_0.out17_4_57": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12360,7 +12360,7 @@ ], "offset": 9216 }, - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3": { + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12392,7 +12392,7 @@ ], "offset": 15360 }, - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4": { + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12408,7 +12408,7 @@ ], "offset": 18432 }, - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1": { + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12440,7 +12440,7 @@ ], "offset": 24576 }, - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2": { + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12488,7 +12488,7 @@ ], "offset": 33792 }, - "encoder_hidden_states.out17_3_0_bfp.out25_0": { + "encoder_hidden_states.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12506,7 +12506,7 @@ ], "offset": 36864 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12524,7 +12524,7 @@ ], "offset": 40960 }, - "/context_embedder/Add_output_0.out17_3_0": { + "/context_embedder/Add_output_0.out17_4_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12800,7 +12800,7 @@ ], "offset": 67072 }, - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8": { + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12908,7 +12908,7 @@ ], "offset": 83968 }, - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7": { + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13184,7 +13184,7 @@ ], "offset": 114688 }, - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12": { + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13310,7 +13310,7 @@ ], "offset": 133120 }, - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13": { + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13586,7 +13586,7 @@ ], "offset": 163840 }, - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17": { + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13712,7 +13712,7 @@ ], "offset": 182272 }, - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18": { + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13988,7 +13988,7 @@ ], "offset": 212992 }, - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22": { + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14114,7 +14114,7 @@ ], "offset": 231424 }, - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23": { + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14390,7 +14390,7 @@ ], "offset": 262144 }, - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27": { + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14516,7 +14516,7 @@ ], "offset": 280576 }, - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28": { + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14792,7 +14792,7 @@ ], "offset": 311296 }, - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31": { + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14854,7 +14854,7 @@ "file_name": "cache/pos_embedprojConv_1.const", "file_size": 135168 }, - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2": { + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14880,7 +14880,7 @@ "file_name": "cache/pos_embedprojConv_3.const", "file_size": 256 }, - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3": { + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14893,7 +14893,7 @@ "file_name": "cache/pos_embedprojConv_4.const", "file_size": 2691072 }, - "time_text_embed.text_embedder.linear_1.weight_5_1_0": { + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14919,7 +14919,7 @@ "file_name": "cache/pos_embedprojConv_6.const", "file_size": 256 }, - "time_text_embed.text_embedder.linear_2.weight_5_1_1": { + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14945,7 +14945,7 @@ "file_name": "cache/pos_embedprojConv_8.const", "file_size": 256 }, - "encoder_hidden_states.out17_3_0_bfp.wts": { + "encoder_hidden_states.out17_4_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14958,7 +14958,7 @@ "file_name": "cache/pos_embedprojConv_9.const", "file_size": 128 }, - "onnx::MatMul_1943": { + "onnx::MatMul_1943_45_1_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14971,7 +14971,7 @@ "file_name": "cache/pos_embedprojConv_10.const", "file_size": 7274496 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14984,7 +14984,7 @@ "file_name": "cache/pos_embedprojConv_11.const", "file_size": 128 }, - "/transformer_blocks.0/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14997,7 +14997,7 @@ "file_name": "cache/pos_embedprojConv_12.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15023,7 +15023,7 @@ "file_name": "cache/pos_embedprojConv_14.const", "file_size": 5382144 }, - "/transformer_blocks.0/norm1/norm/Constant_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15036,7 +15036,7 @@ "file_name": "cache/pos_embedprojConv_15.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15101,7 +15101,7 @@ "file_name": "cache/pos_embedprojConv_20.const", "file_size": 5382144 }, - "onnx::MatMul_1964": { + "onnx::MatMul_1964_45_5_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15127,7 +15127,7 @@ "file_name": "cache/pos_embedprojConv_22.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2_context/Constant_output_0": { + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15140,7 +15140,7 @@ "file_name": "cache/pos_embedprojConv_23.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2_context/Constant_1_output_0": { + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15166,7 +15166,7 @@ "file_name": "cache/pos_embedprojConv_25.const", "file_size": 5382144 }, - "onnx::MatMul_1963": { + "onnx::MatMul_1963_45_5_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15179,7 +15179,7 @@ "file_name": "cache/pos_embedprojConv_26.const", "file_size": 2691072 }, - "onnx::MatMul_1967": { + "onnx::MatMul_1967_45_1_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15192,7 +15192,7 @@ "file_name": "cache/pos_embedprojConv_27.const", "file_size": 10764288 }, - "onnx::MatMul_1968": { + "onnx::MatMul_1968_45_1_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15231,7 +15231,7 @@ "file_name": "cache/pos_embedprojConv_30.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2/Constant_output_0": { + "/transformer_blocks.0/norm2/Constant_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15244,7 +15244,7 @@ "file_name": "cache/pos_embedprojConv_31.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2/Constant_1_output_0": { + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15270,7 +15270,7 @@ "file_name": "cache/pos_embedprojConv_33.const", "file_size": 5382144 }, - "onnx::MatMul_1965": { + "onnx::MatMul_1965_45_1_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15283,7 +15283,7 @@ "file_name": "cache/pos_embedprojConv_34.const", "file_size": 10764288 }, - "onnx::MatMul_1966": { + "onnx::MatMul_1966_45_1_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15309,7 +15309,7 @@ "file_name": "cache/pos_embedprojConv_36.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1/norm/Constant_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15322,7 +15322,7 @@ "file_name": "cache/pos_embedprojConv_37.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15348,7 +15348,7 @@ "file_name": "cache/pos_embedprojConv_39.const", "file_size": 5382144 }, - "onnx::MatMul_2088": { + "onnx::MatMul_2088_45_0_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15361,7 +15361,7 @@ "file_name": "cache/pos_embedprojConv_40.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15374,7 +15374,7 @@ "file_name": "cache/pos_embedprojConv_41.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15439,7 +15439,7 @@ "file_name": "cache/pos_embedprojConv_46.const", "file_size": 5382144 }, - "onnx::MatMul_1988": { + "onnx::MatMul_1988_45_5_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15465,7 +15465,7 @@ "file_name": "cache/pos_embedprojConv_48.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2/Constant_output_0": { + "/transformer_blocks.1/norm2/Constant_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15478,7 +15478,7 @@ "file_name": "cache/pos_embedprojConv_49.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2/Constant_1_output_0": { + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15504,7 +15504,7 @@ "file_name": "cache/pos_embedprojConv_51.const", "file_size": 5382144 }, - "onnx::MatMul_1990": { + "onnx::MatMul_1990_45_1_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15517,7 +15517,7 @@ "file_name": "cache/pos_embedprojConv_52.const", "file_size": 10764288 }, - "onnx::MatMul_1991": { + "onnx::MatMul_1991_45_1_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15543,7 +15543,7 @@ "file_name": "cache/pos_embedprojConv_54.const", "file_size": 2691072 }, - "onnx::MatMul_2089": { + "onnx::MatMul_2089_45_0_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15556,7 +15556,7 @@ "file_name": "cache/pos_embedprojConv_55.const", "file_size": 2691072 }, - "onnx::MatMul_1989": { + "onnx::MatMul_1989_45_5_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15582,7 +15582,7 @@ "file_name": "cache/pos_embedprojConv_57.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2_context/Constant_output_0": { + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15595,7 +15595,7 @@ "file_name": "cache/pos_embedprojConv_58.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2_context/Constant_1_output_0": { + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15621,7 +15621,7 @@ "file_name": "cache/pos_embedprojConv_60.const", "file_size": 5382144 }, - "onnx::MatMul_1992": { + "onnx::MatMul_1992_45_1_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15634,7 +15634,7 @@ "file_name": "cache/pos_embedprojConv_61.const", "file_size": 10764288 }, - "onnx::MatMul_1993": { + "onnx::MatMul_1993_45_1_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15660,7 +15660,7 @@ "file_name": "cache/pos_embedprojConv_63.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm1/norm/Constant_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15673,7 +15673,7 @@ "file_name": "cache/pos_embedprojConv_64.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15699,7 +15699,7 @@ "file_name": "cache/pos_embedprojConv_66.const", "file_size": 5382144 }, - "/transformer_blocks.2/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15712,7 +15712,7 @@ "file_name": "cache/pos_embedprojConv_67.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15777,7 +15777,7 @@ "file_name": "cache/pos_embedprojConv_72.const", "file_size": 5382144 }, - "onnx::MatMul_2013": { + "onnx::MatMul_2013_45_5_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15803,7 +15803,7 @@ "file_name": "cache/pos_embedprojConv_74.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2/Constant_output_0": { + "/transformer_blocks.2/norm2/Constant_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15816,7 +15816,7 @@ "file_name": "cache/pos_embedprojConv_75.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2/Constant_1_output_0": { + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15842,7 +15842,7 @@ "file_name": "cache/pos_embedprojConv_77.const", "file_size": 5382144 }, - "onnx::MatMul_2015": { + "onnx::MatMul_2015_45_1_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15855,7 +15855,7 @@ "file_name": "cache/pos_embedprojConv_78.const", "file_size": 10764288 }, - "onnx::MatMul_2016": { + "onnx::MatMul_2016_45_1_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15881,7 +15881,7 @@ "file_name": "cache/pos_embedprojConv_80.const", "file_size": 2691072 }, - "onnx::MatMul_2090": { + "onnx::MatMul_2090_45_0_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15894,7 +15894,7 @@ "file_name": "cache/pos_embedprojConv_81.const", "file_size": 2691072 }, - "onnx::MatMul_2014": { + "onnx::MatMul_2014_45_5_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15920,7 +15920,7 @@ "file_name": "cache/pos_embedprojConv_83.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2_context/Constant_output_0": { + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15933,7 +15933,7 @@ "file_name": "cache/pos_embedprojConv_84.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2_context/Constant_1_output_0": { + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15959,7 +15959,7 @@ "file_name": "cache/pos_embedprojConv_86.const", "file_size": 5382144 }, - "onnx::MatMul_2017": { + "onnx::MatMul_2017_45_1_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15972,7 +15972,7 @@ "file_name": "cache/pos_embedprojConv_87.const", "file_size": 10764288 }, - "onnx::MatMul_2018": { + "onnx::MatMul_2018_45_1_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15998,7 +15998,7 @@ "file_name": "cache/pos_embedprojConv_89.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm1/norm/Constant_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16011,7 +16011,7 @@ "file_name": "cache/pos_embedprojConv_90.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16037,7 +16037,7 @@ "file_name": "cache/pos_embedprojConv_92.const", "file_size": 5382144 }, - "/transformer_blocks.3/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16050,7 +16050,7 @@ "file_name": "cache/pos_embedprojConv_93.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16115,7 +16115,7 @@ "file_name": "cache/pos_embedprojConv_98.const", "file_size": 5382144 }, - "onnx::MatMul_2038": { + "onnx::MatMul_2038_45_5_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16141,7 +16141,7 @@ "file_name": "cache/pos_embedprojConv_100.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2/Constant_output_0": { + "/transformer_blocks.3/norm2/Constant_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16154,7 +16154,7 @@ "file_name": "cache/pos_embedprojConv_101.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2/Constant_1_output_0": { + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16180,7 +16180,7 @@ "file_name": "cache/pos_embedprojConv_103.const", "file_size": 5382144 }, - "onnx::MatMul_2040": { + "onnx::MatMul_2040_45_1_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16193,7 +16193,7 @@ "file_name": "cache/pos_embedprojConv_104.const", "file_size": 10764288 }, - "onnx::MatMul_2041": { + "onnx::MatMul_2041_45_1_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16219,7 +16219,7 @@ "file_name": "cache/pos_embedprojConv_106.const", "file_size": 2691072 }, - "onnx::MatMul_2091": { + "onnx::MatMul_2091_45_0_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16232,7 +16232,7 @@ "file_name": "cache/pos_embedprojConv_107.const", "file_size": 2691072 }, - "onnx::MatMul_2039": { + "onnx::MatMul_2039_45_5_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16258,7 +16258,7 @@ "file_name": "cache/pos_embedprojConv_109.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2_context/Constant_output_0": { + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16271,7 +16271,7 @@ "file_name": "cache/pos_embedprojConv_110.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2_context/Constant_1_output_0": { + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16297,7 +16297,7 @@ "file_name": "cache/pos_embedprojConv_112.const", "file_size": 5382144 }, - "onnx::MatMul_2042": { + "onnx::MatMul_2042_45_1_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16310,7 +16310,7 @@ "file_name": "cache/pos_embedprojConv_113.const", "file_size": 10764288 }, - "onnx::MatMul_2043": { + "onnx::MatMul_2043_45_1_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16336,7 +16336,7 @@ "file_name": "cache/pos_embedprojConv_115.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm1/norm/Constant_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16349,7 +16349,7 @@ "file_name": "cache/pos_embedprojConv_116.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16375,7 +16375,7 @@ "file_name": "cache/pos_embedprojConv_118.const", "file_size": 5382144 }, - "/transformer_blocks.4/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16388,7 +16388,7 @@ "file_name": "cache/pos_embedprojConv_119.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16453,7 +16453,7 @@ "file_name": "cache/pos_embedprojConv_124.const", "file_size": 5382144 }, - "onnx::MatMul_2063": { + "onnx::MatMul_2063_45_5_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16479,7 +16479,7 @@ "file_name": "cache/pos_embedprojConv_126.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2/Constant_output_0": { + "/transformer_blocks.4/norm2/Constant_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16492,7 +16492,7 @@ "file_name": "cache/pos_embedprojConv_127.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2/Constant_1_output_0": { + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16518,7 +16518,7 @@ "file_name": "cache/pos_embedprojConv_129.const", "file_size": 5382144 }, - "onnx::MatMul_2065": { + "onnx::MatMul_2065_45_1_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16531,7 +16531,7 @@ "file_name": "cache/pos_embedprojConv_130.const", "file_size": 10764288 }, - "onnx::MatMul_2066": { + "onnx::MatMul_2066_45_1_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16557,7 +16557,7 @@ "file_name": "cache/pos_embedprojConv_132.const", "file_size": 2691072 }, - "onnx::MatMul_2092": { + "onnx::MatMul_2092_45_0_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16570,7 +16570,7 @@ "file_name": "cache/pos_embedprojConv_133.const", "file_size": 2691072 }, - "onnx::MatMul_2064": { + "onnx::MatMul_2064_45_5_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16596,7 +16596,7 @@ "file_name": "cache/pos_embedprojConv_135.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2_context/Constant_output_0": { + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16609,7 +16609,7 @@ "file_name": "cache/pos_embedprojConv_136.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2_context/Constant_1_output_0": { + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16635,7 +16635,7 @@ "file_name": "cache/pos_embedprojConv_138.const", "file_size": 5382144 }, - "onnx::MatMul_2067": { + "onnx::MatMul_2067_45_1_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16648,7 +16648,7 @@ "file_name": "cache/pos_embedprojConv_139.const", "file_size": 10764288 }, - "onnx::MatMul_2068": { + "onnx::MatMul_2068_45_1_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16674,7 +16674,7 @@ "file_name": "cache/pos_embedprojConv_141.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm1/norm/Constant_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16687,7 +16687,7 @@ "file_name": "cache/pos_embedprojConv_142.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16713,7 +16713,7 @@ "file_name": "cache/pos_embedprojConv_144.const", "file_size": 5382144 }, - "/transformer_blocks.5/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16726,7 +16726,7 @@ "file_name": "cache/pos_embedprojConv_145.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16791,7 +16791,7 @@ "file_name": "cache/pos_embedprojConv_150.const", "file_size": 5382144 }, - "onnx::MatMul_2085": { + "onnx::MatMul_2085_45_5_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16817,7 +16817,7 @@ "file_name": "cache/pos_embedprojConv_152.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm2/Constant_output_0": { + "/transformer_blocks.5/norm2/Constant_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16830,7 +16830,7 @@ "file_name": "cache/pos_embedprojConv_153.const", "file_size": 3072 }, - "/transformer_blocks.5/norm2/Constant_1_output_0": { + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16856,7 +16856,7 @@ "file_name": "cache/pos_embedprojConv_155.const", "file_size": 5382144 }, - "onnx::MatMul_2086": { + "onnx::MatMul_2086_45_1_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16869,7 +16869,7 @@ "file_name": "cache/pos_embedprojConv_156.const", "file_size": 10764288 }, - "onnx::MatMul_2087": { + "onnx::MatMul_2087_45_1_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16895,7 +16895,7 @@ "file_name": "cache/pos_embedprojConv_158.const", "file_size": 2691072 }, - "onnx::MatMul_2093": { + "onnx::MatMul_2093_45_0_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16912,84 +16912,204 @@ "dynamic_shape_subgraph": true, "dynamic_shape_list": [ { + "w": 64, "floor(h/2)*floor(w/2)": 1024, - "floor(h/2)": 32, - "max_length + floor(h/2)*floor(w/2)": 1184, + "batch_size": 2, "max_length": 160, - "h": 64, + "max_length + floor(h/2)*floor(w/2)": 1184, + "floor(h/2)": 32, "floor(w/2)": 32, + "h": 64 + }, + { + "w": 80, + "floor(h/2)*floor(w/2)": 1600, "batch_size": 2, - "w": 64 + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1760, + "floor(h/2)": 40, + "floor(w/2)": 40, + "h": 80 }, { - "floor(h/2)*floor(w/2)": 1536, - "floor(h/2)": 32, - "max_length + floor(h/2)*floor(w/2)": 1696, + "w": 96, + "floor(h/2)*floor(w/2)": 2304, + "batch_size": 2, "max_length": 160, - "h": 64, + "max_length + floor(h/2)*floor(w/2)": 2464, + "floor(h/2)": 48, "floor(w/2)": 48, + "h": 96 + }, + { + "w": 128, + "floor(h/2)*floor(w/2)": 4096, "batch_size": 2, - "w": 96 + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 4256, + "floor(h/2)": 64, + "floor(w/2)": 64, + "h": 128 }, { - "floor(h/2)*floor(w/2)": 1536, + "w": 48, + "floor(h/2)*floor(w/2)": 768, + "batch_size": 2, + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 928, + "floor(h/2)": 32, + "floor(w/2)": 24, + "h": 64 + }, + { + "w": 60, + "floor(h/2)*floor(w/2)": 1200, + "batch_size": 2, + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1360, + "floor(h/2)": 40, + "floor(w/2)": 30, + "h": 80 + }, + { + "w": 72, + "floor(h/2)*floor(w/2)": 1728, + "batch_size": 2, + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1888, "floor(h/2)": 48, - "max_length + floor(h/2)*floor(w/2)": 1696, + "floor(w/2)": 36, + "h": 96 + }, + { + "w": 96, + "floor(h/2)*floor(w/2)": 3072, + "batch_size": 2, "max_length": 160, - "h": 96, + "max_length + floor(h/2)*floor(w/2)": 3232, + "floor(h/2)": 64, + "floor(w/2)": 48, + "h": 128 + }, + { + "w": 64, + "floor(h/2)*floor(w/2)": 768, + "batch_size": 2, + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 928, + "floor(h/2)": 24, "floor(w/2)": 32, + "h": 48 + }, + { + "w": 80, + "floor(h/2)*floor(w/2)": 1200, "batch_size": 2, - "w": 64 + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1360, + "floor(h/2)": 30, + "floor(w/2)": 40, + "h": 60 }, { - "floor(h/2)*floor(w/2)": 2304, + "w": 96, + "floor(h/2)*floor(w/2)": 1728, + "batch_size": 2, + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1888, "floor(h/2)": 36, - "max_length + floor(h/2)*floor(w/2)": 2464, + "floor(w/2)": 48, + "h": 72 + }, + { + "w": 128, + "floor(h/2)*floor(w/2)": 3072, + "batch_size": 2, "max_length": 160, - "h": 72, + "max_length + floor(h/2)*floor(w/2)": 3232, + "floor(h/2)": 48, "floor(w/2)": 64, + "h": 96 + }, + { + "w": 36, + "floor(h/2)*floor(w/2)": 576, "batch_size": 2, - "w": 128 + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 736, + "floor(h/2)": 32, + "floor(w/2)": 18, + "h": 64 + }, + { + "w": 48, + "floor(h/2)*floor(w/2)": 960, + "batch_size": 2, + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1120, + "floor(h/2)": 40, + "floor(w/2)": 24, + "h": 80 + }, + { + "w": 56, + "floor(h/2)*floor(w/2)": 1344, + "batch_size": 2, + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1504, + "floor(h/2)": 48, + "floor(w/2)": 28, + "h": 96 }, { + "w": 72, "floor(h/2)*floor(w/2)": 2304, - "floor(h/2)": 64, - "max_length + floor(h/2)*floor(w/2)": 2464, + "batch_size": 2, "max_length": 160, - "h": 128, + "max_length + floor(h/2)*floor(w/2)": 2464, + "floor(h/2)": 64, "floor(w/2)": 36, - "batch_size": 2, - "w": 72 + "h": 128 }, { - "floor(h/2)*floor(w/2)": 3072, - "floor(h/2)": 48, - "max_length + floor(h/2)*floor(w/2)": 3232, + "w": 64, + "floor(h/2)*floor(w/2)": 576, + "batch_size": 2, "max_length": 160, - "h": 96, - "floor(w/2)": 64, + "max_length + floor(h/2)*floor(w/2)": 736, + "floor(h/2)": 18, + "floor(w/2)": 32, + "h": 36 + }, + { + "w": 80, + "floor(h/2)*floor(w/2)": 960, "batch_size": 2, - "w": 128 + "max_length": 160, + "max_length + floor(h/2)*floor(w/2)": 1120, + "floor(h/2)": 24, + "floor(w/2)": 40, + "h": 48 }, { - "floor(h/2)*floor(w/2)": 3072, - "floor(h/2)": 64, - "max_length + floor(h/2)*floor(w/2)": 3232, + "w": 96, + "floor(h/2)*floor(w/2)": 1344, + "batch_size": 2, "max_length": 160, - "h": 128, + "max_length + floor(h/2)*floor(w/2)": 1504, + "floor(h/2)": 28, "floor(w/2)": 48, - "batch_size": 2, - "w": 96 + "h": 56 }, { - "floor(h/2)*floor(w/2)": 4096, - "floor(h/2)": 64, - "max_length + floor(h/2)*floor(w/2)": 4256, + "w": 128, + "floor(h/2)*floor(w/2)": 2304, + "batch_size": 2, "max_length": 160, - "h": 128, + "max_length + floor(h/2)*floor(w/2)": 2464, + "floor(h/2)": 36, "floor(w/2)": 64, - "batch_size": 2, - "w": 128 + "h": 72 } ], "aux_info": {} diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..2a580fbc555c82441fefa050e30b00719de0608a --- /dev/null +++ b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:442f8519fd1d1db6c01a6fed2615b9a6a46d4d2c3a5efc0062b93d54b9202ceb +size 607605147 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state index e17b97296528743c8e16dcd0fe77804688d06d81..003f6f32e40b85f3a335457ccb7f142c16eaa836 100644 Binary files a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state and b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 1cd68c968b9251ce8eac5df6973a3634f478caf6..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:3f94fdc43a513f1268b7a97d036c4745120b77760c15a6440a030b428a2e6732 -size 44376912 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta deleted file mode 100644 index 758f2b7ec50b163695369ad750e4f213eb5f1f50..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:1e063714046f3301c6bd0e6bcab260c1e348bf081a8e2f6f4b386386b10650c9 -size 6965646 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt deleted file mode 100644 index fbc488e06e58d587ddbe95ac304a43b9672e039c..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:e76618f00493c0f8ccc5a822fb2c9b3f053e7bdfd685486ea40dfc62ab43c2a1 -size 28207344 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta deleted file mode 100644 index 185ca30a49e5dd196a965ec92fc4ccc264e650c9..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:3bbf2443801834a30c7a8a460238b76e05290f16c934222c057acff10b0e7a58 -size 4252364 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index f4f9c18ff0d9440a691aa6b202e5278296ce2ffb..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:838a135e57ed85c5159646a9c75e91d1cd4c5ee6325b5d7a3f260fe847e9e310 -size 35758704 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta deleted file mode 100644 index 993580f3ecec39aaf635647e8b12ac0014fa6fb5..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c6260fd40b4525d0a742973f97c68ab310808eb376b20192298ccce065d2296b -size 5549773 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index 65909b06a92e9f4a5475a3f83e182ef6a47125f8..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:cc415bb65aff52bf3abaab0ce294d30c180bde5e423873d736123b0450ea46d7 -size 17375184 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta deleted file mode 100644 index b18da738f52165e45e94d11fd4f75a74254cda0c..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:feaf825ab41c34908237161b4ac52cd0cf98ed10bcb3fb71c39eea306e21d1ab -size 2453156 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index 70877fbc8bee4e4f550b87bf47c598422af3019c..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:39ee8075a880e02c8ca14caf2577db95d1d575c686c5724edafcb32da4487979 -size 21936336 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta deleted file mode 100644 index 2413aea432939bb71424ccdf0766f2e2ff826d4b..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:79e5d2e0bdaae7d422d7d29d6b4747d94432754984b5a394955cc10d6ab539a8 -size 3172016 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 51ed659a665e0a73090a2e04c6cc05f9bbae1c7c..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a70cfe7ea32c5b54e7dc323b7e8a69682fd2db68d51252c79dcc528e93cde4aa -size 28093680 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta deleted file mode 100644 index e8b22ebbf9a47ec1efceb1419d5730844a19bbeb..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:661cb4ab784e639ba1788764ecf1fe44fa0de862833feb423aa5939134002b43 -size 4246470 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 3269ad04ae314c9e26aa80479927e1317cd75e0d..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:3de1696ff7a340201a210a6ff2b834c8650ddca0a227dd43bb05e88fe2f434e2 -size 35708592 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta deleted file mode 100644 index 2272eaa4811520b39d3f60639fd2d2b92dffae77..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a6974462bac6c19cbf8e1c30a154c971c397ffb059c71423b902b9990bb00a01 -size 5562534 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index 5e6651bb20823b32e68ff822a1b8d05686bf0f5a..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ae6e94e05882935a8a881566a1d72bb162d35146c750e2be29bdf1e3bf41f1ea -size 21936336 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta deleted file mode 100644 index ac27b86992fd65cb10b1ae7c424811e15ad5ad22..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:b88326629f17106047c9c05670fb119a7eea247a7552f4db015b58960e48c702 -size 3172016 diff --git a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super b/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-canny/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-canny/dynamic/dd/replaced.onnx b/normal/controlnet-canny/dynamic/dd/replaced.onnx index ff0b78c9f208999cb25f6a118b41575589284a9e..e84c17c4fbebe8f5d410c4029edfa36eb032145a 100644 --- a/normal/controlnet-canny/dynamic/dd/replaced.onnx +++ b/normal/controlnet-canny/dynamic/dd/replaced.onnx @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2d71a794b89f837411a5040121b729334873cf53a977ac9355cb58b9b31d7b53 -size 256940 +oid sha256:645a41d8e5dc9f68a06c1ca66f7daa3f1c7089849d466f93e2eece530b81ef5d +size 257046 diff --git a/normal/controlnet-depth/dynamic/dd/cache/pos_embedprojConv_meta.json b/normal/controlnet-depth/dynamic/dd/cache/pos_embedprojConv_meta.json index 5293628b776fb0a7d9c1a0a20e4e61322386cb18..10f3f83edb10f96b5821bb21ecbf1f74d74ab8c7 100644 --- a/normal/controlnet-depth/dynamic/dd/cache/pos_embedprojConv_meta.json +++ b/normal/controlnet-depth/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -341,13 +341,13 @@ "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "/time_text_embed/Cast_output_0.out17_3_3" + "/time_text_embed/Cast_output_0.out17_4_3" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2" + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "attrs": { "input_shape": { @@ -405,7 +405,7 @@ "name": "/time_text_embed/timestep_embedder/act/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "const_args": [ "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1" @@ -461,10 +461,10 @@ "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3" + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4" ], "attrs": { "input_shape": { @@ -522,13 +522,13 @@ "name": "/time_text_embed/text_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "pooled_projections.out17_3_1" + "pooled_projections.out17_4_1" ], "const_args": [ - "time_text_embed.text_embedder.linear_1.weight_5_1_0" + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0" ], "out_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "attrs": { "input_shape": { @@ -586,7 +586,7 @@ "name": "/time_text_embed/text_embedder/act_1/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "const_args": [ "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0" @@ -642,10 +642,10 @@ "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" ], "const_args": [ - "time_text_embed.text_embedder.linear_2.weight_5_1_1" + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1" ], "out_args": [ - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "attrs": { "input_shape": { @@ -703,8 +703,8 @@ "name": "/time_text_embed/Add", "type": "SDAdd", "in_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "const_args": [], "out_args": [ @@ -813,16 +813,16 @@ } }, { - "name": "encoder_hidden_states.out17_3_0_SDCastBf2Bfp", + "name": "encoder_hidden_states.out17_4_0_SDCastBf2Bfp", "type": "SDCastBf2Bfp", "in_args": [ - "encoder_hidden_states.out17_3_0" + "encoder_hidden_states.out17_4_0" ], "const_args": [ - "encoder_hidden_states.out17_3_0_bfp.wts" + "encoder_hidden_states.out17_4_0_bfp.wts" ], "out_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -865,13 +865,13 @@ "name": "/context_embedder/MatMul", "type": "SDGemm_bfp", "in_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "const_args": [ - "onnx::MatMul_3779" + "onnx::MatMul_3779_45_1_0" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -926,16 +926,16 @@ } }, { - "name": "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_SDCastBfp2Bf", + "name": "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_SDCastBfp2Bf", "type": "SDCastBfp2Bf", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "const_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "attrs": { "input_shape": { @@ -978,11 +978,11 @@ "name": "/transformer_blocks.0/norm1_context/norm/LayerNormalization", "type": "SDLayerNorm_bfbfp", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0" ], "out_args": [ "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" @@ -1142,8 +1142,8 @@ "/Add_output_0.out_35_1_3" ], "const_args": [ - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1" ], "out_args": [ "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" @@ -1342,12 +1342,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -1380,6 +1374,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -1448,12 +1448,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -1486,6 +1480,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -1548,12 +1548,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -1586,6 +1580,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -1688,7 +1688,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_3800" + "onnx::MatMul_3800_45_5_1" ], "out_args": [ "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1" @@ -1783,7 +1783,7 @@ "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_2_gma" @@ -1867,8 +1867,8 @@ "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0" + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3" ], "out_args": [ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" @@ -2028,7 +2028,7 @@ "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" ], "const_args": [ - "onnx::MatMul_3803" + "onnx::MatMul_3803_45_1_2" ], "out_args": [ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" @@ -2098,10 +2098,10 @@ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" ], "const_args": [ - "onnx::MatMul_3804" + "onnx::MatMul_3804_45_1_4" ], "out_args": [ - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8" + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8" ], "attrs": { "input_shape": { @@ -2160,7 +2160,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ @@ -2245,7 +2245,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_3799" + "onnx::MatMul_3799_45_5_0" ], "out_args": [ "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0" @@ -2424,8 +2424,8 @@ "/transformer_blocks.0/Add_output_0.out10_44" ], "const_args": [ - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0" + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2" ], "out_args": [ "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" @@ -2585,7 +2585,7 @@ "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_46" ], "const_args": [ - "onnx::MatMul_3801" + "onnx::MatMul_3801_45_1_1" ], "out_args": [ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" @@ -2655,10 +2655,10 @@ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" ], "const_args": [ - "onnx::MatMul_3802" + "onnx::MatMul_3802_45_1_3" ], "out_args": [ - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7" + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7" ], "attrs": { "input_shape": { @@ -2717,7 +2717,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_output_0.out10_44" ], "const_args": [ @@ -2802,8 +2802,8 @@ "/transformer_blocks.0/Add_3_output_0.out10_45" ], "const_args": [ - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4" ], "out_args": [ "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" @@ -2963,10 +2963,10 @@ "/transformer_blocks.0/Add_3_output_0.out10_45" ], "const_args": [ - "onnx::MatMul_4074" + "onnx::MatMul_4074_45_0_4" ], "out_args": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13" + "/controlnet_blocks.0/Add_output_0.out17_4_13" ], "attrs": { "input_shape": { @@ -3027,8 +3027,8 @@ "/transformer_blocks.0/Add_7_output_0.out10_1" ], "const_args": [ - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5" ], "out_args": [ "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" @@ -3227,12 +3227,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -3265,6 +3259,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -3333,12 +3333,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -3371,6 +3365,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -3433,12 +3433,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -3471,6 +3465,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -3573,7 +3573,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_3824" + "onnx::MatMul_3824_45_5_2" ], "out_args": [ "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2" @@ -3752,8 +3752,8 @@ "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0" + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6" ], "out_args": [ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" @@ -3913,7 +3913,7 @@ "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" ], "const_args": [ - "onnx::MatMul_3826" + "onnx::MatMul_3826_45_1_5" ], "out_args": [ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" @@ -3983,10 +3983,10 @@ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" ], "const_args": [ - "onnx::MatMul_3827" + "onnx::MatMul_3827_45_1_7" ], "out_args": [ - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12" + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12" ], "attrs": { "input_shape": { @@ -4045,7 +4045,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ @@ -4130,10 +4130,10 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "onnx::MatMul_4075" + "onnx::MatMul_4075_45_0_5" ], "out_args": [ - "/controlnet_blocks.1/Add_output_0.out17_3_22" + "/controlnet_blocks.1/Add_output_0.out17_4_22" ], "attrs": { "input_shape": { @@ -4194,7 +4194,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_3825" + "onnx::MatMul_3825_45_5_3" ], "out_args": [ "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3" @@ -4373,8 +4373,8 @@ "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0" + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7" ], "out_args": [ "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" @@ -4534,7 +4534,7 @@ "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" ], "const_args": [ - "onnx::MatMul_3828" + "onnx::MatMul_3828_45_1_6" ], "out_args": [ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" @@ -4604,10 +4604,10 @@ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" ], "const_args": [ - "onnx::MatMul_3829" + "onnx::MatMul_3829_45_1_8" ], "out_args": [ - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13" + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13" ], "attrs": { "input_shape": { @@ -4666,7 +4666,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ @@ -4751,8 +4751,8 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8" ], "out_args": [ "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" @@ -4832,8 +4832,8 @@ "/transformer_blocks.1/Add_7_output_0.out10_5" ], "const_args": [ - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9" ], "out_args": [ "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" @@ -5112,12 +5112,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -5150,6 +5144,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -5218,12 +5218,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -5256,6 +5250,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -5318,12 +5318,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -5356,6 +5350,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -5458,7 +5458,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_3850" + "onnx::MatMul_3850_45_5_5" ], "out_args": [ "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5" @@ -5637,7 +5637,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_3849" + "onnx::MatMul_3849_45_5_4" ], "out_args": [ "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4" @@ -5816,8 +5816,8 @@ "/transformer_blocks.2/Add_output_0.out10_12" ], "const_args": [ - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0" + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10" ], "out_args": [ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" @@ -5977,7 +5977,7 @@ "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_14" ], "const_args": [ - "onnx::MatMul_3851" + "onnx::MatMul_3851_45_1_9" ], "out_args": [ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" @@ -6047,10 +6047,10 @@ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" ], "const_args": [ - "onnx::MatMul_3852" + "onnx::MatMul_3852_45_1_11" ], "out_args": [ - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17" + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17" ], "attrs": { "input_shape": { @@ -6109,7 +6109,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_output_0.out10_12" ], "const_args": [ @@ -6194,10 +6194,10 @@ "/transformer_blocks.2/Add_3_output_0.out10_13" ], "const_args": [ - "onnx::MatMul_4076" + "onnx::MatMul_4076_45_0_6" ], "out_args": [ - "/controlnet_blocks.2/Add_output_0.out17_3_31" + "/controlnet_blocks.2/Add_output_0.out17_4_31" ], "attrs": { "input_shape": { @@ -6258,8 +6258,8 @@ "/transformer_blocks.2/Add_4_output_0.out10_14" ], "const_args": [ - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0" + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11" ], "out_args": [ "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" @@ -6419,7 +6419,7 @@ "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_16" ], "const_args": [ - "onnx::MatMul_3853" + "onnx::MatMul_3853_45_1_10" ], "out_args": [ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" @@ -6489,10 +6489,10 @@ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" ], "const_args": [ - "onnx::MatMul_3854" + "onnx::MatMul_3854_45_1_12" ], "out_args": [ - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18" + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18" ], "attrs": { "input_shape": { @@ -6551,7 +6551,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_4_output_0.out10_14" ], "const_args": [ @@ -6636,8 +6636,8 @@ "/transformer_blocks.2/Add_3_output_0.out10_13" ], "const_args": [ - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12" ], "out_args": [ "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" @@ -6717,8 +6717,8 @@ "/transformer_blocks.2/Add_7_output_0.out10_15" ], "const_args": [ - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13" ], "out_args": [ "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" @@ -6997,12 +6997,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -7035,6 +7029,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -7103,12 +7103,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -7141,6 +7135,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -7203,12 +7203,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -7241,6 +7235,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -7343,7 +7343,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_3875" + "onnx::MatMul_3875_45_5_7" ], "out_args": [ "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7" @@ -7522,7 +7522,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_3874" + "onnx::MatMul_3874_45_5_6" ], "out_args": [ "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6" @@ -7701,8 +7701,8 @@ "/transformer_blocks.3/Add_output_0.out10_16" ], "const_args": [ - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0" + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14" ], "out_args": [ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" @@ -7862,7 +7862,7 @@ "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_18" ], "const_args": [ - "onnx::MatMul_3876" + "onnx::MatMul_3876_45_1_13" ], "out_args": [ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" @@ -7932,10 +7932,10 @@ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" ], "const_args": [ - "onnx::MatMul_3877" + "onnx::MatMul_3877_45_1_15" ], "out_args": [ - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22" + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22" ], "attrs": { "input_shape": { @@ -7994,7 +7994,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_output_0.out10_16" ], "const_args": [ @@ -8079,10 +8079,10 @@ "/transformer_blocks.3/Add_3_output_0.out10_17" ], "const_args": [ - "onnx::MatMul_4077" + "onnx::MatMul_4077_45_0_7" ], "out_args": [ - "/controlnet_blocks.3/Add_output_0.out17_3_40" + "/controlnet_blocks.3/Add_output_0.out17_4_40" ], "attrs": { "input_shape": { @@ -8143,8 +8143,8 @@ "/transformer_blocks.3/Add_4_output_0.out10_18" ], "const_args": [ - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0" + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15" ], "out_args": [ "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" @@ -8304,7 +8304,7 @@ "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_20" ], "const_args": [ - "onnx::MatMul_3878" + "onnx::MatMul_3878_45_1_14" ], "out_args": [ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" @@ -8374,10 +8374,10 @@ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" ], "const_args": [ - "onnx::MatMul_3879" + "onnx::MatMul_3879_45_1_16" ], "out_args": [ - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23" + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23" ], "attrs": { "input_shape": { @@ -8436,7 +8436,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_4_output_0.out10_18" ], "const_args": [ @@ -8521,8 +8521,8 @@ "/transformer_blocks.3/Add_3_output_0.out10_17" ], "const_args": [ - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16" ], "out_args": [ "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" @@ -8602,8 +8602,8 @@ "/transformer_blocks.3/Add_7_output_0.out10_19" ], "const_args": [ - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17" ], "out_args": [ "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" @@ -8882,12 +8882,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -8920,6 +8914,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -8988,12 +8988,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -9026,6 +9020,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -9088,12 +9088,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -9126,6 +9120,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -9228,7 +9228,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_3900" + "onnx::MatMul_3900_45_5_9" ], "out_args": [ "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9" @@ -9407,7 +9407,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_3899" + "onnx::MatMul_3899_45_5_8" ], "out_args": [ "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8" @@ -9586,8 +9586,8 @@ "/transformer_blocks.4/Add_output_0.out10_20" ], "const_args": [ - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0" + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18" ], "out_args": [ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" @@ -9747,7 +9747,7 @@ "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_22" ], "const_args": [ - "onnx::MatMul_3901" + "onnx::MatMul_3901_45_1_17" ], "out_args": [ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" @@ -9817,10 +9817,10 @@ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" ], "const_args": [ - "onnx::MatMul_3902" + "onnx::MatMul_3902_45_1_19" ], "out_args": [ - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27" + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27" ], "attrs": { "input_shape": { @@ -9879,7 +9879,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_output_0.out10_20" ], "const_args": [ @@ -9964,10 +9964,10 @@ "/transformer_blocks.4/Add_3_output_0.out10_21" ], "const_args": [ - "onnx::MatMul_4078" + "onnx::MatMul_4078_45_0_8" ], "out_args": [ - "/controlnet_blocks.4/Add_output_0.out17_3_49" + "/controlnet_blocks.4/Add_output_0.out17_4_49" ], "attrs": { "input_shape": { @@ -10028,8 +10028,8 @@ "/transformer_blocks.4/Add_4_output_0.out10_22" ], "const_args": [ - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0" + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19" ], "out_args": [ "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" @@ -10189,7 +10189,7 @@ "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_24" ], "const_args": [ - "onnx::MatMul_3903" + "onnx::MatMul_3903_45_1_18" ], "out_args": [ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" @@ -10259,10 +10259,10 @@ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" ], "const_args": [ - "onnx::MatMul_3904" + "onnx::MatMul_3904_45_1_20" ], "out_args": [ - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28" + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28" ], "attrs": { "input_shape": { @@ -10321,7 +10321,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_4_output_0.out10_22" ], "const_args": [ @@ -10406,8 +10406,8 @@ "/transformer_blocks.4/Add_3_output_0.out10_21" ], "const_args": [ - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20" ], "out_args": [ "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" @@ -10487,8 +10487,8 @@ "/transformer_blocks.4/Add_7_output_0.out10_23" ], "const_args": [ - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21" ], "out_args": [ "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" @@ -10767,12 +10767,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -10805,6 +10799,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -10873,12 +10873,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -10911,6 +10905,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -10973,12 +10973,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -11011,6 +11005,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -11113,7 +11113,7 @@ "/transformer_blocks.5/attn/Reshape_3_output_0.out22_5_bfp.out27_0_5" ], "const_args": [ - "onnx::MatMul_3925" + "onnx::MatMul_3925_45_5_11" ], "out_args": [ "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11" @@ -11292,7 +11292,7 @@ "/transformer_blocks.5/attn/Reshape_3_output_0.out22_5_bfp.out27_0_5" ], "const_args": [ - "onnx::MatMul_3924" + "onnx::MatMul_3924_45_5_10" ], "out_args": [ "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10" @@ -11471,8 +11471,8 @@ "/transformer_blocks.5/Add_output_0.out10_24" ], "const_args": [ - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0" + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22" ], "out_args": [ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" @@ -11632,7 +11632,7 @@ "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_26" ], "const_args": [ - "onnx::MatMul_3926" + "onnx::MatMul_3926_45_1_21" ], "out_args": [ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" @@ -11702,10 +11702,10 @@ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" ], "const_args": [ - "onnx::MatMul_3927" + "onnx::MatMul_3927_45_1_23" ], "out_args": [ - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_32" + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32" ], "attrs": { "input_shape": { @@ -11764,7 +11764,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_32", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32", "/transformer_blocks.5/Add_output_0.out10_24" ], "const_args": [ @@ -11849,10 +11849,10 @@ "/transformer_blocks.5/Add_3_output_0.out10_25" ], "const_args": [ - "onnx::MatMul_4079" + "onnx::MatMul_4079_45_0_9" ], "out_args": [ - "/controlnet_blocks.5/Add_output_0.out17_3_58" + "/controlnet_blocks.5/Add_output_0.out17_4_58" ], "attrs": { "input_shape": { @@ -11913,8 +11913,8 @@ "/transformer_blocks.5/Add_4_output_0.out10_26" ], "const_args": [ - "/transformer_blocks.5/norm2_context/Constant_output_0", - "/transformer_blocks.5/norm2_context/Constant_1_output_0" + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23" ], "out_args": [ "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23" @@ -12074,7 +12074,7 @@ "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_28" ], "const_args": [ - "onnx::MatMul_3928" + "onnx::MatMul_3928_45_1_22" ], "out_args": [ "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_31" @@ -12144,10 +12144,10 @@ "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_31" ], "const_args": [ - "onnx::MatMul_3929" + "onnx::MatMul_3929_45_1_24" ], "out_args": [ - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_57_bfp.out25_33" + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33" ], "attrs": { "input_shape": { @@ -12206,7 +12206,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_57_bfp.out25_33", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33", "/transformer_blocks.5/Add_4_output_0.out10_26" ], "const_args": [ @@ -12291,8 +12291,8 @@ "/transformer_blocks.5/Add_3_output_0.out10_25" ], "const_args": [ - "/transformer_blocks.6/norm1/norm/Constant_output_0", - "/transformer_blocks.6/norm1/norm/Constant_1_output_0" + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_24" ], "out_args": [ "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_24_bfp.out15_24" @@ -12372,8 +12372,8 @@ "/transformer_blocks.5/Add_7_output_0.out10_27" ], "const_args": [ - "/transformer_blocks.6/norm1_context/norm/Constant_output_0", - "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_25" ], "out_args": [ "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_25_bfp.out15_25" @@ -12652,12 +12652,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -12690,6 +12684,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -12758,12 +12758,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -12796,6 +12790,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -12858,12 +12858,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -12896,6 +12890,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -12998,7 +12998,7 @@ "/transformer_blocks.6/attn/Reshape_3_output_0.out22_6_bfp.out27_0_6" ], "const_args": [ - "onnx::MatMul_3950" + "onnx::MatMul_3950_45_5_13" ], "out_args": [ "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13" @@ -13177,7 +13177,7 @@ "/transformer_blocks.6/attn/Reshape_3_output_0.out22_6_bfp.out27_0_6" ], "const_args": [ - "onnx::MatMul_3949" + "onnx::MatMul_3949_45_5_12" ], "out_args": [ "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12" @@ -13356,8 +13356,8 @@ "/transformer_blocks.6/Add_output_0.out10_28" ], "const_args": [ - "/transformer_blocks.6/norm2/Constant_output_0", - "/transformer_blocks.6/norm2/Constant_1_output_0" + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26" ], "out_args": [ "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26" @@ -13517,7 +13517,7 @@ "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_30" ], "const_args": [ - "onnx::MatMul_3951" + "onnx::MatMul_3951_45_1_25" ], "out_args": [ "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_35" @@ -13587,10 +13587,10 @@ "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_35" ], "const_args": [ - "onnx::MatMul_3952" + "onnx::MatMul_3952_45_1_27" ], "out_args": [ - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_65_bfp.out25_37" + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37" ], "attrs": { "input_shape": { @@ -13649,7 +13649,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_65_bfp.out25_37", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37", "/transformer_blocks.6/Add_output_0.out10_28" ], "const_args": [ @@ -13734,10 +13734,10 @@ "/transformer_blocks.6/Add_3_output_0.out10_29" ], "const_args": [ - "onnx::MatMul_4080" + "onnx::MatMul_4080_45_0_10" ], "out_args": [ - "/controlnet_blocks.6/Add_output_0.out17_3_67" + "/controlnet_blocks.6/Add_output_0.out17_4_67" ], "attrs": { "input_shape": { @@ -13798,8 +13798,8 @@ "/transformer_blocks.6/Add_4_output_0.out10_30" ], "const_args": [ - "/transformer_blocks.6/norm2_context/Constant_output_0", - "/transformer_blocks.6/norm2_context/Constant_1_output_0" + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27" ], "out_args": [ "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27" @@ -13959,7 +13959,7 @@ "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_32" ], "const_args": [ - "onnx::MatMul_3953" + "onnx::MatMul_3953_45_1_26" ], "out_args": [ "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_36" @@ -14029,10 +14029,10 @@ "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_36" ], "const_args": [ - "onnx::MatMul_3954" + "onnx::MatMul_3954_45_1_28" ], "out_args": [ - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_66_bfp.out25_38" + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38" ], "attrs": { "input_shape": { @@ -14091,7 +14091,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_66_bfp.out25_38", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38", "/transformer_blocks.6/Add_4_output_0.out10_30" ], "const_args": [ @@ -14176,8 +14176,8 @@ "/transformer_blocks.6/Add_3_output_0.out10_29" ], "const_args": [ - "/transformer_blocks.7/norm1/norm/Constant_output_0", - "/transformer_blocks.7/norm1/norm/Constant_1_output_0" + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_28" ], "out_args": [ "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_28_bfp.out15_28" @@ -14257,8 +14257,8 @@ "/transformer_blocks.6/Add_7_output_0.out10_31" ], "const_args": [ - "/transformer_blocks.7/norm1_context/norm/Constant_output_0", - "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_29" ], "out_args": [ "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_29_bfp.out15_29" @@ -14537,12 +14537,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -14575,6 +14569,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -14643,12 +14643,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -14676,9 +14670,15 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "batch_size", + "floor(h/2)*floor(w/2)", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" ] }, "in_dtypes": { @@ -14743,12 +14743,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -14781,6 +14775,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -14883,7 +14883,7 @@ "/transformer_blocks.7/attn/Reshape_3_output_0.out22_7_bfp.out27_0_7" ], "const_args": [ - "onnx::MatMul_3975" + "onnx::MatMul_3975_45_5_15" ], "out_args": [ "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15" @@ -15062,7 +15062,7 @@ "/transformer_blocks.7/attn/Reshape_3_output_0.out22_7_bfp.out27_0_7" ], "const_args": [ - "onnx::MatMul_3974" + "onnx::MatMul_3974_45_5_14" ], "out_args": [ "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14" @@ -15241,8 +15241,8 @@ "/transformer_blocks.7/Add_output_0.out10_32" ], "const_args": [ - "/transformer_blocks.7/norm2/Constant_output_0", - "/transformer_blocks.7/norm2/Constant_1_output_0" + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30" ], "out_args": [ "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30" @@ -15402,7 +15402,7 @@ "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_34" ], "const_args": [ - "onnx::MatMul_3976" + "onnx::MatMul_3976_45_1_29" ], "out_args": [ "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_40" @@ -15472,10 +15472,10 @@ "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_40" ], "const_args": [ - "onnx::MatMul_3977" + "onnx::MatMul_3977_45_1_31" ], "out_args": [ - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_74_bfp.out25_42" + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42" ], "attrs": { "input_shape": { @@ -15534,7 +15534,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_74_bfp.out25_42", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42", "/transformer_blocks.7/Add_output_0.out10_32" ], "const_args": [ @@ -15619,10 +15619,10 @@ "/transformer_blocks.7/Add_3_output_0.out10_33" ], "const_args": [ - "onnx::MatMul_4081" + "onnx::MatMul_4081_45_0_11" ], "out_args": [ - "/controlnet_blocks.7/Add_output_0.out17_3_76" + "/controlnet_blocks.7/Add_output_0.out17_4_76" ], "attrs": { "input_shape": { @@ -15683,8 +15683,8 @@ "/transformer_blocks.7/Add_4_output_0.out10_34" ], "const_args": [ - "/transformer_blocks.7/norm2_context/Constant_output_0", - "/transformer_blocks.7/norm2_context/Constant_1_output_0" + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31" ], "out_args": [ "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31" @@ -15844,7 +15844,7 @@ "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_36" ], "const_args": [ - "onnx::MatMul_3978" + "onnx::MatMul_3978_45_1_30" ], "out_args": [ "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_41" @@ -15914,10 +15914,10 @@ "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_41" ], "const_args": [ - "onnx::MatMul_3979" + "onnx::MatMul_3979_45_1_32" ], "out_args": [ - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_75_bfp.out25_43" + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43" ], "attrs": { "input_shape": { @@ -15976,7 +15976,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_75_bfp.out25_43", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43", "/transformer_blocks.7/Add_4_output_0.out10_34" ], "const_args": [ @@ -16061,8 +16061,8 @@ "/transformer_blocks.7/Add_3_output_0.out10_33" ], "const_args": [ - "/transformer_blocks.8/norm1/norm/Constant_output_0", - "/transformer_blocks.8/norm1/norm/Constant_1_output_0" + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_32" ], "out_args": [ "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_32_bfp.out15_32" @@ -16142,8 +16142,8 @@ "/transformer_blocks.7/Add_7_output_0.out10_35" ], "const_args": [ - "/transformer_blocks.8/norm1_context/norm/Constant_output_0", - "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_33" ], "out_args": [ "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_33_bfp.out15_33" @@ -16422,12 +16422,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -16460,6 +16454,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -16528,12 +16528,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -16566,6 +16560,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -16628,12 +16628,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -16666,6 +16660,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -16768,7 +16768,7 @@ "/transformer_blocks.8/attn/Reshape_3_output_0.out22_8_bfp.out27_0_8" ], "const_args": [ - "onnx::MatMul_4000" + "onnx::MatMul_4000_45_5_17" ], "out_args": [ "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17" @@ -16947,7 +16947,7 @@ "/transformer_blocks.8/attn/Reshape_3_output_0.out22_8_bfp.out27_0_8" ], "const_args": [ - "onnx::MatMul_3999" + "onnx::MatMul_3999_45_5_16" ], "out_args": [ "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16" @@ -17126,8 +17126,8 @@ "/transformer_blocks.8/Add_output_0.out10_36" ], "const_args": [ - "/transformer_blocks.8/norm2/Constant_output_0", - "/transformer_blocks.8/norm2/Constant_1_output_0" + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34" ], "out_args": [ "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34" @@ -17287,7 +17287,7 @@ "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_38" ], "const_args": [ - "onnx::MatMul_4001" + "onnx::MatMul_4001_45_1_33" ], "out_args": [ "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_45" @@ -17357,10 +17357,10 @@ "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_45" ], "const_args": [ - "onnx::MatMul_4002" + "onnx::MatMul_4002_45_1_35" ], "out_args": [ - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_83_bfp.out25_47" + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47" ], "attrs": { "input_shape": { @@ -17419,7 +17419,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_83_bfp.out25_47", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47", "/transformer_blocks.8/Add_output_0.out10_36" ], "const_args": [ @@ -17504,10 +17504,10 @@ "/transformer_blocks.8/Add_3_output_0.out10_37" ], "const_args": [ - "onnx::MatMul_4082" + "onnx::MatMul_4082_45_0_12" ], "out_args": [ - "/controlnet_blocks.8/Add_output_0.out17_3_85" + "/controlnet_blocks.8/Add_output_0.out17_4_85" ], "attrs": { "input_shape": { @@ -17568,8 +17568,8 @@ "/transformer_blocks.8/Add_4_output_0.out10_38" ], "const_args": [ - "/transformer_blocks.8/norm2_context/Constant_output_0", - "/transformer_blocks.8/norm2_context/Constant_1_output_0" + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35" ], "out_args": [ "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35" @@ -17729,7 +17729,7 @@ "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_40" ], "const_args": [ - "onnx::MatMul_4003" + "onnx::MatMul_4003_45_1_34" ], "out_args": [ "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_46" @@ -17799,10 +17799,10 @@ "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_46" ], "const_args": [ - "onnx::MatMul_4004" + "onnx::MatMul_4004_45_1_36" ], "out_args": [ - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_48" + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48" ], "attrs": { "input_shape": { @@ -17861,7 +17861,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_48", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48", "/transformer_blocks.8/Add_4_output_0.out10_38" ], "const_args": [ @@ -17946,8 +17946,8 @@ "/transformer_blocks.8/Add_3_output_0.out10_37" ], "const_args": [ - "/transformer_blocks.9/norm1/norm/Constant_output_0", - "/transformer_blocks.9/norm1/norm/Constant_1_output_0" + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_36" ], "out_args": [ "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_36_bfp.out15_36" @@ -18027,8 +18027,8 @@ "/transformer_blocks.8/Add_7_output_0.out10_39" ], "const_args": [ - "/transformer_blocks.9/norm1_context/norm/Constant_output_0", - "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_37" ], "out_args": [ "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_37_bfp.out15_37" @@ -18307,12 +18307,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -18345,6 +18339,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -18413,12 +18413,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -18451,6 +18445,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -18513,12 +18513,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -18551,6 +18545,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -18653,7 +18653,7 @@ "/transformer_blocks.9/attn/Reshape_3_output_0.out22_9_bfp.out27_0_9" ], "const_args": [ - "onnx::MatMul_4025" + "onnx::MatMul_4025_45_5_19" ], "out_args": [ "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19" @@ -18832,7 +18832,7 @@ "/transformer_blocks.9/attn/Reshape_3_output_0.out22_9_bfp.out27_0_9" ], "const_args": [ - "onnx::MatMul_4024" + "onnx::MatMul_4024_45_5_18" ], "out_args": [ "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18" @@ -19011,8 +19011,8 @@ "/transformer_blocks.9/Add_output_0.out10_40" ], "const_args": [ - "/transformer_blocks.9/norm2/Constant_output_0", - "/transformer_blocks.9/norm2/Constant_1_output_0" + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38" ], "out_args": [ "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38" @@ -19172,7 +19172,7 @@ "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_42" ], "const_args": [ - "onnx::MatMul_4026" + "onnx::MatMul_4026_45_1_37" ], "out_args": [ "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_50" @@ -19242,10 +19242,10 @@ "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_50" ], "const_args": [ - "onnx::MatMul_4027" + "onnx::MatMul_4027_45_1_39" ], "out_args": [ - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_92_bfp.out25_52" + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52" ], "attrs": { "input_shape": { @@ -19304,7 +19304,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_92_bfp.out25_52", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52", "/transformer_blocks.9/Add_output_0.out10_40" ], "const_args": [ @@ -19389,8 +19389,8 @@ "/transformer_blocks.9/Add_3_output_0.out10_41" ], "const_args": [ - "/transformer_blocks.10/norm1/norm/Constant_output_0", - "/transformer_blocks.10/norm1/norm/Constant_1_output_0" + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40" ], "out_args": [ "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40" @@ -19550,10 +19550,10 @@ "/transformer_blocks.9/Add_3_output_0.out10_41" ], "const_args": [ - "onnx::MatMul_4083" + "onnx::MatMul_4083_45_0_13" ], "out_args": [ - "/controlnet_blocks.9/Add_output_0.out17_3_94" + "/controlnet_blocks.9/Add_output_0.out17_4_94" ], "attrs": { "input_shape": { @@ -19614,8 +19614,8 @@ "/transformer_blocks.9/Add_4_output_0.out10_42" ], "const_args": [ - "/transformer_blocks.9/norm2_context/Constant_output_0", - "/transformer_blocks.9/norm2_context/Constant_1_output_0" + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39" ], "out_args": [ "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39" @@ -19775,7 +19775,7 @@ "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_44" ], "const_args": [ - "onnx::MatMul_4028" + "onnx::MatMul_4028_45_1_38" ], "out_args": [ "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_51" @@ -19845,10 +19845,10 @@ "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_51" ], "const_args": [ - "onnx::MatMul_4029" + "onnx::MatMul_4029_45_1_40" ], "out_args": [ - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_93_bfp.out25_53" + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53" ], "attrs": { "input_shape": { @@ -19907,7 +19907,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_93_bfp.out25_53", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53", "/transformer_blocks.9/Add_4_output_0.out10_42" ], "const_args": [ @@ -19992,8 +19992,8 @@ "/transformer_blocks.9/Add_7_output_0.out10_43" ], "const_args": [ - "/transformer_blocks.10/norm1_context/norm/Constant_output_0", - "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41" ], "out_args": [ "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41" @@ -20192,12 +20192,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -20230,6 +20224,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -20298,12 +20298,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -20336,6 +20330,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -20398,12 +20398,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -20436,6 +20430,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -20538,7 +20538,7 @@ "/transformer_blocks.10/attn/Reshape_3_output_0.out22_10_bfp.out27_0_10" ], "const_args": [ - "onnx::MatMul_4049" + "onnx::MatMul_4049_45_5_20" ], "out_args": [ "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20" @@ -20717,8 +20717,8 @@ "/transformer_blocks.10/Add_output_0.out10_6" ], "const_args": [ - "/transformer_blocks.10/norm2/Constant_output_0", - "/transformer_blocks.10/norm2/Constant_1_output_0" + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42" ], "out_args": [ "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42" @@ -20878,7 +20878,7 @@ "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7" ], "const_args": [ - "onnx::MatMul_4051" + "onnx::MatMul_4051_45_1_41" ], "out_args": [ "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_55" @@ -20948,10 +20948,10 @@ "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_55" ], "const_args": [ - "onnx::MatMul_4052" + "onnx::MatMul_4052_45_1_43" ], "out_args": [ - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_101_bfp.out25_57" + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57" ], "attrs": { "input_shape": { @@ -21010,7 +21010,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_101_bfp.out25_57", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57", "/transformer_blocks.10/Add_output_0.out10_6" ], "const_args": [ @@ -21095,10 +21095,10 @@ "/transformer_blocks.10/Add_3_output_0.out10_7" ], "const_args": [ - "onnx::MatMul_4084" + "onnx::MatMul_4084_45_0_14" ], "out_args": [ - "/controlnet_blocks.10/Add_output_0.out17_3_103" + "/controlnet_blocks.10/Add_output_0.out17_4_103" ], "attrs": { "input_shape": { @@ -21159,7 +21159,7 @@ "/transformer_blocks.10/attn/Reshape_3_output_0.out22_10_bfp.out27_0_10" ], "const_args": [ - "onnx::MatMul_4050" + "onnx::MatMul_4050_45_5_21" ], "out_args": [ "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21" @@ -21338,8 +21338,8 @@ "/transformer_blocks.10/Add_4_output_0.out10_8" ], "const_args": [ - "/transformer_blocks.10/norm2_context/Constant_output_0", - "/transformer_blocks.10/norm2_context/Constant_1_output_0" + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43" ], "out_args": [ "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43" @@ -21499,7 +21499,7 @@ "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9" ], "const_args": [ - "onnx::MatMul_4053" + "onnx::MatMul_4053_45_1_42" ], "out_args": [ "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_56" @@ -21569,10 +21569,10 @@ "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_56" ], "const_args": [ - "onnx::MatMul_4054" + "onnx::MatMul_4054_45_1_44" ], "out_args": [ - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_102_bfp.out25_58" + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58" ], "attrs": { "input_shape": { @@ -21631,7 +21631,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_102_bfp.out25_58", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58", "/transformer_blocks.10/Add_4_output_0.out10_8" ], "const_args": [ @@ -21716,8 +21716,8 @@ "/transformer_blocks.10/Add_3_output_0.out10_7" ], "const_args": [ - "/transformer_blocks.11/norm1/norm/Constant_output_0", - "/transformer_blocks.11/norm1/norm/Constant_1_output_0" + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_44" ], "out_args": [ "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_44_bfp.out15_44" @@ -21877,8 +21877,8 @@ "/transformer_blocks.10/Add_7_output_0.out10_9" ], "const_args": [ - "/transformer_blocks.11/norm1_context/norm/Constant_output_0", - "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_45" ], "out_args": [ "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_45_bfp.out15_45" @@ -22077,12 +22077,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -22115,6 +22109,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -22183,12 +22183,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -22221,6 +22215,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -22283,12 +22283,6 @@ "1536" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "output_shape_0": { "type": "str", "value": [ @@ -22321,6 +22315,12 @@ "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -22423,7 +22423,7 @@ "/transformer_blocks.11/attn/Reshape_3_output_0.out22_11_bfp.out27_0_11" ], "const_args": [ - "onnx::MatMul_4071" + "onnx::MatMul_4071_45_5_22" ], "out_args": [ "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22" @@ -22602,8 +22602,8 @@ "/transformer_blocks.11/Add_output_0.out10_10" ], "const_args": [ - "/transformer_blocks.11/norm2/Constant_output_0", - "/transformer_blocks.11/norm2/Constant_1_output_0" + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46" ], "out_args": [ "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46" @@ -22763,7 +22763,7 @@ "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11" ], "const_args": [ - "onnx::MatMul_4072" + "onnx::MatMul_4072_45_1_45" ], "out_args": [ "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_60" @@ -22833,10 +22833,10 @@ "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_60" ], "const_args": [ - "onnx::MatMul_4073" + "onnx::MatMul_4073_45_1_46" ], "out_args": [ - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_110_bfp.out25_61" + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_61" ], "attrs": { "input_shape": { @@ -22895,7 +22895,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_110_bfp.out25_61", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_61", "/transformer_blocks.11/Add_output_0.out10_10" ], "const_args": [ @@ -22980,10 +22980,10 @@ "/transformer_blocks.11/Add_3_output_0.out10_11" ], "const_args": [ - "onnx::MatMul_4085" + "onnx::MatMul_4085_45_0_15" ], "out_args": [ - "/controlnet_blocks.11/Add_output_0.out17_3_111" + "/controlnet_blocks.11/Add_output_0.out17_4_111" ], "attrs": { "input_shape": { @@ -23046,27 +23046,27 @@ "hidden_states_nhwc.out5_0_0", "/pos_embed/Reshape_1_output_0.out_35_1_2", "controlnet_cond_nhwc.out5_0_1", - "/time_text_embed/Cast_output_0.out17_3_3", - "pooled_projections.out17_3_1", - "encoder_hidden_states.out17_3_0" + "/time_text_embed/Cast_output_0.out17_4_3", + "pooled_projections.out17_4_1", + "encoder_hidden_states.out17_4_0" ] }, "out": { "buffer_size": 36864, "xrt_arg_id": 1, "packed_tensors": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13", - "/controlnet_blocks.1/Add_output_0.out17_3_22", - "/controlnet_blocks.2/Add_output_0.out17_3_31", - "/controlnet_blocks.3/Add_output_0.out17_3_40", - "/controlnet_blocks.4/Add_output_0.out17_3_49", - "/controlnet_blocks.5/Add_output_0.out17_3_58", - "/controlnet_blocks.6/Add_output_0.out17_3_67", - "/controlnet_blocks.7/Add_output_0.out17_3_76", - "/controlnet_blocks.8/Add_output_0.out17_3_85", - "/controlnet_blocks.9/Add_output_0.out17_3_94", - "/controlnet_blocks.10/Add_output_0.out17_3_103", - "/controlnet_blocks.11/Add_output_0.out17_3_111" + "/controlnet_blocks.0/Add_output_0.out17_4_13", + "/controlnet_blocks.1/Add_output_0.out17_4_22", + "/controlnet_blocks.2/Add_output_0.out17_4_31", + "/controlnet_blocks.3/Add_output_0.out17_4_40", + "/controlnet_blocks.4/Add_output_0.out17_4_49", + "/controlnet_blocks.5/Add_output_0.out17_4_58", + "/controlnet_blocks.6/Add_output_0.out17_4_67", + "/controlnet_blocks.7/Add_output_0.out17_4_76", + "/controlnet_blocks.8/Add_output_0.out17_4_85", + "/controlnet_blocks.9/Add_output_0.out17_4_94", + "/controlnet_blocks.10/Add_output_0.out17_4_103", + "/controlnet_blocks.11/Add_output_0.out17_4_111" ] }, "scratch": { @@ -23077,17 +23077,17 @@ "/pos_embed/Add_2_output_0.out_35_1_2", "/pos_embed_input/Transpose_output_0.out5_0_1", "/Add_output_0.out_35_1_3", - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3", "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1", - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1", "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2", "/time_text_embed/Add_output_0.out_35_1_4", "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "encoder_hidden_states.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0", + "encoder_hidden_states.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0", "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0", "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1", @@ -23101,14 +23101,14 @@ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3", "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1", "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_7_output_0.out10_1", "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", "/transformer_blocks.0/Add_output_0.out10_44", "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2", "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_46", "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_3_output_0.out10_45", "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4", "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2", @@ -23123,14 +23123,14 @@ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6", "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3", "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_3_output_0.out10_3", "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", "/transformer_blocks.1/Add_4_output_0.out10_4", "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7", "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5", "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_7_output_0.out10_5", "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8", "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9", @@ -23147,12 +23147,12 @@ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10", "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_14", "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_3_output_0.out10_13", "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11", "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_16", "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_7_output_0.out10_15", "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12", "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13", @@ -23169,12 +23169,12 @@ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14", "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_18", "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_3_output_0.out10_17", "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15", "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_20", "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_7_output_0.out10_19", "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16", "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17", @@ -23191,12 +23191,12 @@ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18", "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_22", "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_3_output_0.out10_21", "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19", "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_24", "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_7_output_0.out10_23", "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20", "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21", @@ -23213,12 +23213,12 @@ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22", "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_26", "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_32", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32", "/transformer_blocks.5/Add_3_output_0.out10_25", "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23", "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_28", "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_31", - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_57_bfp.out25_33", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33", "/transformer_blocks.5/Add_7_output_0.out10_27", "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_24_bfp.out15_24", "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_25_bfp.out15_25", @@ -23235,12 +23235,12 @@ "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26", "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_30", "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_35", - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_65_bfp.out25_37", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37", "/transformer_blocks.6/Add_3_output_0.out10_29", "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27", "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_32", "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_36", - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_66_bfp.out25_38", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38", "/transformer_blocks.6/Add_7_output_0.out10_31", "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_28_bfp.out15_28", "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_29_bfp.out15_29", @@ -23257,12 +23257,12 @@ "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30", "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_34", "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_40", - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_74_bfp.out25_42", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42", "/transformer_blocks.7/Add_3_output_0.out10_33", "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31", "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_36", "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_41", - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_75_bfp.out25_43", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43", "/transformer_blocks.7/Add_7_output_0.out10_35", "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_32_bfp.out15_32", "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_33_bfp.out15_33", @@ -23279,12 +23279,12 @@ "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34", "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_38", "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_45", - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_83_bfp.out25_47", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47", "/transformer_blocks.8/Add_3_output_0.out10_37", "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35", "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_40", "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_46", - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_48", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48", "/transformer_blocks.8/Add_7_output_0.out10_39", "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_36_bfp.out15_36", "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_37_bfp.out15_37", @@ -23301,14 +23301,14 @@ "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38", "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_42", "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_50", - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_92_bfp.out25_52", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52", "/transformer_blocks.9/Add_3_output_0.out10_41", "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40", "/transformer_blocks.10/norm1/Add_2_output_0.out0_0_40_bfp.out1_6", "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39", "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_44", "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_51", - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_93_bfp.out25_53", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53", "/transformer_blocks.9/Add_7_output_0.out10_43", "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41", "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", @@ -23321,14 +23321,14 @@ "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42", "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7", "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_55", - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_101_bfp.out25_57", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57", "/transformer_blocks.10/Add_3_output_0.out10_7", "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21", "/transformer_blocks.10/Add_4_output_0.out10_8", "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43", "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9", "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_56", - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_102_bfp.out25_58", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58", "/transformer_blocks.10/Add_7_output_0.out10_9", "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_44_bfp.out15_44", "/transformer_blocks.11/norm1/Add_2_output_0.out0_0_44_bfp.out1_10", @@ -23343,7 +23343,7 @@ "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46", "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11", "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_60", - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_110_bfp.out25_61", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_61", "/transformer_blocks.11/Add_3_output_0.out10_11" ] }, @@ -23353,320 +23353,320 @@ "packed_tensors": [ "pos_embed.proj.weight", "pos_embed_input.proj.weight", - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2", + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2", "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1", - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3", - "time_text_embed.text_embedder.linear_1.weight_5_1_0", + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3", + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0", "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0", - "time_text_embed.text_embedder.linear_2.weight_5_1_1", + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1", "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2", - "encoder_hidden_states.out17_3_0_bfp.wts", - "onnx::MatMul_3779", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts", - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0", + "encoder_hidden_states.out17_4_0_bfp.wts", + "onnx::MatMul_3779_45_1_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts", + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_1_transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_0", - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0", + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_1_transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_0", "onnx::MatMul_3783_onnx::MatMul_3780", "onnx::MatMul_3784_onnx::MatMul_3781", "onnx::MatMul_3785_onnx::MatMul_3782", - "onnx::MatMul_3800", + "onnx::MatMul_3800_45_5_1", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_2_gma", - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0", + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_3_transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_4", - "onnx::MatMul_3803", - "onnx::MatMul_3804", + "onnx::MatMul_3803_45_1_2", + "onnx::MatMul_3804_45_1_4", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_5_gma", - "onnx::MatMul_3799", + "onnx::MatMul_3799_45_5_0", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_2_gma", - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0", + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_3_transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_4", - "onnx::MatMul_3801", - "onnx::MatMul_3802", + "onnx::MatMul_3801_45_1_1", + "onnx::MatMul_3802_45_1_3", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_5_gma", - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0", + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_0_transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_1", - "onnx::MatMul_4074", - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0", + "onnx::MatMul_4074_45_0_4", + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_0_transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_1", "onnx::MatMul_3808_onnx::MatMul_3805", "onnx::MatMul_3809_onnx::MatMul_3806", "onnx::MatMul_3810_onnx::MatMul_3807", - "onnx::MatMul_3824", + "onnx::MatMul_3824_45_5_2", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_2_gma", - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0", + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_3_transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_4", - "onnx::MatMul_3826", - "onnx::MatMul_3827", + "onnx::MatMul_3826_45_1_5", + "onnx::MatMul_3827_45_1_7", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_5_gma", - "onnx::MatMul_4075", - "onnx::MatMul_3825", + "onnx::MatMul_4075_45_0_5", + "onnx::MatMul_3825_45_5_3", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_2_gma", - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0", + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_3_transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_4", - "onnx::MatMul_3828", - "onnx::MatMul_3829", + "onnx::MatMul_3828_45_1_6", + "onnx::MatMul_3829_45_1_8", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_5_gma", - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8", + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_0_transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_1", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_0_transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_1", "onnx::MatMul_3833_onnx::MatMul_3830", "onnx::MatMul_3834_onnx::MatMul_3831", "onnx::MatMul_3835_onnx::MatMul_3832", - "onnx::MatMul_3850", + "onnx::MatMul_3850_45_5_5", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_2_gma", - "onnx::MatMul_3849", + "onnx::MatMul_3849_45_5_4", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_2_gma", - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0", + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_3_transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_4", - "onnx::MatMul_3851", - "onnx::MatMul_3852", + "onnx::MatMul_3851_45_1_9", + "onnx::MatMul_3852_45_1_11", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_5_gma", - "onnx::MatMul_4076", - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0", + "onnx::MatMul_4076_45_0_6", + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_3_transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_4", - "onnx::MatMul_3853", - "onnx::MatMul_3854", + "onnx::MatMul_3853_45_1_10", + "onnx::MatMul_3854_45_1_12", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_5_gma", - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12", + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_0_transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_1", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_0_transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_1", "onnx::MatMul_3858_onnx::MatMul_3855", "onnx::MatMul_3859_onnx::MatMul_3856", "onnx::MatMul_3860_onnx::MatMul_3857", - "onnx::MatMul_3875", + "onnx::MatMul_3875_45_5_7", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_2_gma", - "onnx::MatMul_3874", + "onnx::MatMul_3874_45_5_6", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_2_gma", - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0", + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_3_transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_4", - "onnx::MatMul_3876", - "onnx::MatMul_3877", + "onnx::MatMul_3876_45_1_13", + "onnx::MatMul_3877_45_1_15", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_5_gma", - "onnx::MatMul_4077", - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0", + "onnx::MatMul_4077_45_0_7", + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_3_transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_4", - "onnx::MatMul_3878", - "onnx::MatMul_3879", + "onnx::MatMul_3878_45_1_14", + "onnx::MatMul_3879_45_1_16", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_5_gma", - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16", + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_0_transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_1", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_0_transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_1", "onnx::MatMul_3883_onnx::MatMul_3880", "onnx::MatMul_3884_onnx::MatMul_3881", "onnx::MatMul_3885_onnx::MatMul_3882", - "onnx::MatMul_3900", + "onnx::MatMul_3900_45_5_9", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_2_gma", - "onnx::MatMul_3899", + "onnx::MatMul_3899_45_5_8", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_2_gma", - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0", + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_3_transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_4", - "onnx::MatMul_3901", - "onnx::MatMul_3902", + "onnx::MatMul_3901_45_1_17", + "onnx::MatMul_3902_45_1_19", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_5_gma", - "onnx::MatMul_4078", - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0", + "onnx::MatMul_4078_45_0_8", + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_3_transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_4", - "onnx::MatMul_3903", - "onnx::MatMul_3904", + "onnx::MatMul_3903_45_1_18", + "onnx::MatMul_3904_45_1_20", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_5_gma", - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20", + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_0_transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_1", "transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_0_transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_1", "onnx::MatMul_3908_onnx::MatMul_3905", "onnx::MatMul_3909_onnx::MatMul_3906", "onnx::MatMul_3910_onnx::MatMul_3907", - "onnx::MatMul_3925", + "onnx::MatMul_3925_45_5_11", "transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_2_gma", - "onnx::MatMul_3924", + "onnx::MatMul_3924_45_5_10", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_2_gma", - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0", + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_3_transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_4", - "onnx::MatMul_3926", - "onnx::MatMul_3927", + "onnx::MatMul_3926_45_1_21", + "onnx::MatMul_3927_45_1_23", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_5_gma", - "onnx::MatMul_4079", - "/transformer_blocks.5/norm2_context/Constant_output_0", - "/transformer_blocks.5/norm2_context/Constant_1_output_0", + "onnx::MatMul_4079_45_0_9", + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23", "transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_3_transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_4", - "onnx::MatMul_3928", - "onnx::MatMul_3929", + "onnx::MatMul_3928_45_1_22", + "onnx::MatMul_3929_45_1_24", "transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_5_gma", - "/transformer_blocks.6/norm1/norm/Constant_output_0", - "/transformer_blocks.6/norm1/norm/Constant_1_output_0", - "/transformer_blocks.6/norm1_context/norm/Constant_output_0", - "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_24", + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_25", "transformer_blocks.6.norm1.linear.weight_5_1_17_27_13_0_transformer_blocks.6.norm1.linear.weight_5_1_17_27_13_1", "transformer_blocks.6.norm1_context.linear.weight_5_1_16_27_12_0_transformer_blocks.6.norm1_context.linear.weight_5_1_16_27_12_1", "onnx::MatMul_3933_onnx::MatMul_3930", "onnx::MatMul_3934_onnx::MatMul_3931", "onnx::MatMul_3935_onnx::MatMul_3932", - "onnx::MatMul_3950", + "onnx::MatMul_3950_45_5_13", "transformer_blocks.6.norm1_context.linear.weight_5_1_16_27_12_2_gma", - "onnx::MatMul_3949", + "onnx::MatMul_3949_45_5_12", "transformer_blocks.6.norm1.linear.weight_5_1_17_27_13_2_gma", - "/transformer_blocks.6/norm2/Constant_output_0", - "/transformer_blocks.6/norm2/Constant_1_output_0", + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26", "transformer_blocks.6.norm1.linear.weight_5_1_17_27_13_3_transformer_blocks.6.norm1.linear.weight_5_1_17_27_13_4", - "onnx::MatMul_3951", - "onnx::MatMul_3952", + "onnx::MatMul_3951_45_1_25", + "onnx::MatMul_3952_45_1_27", "transformer_blocks.6.norm1.linear.weight_5_1_17_27_13_5_gma", - "onnx::MatMul_4080", - "/transformer_blocks.6/norm2_context/Constant_output_0", - "/transformer_blocks.6/norm2_context/Constant_1_output_0", + "onnx::MatMul_4080_45_0_10", + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27", "transformer_blocks.6.norm1_context.linear.weight_5_1_16_27_12_3_transformer_blocks.6.norm1_context.linear.weight_5_1_16_27_12_4", - "onnx::MatMul_3953", - "onnx::MatMul_3954", + "onnx::MatMul_3953_45_1_26", + "onnx::MatMul_3954_45_1_28", "transformer_blocks.6.norm1_context.linear.weight_5_1_16_27_12_5_gma", - "/transformer_blocks.7/norm1/norm/Constant_output_0", - "/transformer_blocks.7/norm1/norm/Constant_1_output_0", - "/transformer_blocks.7/norm1_context/norm/Constant_output_0", - "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_28", + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_29", "transformer_blocks.7.norm1.linear.weight_5_1_19_27_15_0_transformer_blocks.7.norm1.linear.weight_5_1_19_27_15_1", "transformer_blocks.7.norm1_context.linear.weight_5_1_18_27_14_0_transformer_blocks.7.norm1_context.linear.weight_5_1_18_27_14_1", "onnx::MatMul_3958_onnx::MatMul_3955", "onnx::MatMul_3959_onnx::MatMul_3956", "onnx::MatMul_3960_onnx::MatMul_3957", - "onnx::MatMul_3975", + "onnx::MatMul_3975_45_5_15", "transformer_blocks.7.norm1_context.linear.weight_5_1_18_27_14_2_gma", - "onnx::MatMul_3974", + "onnx::MatMul_3974_45_5_14", "transformer_blocks.7.norm1.linear.weight_5_1_19_27_15_2_gma", - "/transformer_blocks.7/norm2/Constant_output_0", - "/transformer_blocks.7/norm2/Constant_1_output_0", + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30", "transformer_blocks.7.norm1.linear.weight_5_1_19_27_15_3_transformer_blocks.7.norm1.linear.weight_5_1_19_27_15_4", - "onnx::MatMul_3976", - "onnx::MatMul_3977", + "onnx::MatMul_3976_45_1_29", + "onnx::MatMul_3977_45_1_31", "transformer_blocks.7.norm1.linear.weight_5_1_19_27_15_5_gma", - "onnx::MatMul_4081", - "/transformer_blocks.7/norm2_context/Constant_output_0", - "/transformer_blocks.7/norm2_context/Constant_1_output_0", + "onnx::MatMul_4081_45_0_11", + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31", "transformer_blocks.7.norm1_context.linear.weight_5_1_18_27_14_3_transformer_blocks.7.norm1_context.linear.weight_5_1_18_27_14_4", - "onnx::MatMul_3978", - "onnx::MatMul_3979", + "onnx::MatMul_3978_45_1_30", + "onnx::MatMul_3979_45_1_32", "transformer_blocks.7.norm1_context.linear.weight_5_1_18_27_14_5_gma", - "/transformer_blocks.8/norm1/norm/Constant_output_0", - "/transformer_blocks.8/norm1/norm/Constant_1_output_0", - "/transformer_blocks.8/norm1_context/norm/Constant_output_0", - "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_32", + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_33", "transformer_blocks.8.norm1.linear.weight_5_1_21_27_17_0_transformer_blocks.8.norm1.linear.weight_5_1_21_27_17_1", "transformer_blocks.8.norm1_context.linear.weight_5_1_20_27_16_0_transformer_blocks.8.norm1_context.linear.weight_5_1_20_27_16_1", "onnx::MatMul_3983_onnx::MatMul_3980", "onnx::MatMul_3984_onnx::MatMul_3981", "onnx::MatMul_3985_onnx::MatMul_3982", - "onnx::MatMul_4000", + "onnx::MatMul_4000_45_5_17", "transformer_blocks.8.norm1_context.linear.weight_5_1_20_27_16_2_gma", - "onnx::MatMul_3999", + "onnx::MatMul_3999_45_5_16", "transformer_blocks.8.norm1.linear.weight_5_1_21_27_17_2_gma", - "/transformer_blocks.8/norm2/Constant_output_0", - "/transformer_blocks.8/norm2/Constant_1_output_0", + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34", "transformer_blocks.8.norm1.linear.weight_5_1_21_27_17_3_transformer_blocks.8.norm1.linear.weight_5_1_21_27_17_4", - "onnx::MatMul_4001", - "onnx::MatMul_4002", + "onnx::MatMul_4001_45_1_33", + "onnx::MatMul_4002_45_1_35", "transformer_blocks.8.norm1.linear.weight_5_1_21_27_17_5_gma", - "onnx::MatMul_4082", - "/transformer_blocks.8/norm2_context/Constant_output_0", - "/transformer_blocks.8/norm2_context/Constant_1_output_0", + "onnx::MatMul_4082_45_0_12", + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35", "transformer_blocks.8.norm1_context.linear.weight_5_1_20_27_16_3_transformer_blocks.8.norm1_context.linear.weight_5_1_20_27_16_4", - "onnx::MatMul_4003", - "onnx::MatMul_4004", + "onnx::MatMul_4003_45_1_34", + "onnx::MatMul_4004_45_1_36", "transformer_blocks.8.norm1_context.linear.weight_5_1_20_27_16_5_gma", - "/transformer_blocks.9/norm1/norm/Constant_output_0", - "/transformer_blocks.9/norm1/norm/Constant_1_output_0", - "/transformer_blocks.9/norm1_context/norm/Constant_output_0", - "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_36", + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_37", "transformer_blocks.9.norm1.linear.weight_5_1_23_27_19_0_transformer_blocks.9.norm1.linear.weight_5_1_23_27_19_1", "transformer_blocks.9.norm1_context.linear.weight_5_1_22_27_18_0_transformer_blocks.9.norm1_context.linear.weight_5_1_22_27_18_1", "onnx::MatMul_4008_onnx::MatMul_4005", "onnx::MatMul_4009_onnx::MatMul_4006", "onnx::MatMul_4010_onnx::MatMul_4007", - "onnx::MatMul_4025", + "onnx::MatMul_4025_45_5_19", "transformer_blocks.9.norm1_context.linear.weight_5_1_22_27_18_2_gma", - "onnx::MatMul_4024", + "onnx::MatMul_4024_45_5_18", "transformer_blocks.9.norm1.linear.weight_5_1_23_27_19_2_gma", - "/transformer_blocks.9/norm2/Constant_output_0", - "/transformer_blocks.9/norm2/Constant_1_output_0", + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38", "transformer_blocks.9.norm1.linear.weight_5_1_23_27_19_3_transformer_blocks.9.norm1.linear.weight_5_1_23_27_19_4", - "onnx::MatMul_4026", - "onnx::MatMul_4027", + "onnx::MatMul_4026_45_1_37", + "onnx::MatMul_4027_45_1_39", "transformer_blocks.9.norm1.linear.weight_5_1_23_27_19_5_gma", - "/transformer_blocks.10/norm1/norm/Constant_output_0", - "/transformer_blocks.10/norm1/norm/Constant_1_output_0", + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40", "transformer_blocks.10.norm1.linear.weight_5_1_25_27_21_0_transformer_blocks.10.norm1.linear.weight_5_1_25_27_21_1", - "onnx::MatMul_4083", - "/transformer_blocks.9/norm2_context/Constant_output_0", - "/transformer_blocks.9/norm2_context/Constant_1_output_0", + "onnx::MatMul_4083_45_0_13", + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39", "transformer_blocks.9.norm1_context.linear.weight_5_1_22_27_18_3_transformer_blocks.9.norm1_context.linear.weight_5_1_22_27_18_4", - "onnx::MatMul_4028", - "onnx::MatMul_4029", + "onnx::MatMul_4028_45_1_38", + "onnx::MatMul_4029_45_1_40", "transformer_blocks.9.norm1_context.linear.weight_5_1_22_27_18_5_gma", - "/transformer_blocks.10/norm1_context/norm/Constant_output_0", - "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41", "transformer_blocks.10.norm1_context.linear.weight_5_1_24_27_20_0_transformer_blocks.10.norm1_context.linear.weight_5_1_24_27_20_1", "onnx::MatMul_4033_onnx::MatMul_4030", "onnx::MatMul_4034_onnx::MatMul_4031", "onnx::MatMul_4035_onnx::MatMul_4032", - "onnx::MatMul_4049", + "onnx::MatMul_4049_45_5_20", "transformer_blocks.10.norm1.linear.weight_5_1_25_27_21_2_gma", - "/transformer_blocks.10/norm2/Constant_output_0", - "/transformer_blocks.10/norm2/Constant_1_output_0", + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42", "transformer_blocks.10.norm1.linear.weight_5_1_25_27_21_3_transformer_blocks.10.norm1.linear.weight_5_1_25_27_21_4", - "onnx::MatMul_4051", - "onnx::MatMul_4052", + "onnx::MatMul_4051_45_1_41", + "onnx::MatMul_4052_45_1_43", "transformer_blocks.10.norm1.linear.weight_5_1_25_27_21_5_gma", - "onnx::MatMul_4084", - "onnx::MatMul_4050", + "onnx::MatMul_4084_45_0_14", + "onnx::MatMul_4050_45_5_21", "transformer_blocks.10.norm1_context.linear.weight_5_1_24_27_20_2_gma", - "/transformer_blocks.10/norm2_context/Constant_output_0", - "/transformer_blocks.10/norm2_context/Constant_1_output_0", + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43", "transformer_blocks.10.norm1_context.linear.weight_5_1_24_27_20_3_transformer_blocks.10.norm1_context.linear.weight_5_1_24_27_20_4", - "onnx::MatMul_4053", - "onnx::MatMul_4054", + "onnx::MatMul_4053_45_1_42", + "onnx::MatMul_4054_45_1_44", "transformer_blocks.10.norm1_context.linear.weight_5_1_24_27_20_5_gma", - "/transformer_blocks.11/norm1/norm/Constant_output_0", - "/transformer_blocks.11/norm1/norm/Constant_1_output_0", + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_44", "transformer_blocks.11.norm1.linear.weight_5_1_27_27_23_0_transformer_blocks.11.norm1.linear.weight_5_1_27_27_23_1", - "/transformer_blocks.11/norm1_context/norm/Constant_output_0", - "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_45", "transformer_blocks.11.norm1_context.linear.weight_5_1_26_27_22_0_transformer_blocks.11.norm1_context.linear.weight_5_1_26_27_22_1", "onnx::MatMul_4058_onnx::MatMul_4055", "onnx::MatMul_4059_onnx::MatMul_4056", "onnx::MatMul_4060_onnx::MatMul_4057", - "onnx::MatMul_4071", + "onnx::MatMul_4071_45_5_22", "transformer_blocks.11.norm1.linear.weight_5_1_27_27_23_2_gma", - "/transformer_blocks.11/norm2/Constant_output_0", - "/transformer_blocks.11/norm2/Constant_1_output_0", + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46", "transformer_blocks.11.norm1.linear.weight_5_1_27_27_23_3_transformer_blocks.11.norm1.linear.weight_5_1_27_27_23_4", - "onnx::MatMul_4072", - "onnx::MatMul_4073", + "onnx::MatMul_4072_45_1_45", + "onnx::MatMul_4073_45_1_46", "transformer_blocks.11.norm1.linear.weight_5_1_27_27_23_5_gma", - "onnx::MatMul_4085" + "onnx::MatMul_4085_45_0_15" ] }, "super_instr": { @@ -23734,7 +23734,7 @@ ], "offset": 3104 }, - "/time_text_embed/Cast_output_0.out17_3_3": { + "/time_text_embed/Cast_output_0.out17_4_3": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -23750,7 +23750,7 @@ ], "offset": 3136 }, - "pooled_projections.out17_3_1": { + "pooled_projections.out17_4_1": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -23766,7 +23766,7 @@ ], "offset": 3648 }, - "encoder_hidden_states.out17_3_0": { + "encoder_hidden_states.out17_4_0": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -23784,7 +23784,7 @@ ], "offset": 7744 }, - "/controlnet_blocks.0/Add_output_0.out17_3_13": { + "/controlnet_blocks.0/Add_output_0.out17_4_13": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23802,7 +23802,7 @@ ], "offset": 0 }, - "/controlnet_blocks.1/Add_output_0.out17_3_22": { + "/controlnet_blocks.1/Add_output_0.out17_4_22": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23820,7 +23820,7 @@ ], "offset": 3072 }, - "/controlnet_blocks.2/Add_output_0.out17_3_31": { + "/controlnet_blocks.2/Add_output_0.out17_4_31": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23838,7 +23838,7 @@ ], "offset": 6144 }, - "/controlnet_blocks.3/Add_output_0.out17_3_40": { + "/controlnet_blocks.3/Add_output_0.out17_4_40": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23856,7 +23856,7 @@ ], "offset": 9216 }, - "/controlnet_blocks.4/Add_output_0.out17_3_49": { + "/controlnet_blocks.4/Add_output_0.out17_4_49": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23874,7 +23874,7 @@ ], "offset": 12288 }, - "/controlnet_blocks.5/Add_output_0.out17_3_58": { + "/controlnet_blocks.5/Add_output_0.out17_4_58": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23892,7 +23892,7 @@ ], "offset": 15360 }, - "/controlnet_blocks.6/Add_output_0.out17_3_67": { + "/controlnet_blocks.6/Add_output_0.out17_4_67": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23910,7 +23910,7 @@ ], "offset": 18432 }, - "/controlnet_blocks.7/Add_output_0.out17_3_76": { + "/controlnet_blocks.7/Add_output_0.out17_4_76": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23928,7 +23928,7 @@ ], "offset": 21504 }, - "/controlnet_blocks.8/Add_output_0.out17_3_85": { + "/controlnet_blocks.8/Add_output_0.out17_4_85": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23946,7 +23946,7 @@ ], "offset": 24576 }, - "/controlnet_blocks.9/Add_output_0.out17_3_94": { + "/controlnet_blocks.9/Add_output_0.out17_4_94": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23964,7 +23964,7 @@ ], "offset": 27648 }, - "/controlnet_blocks.10/Add_output_0.out17_3_103": { + "/controlnet_blocks.10/Add_output_0.out17_4_103": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -23982,7 +23982,7 @@ ], "offset": 30720 }, - "/controlnet_blocks.11/Add_output_0.out17_3_111": { + "/controlnet_blocks.11/Add_output_0.out17_4_111": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -24072,7 +24072,7 @@ ], "offset": 9216 }, - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3": { + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -24104,7 +24104,7 @@ ], "offset": 15360 }, - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4": { + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -24120,7 +24120,7 @@ ], "offset": 18432 }, - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1": { + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -24152,7 +24152,7 @@ ], "offset": 24576 }, - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2": { + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -24200,7 +24200,7 @@ ], "offset": 33792 }, - "encoder_hidden_states.out17_3_0_bfp.out25_0": { + "encoder_hidden_states.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -24218,7 +24218,7 @@ ], "offset": 36864 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -24236,7 +24236,7 @@ ], "offset": 40960 }, - "/context_embedder/Add_output_0.out17_3_0": { + "/context_embedder/Add_output_0.out17_4_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -24494,7 +24494,7 @@ ], "offset": 65536 }, - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8": { + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -24620,7 +24620,7 @@ ], "offset": 83968 }, - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7": { + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -24896,7 +24896,7 @@ ], "offset": 114688 }, - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12": { + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -25022,7 +25022,7 @@ ], "offset": 133120 }, - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13": { + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -25334,7 +25334,7 @@ ], "offset": 168448 }, - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17": { + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -25424,7 +25424,7 @@ ], "offset": 182272 }, - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18": { + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -25736,7 +25736,7 @@ ], "offset": 217600 }, - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22": { + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -25826,7 +25826,7 @@ ], "offset": 231424 }, - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23": { + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -26138,7 +26138,7 @@ ], "offset": 266752 }, - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27": { + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -26228,7 +26228,7 @@ ], "offset": 280576 }, - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28": { + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -26540,7 +26540,7 @@ ], "offset": 315904 }, - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_32": { + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_32": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -26630,7 +26630,7 @@ ], "offset": 329728 }, - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_57_bfp.out25_33": { + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_57_bfp.out25_33": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -26942,7 +26942,7 @@ ], "offset": 365056 }, - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_65_bfp.out25_37": { + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_65_bfp.out25_37": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -27032,7 +27032,7 @@ ], "offset": 378880 }, - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_66_bfp.out25_38": { + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_66_bfp.out25_38": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -27344,7 +27344,7 @@ ], "offset": 414208 }, - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_74_bfp.out25_42": { + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_74_bfp.out25_42": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -27434,7 +27434,7 @@ ], "offset": 428032 }, - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_75_bfp.out25_43": { + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_75_bfp.out25_43": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -27746,7 +27746,7 @@ ], "offset": 463360 }, - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_83_bfp.out25_47": { + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_83_bfp.out25_47": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -27836,7 +27836,7 @@ ], "offset": 477184 }, - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_48": { + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_48": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -28148,7 +28148,7 @@ ], "offset": 512512 }, - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_92_bfp.out25_52": { + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_92_bfp.out25_52": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -28274,7 +28274,7 @@ ], "offset": 529408 }, - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_93_bfp.out25_53": { + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_93_bfp.out25_53": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -28514,7 +28514,7 @@ ], "offset": 557056 }, - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_101_bfp.out25_57": { + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_101_bfp.out25_57": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -28640,7 +28640,7 @@ ], "offset": 575488 }, - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_102_bfp.out25_58": { + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_102_bfp.out25_58": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -28916,7 +28916,7 @@ ], "offset": 606208 }, - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_110_bfp.out25_61": { + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_110_bfp.out25_61": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -28978,7 +28978,7 @@ "file_name": "cache/pos_embedprojConv_1.const", "file_size": 135168 }, - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2": { + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29004,7 +29004,7 @@ "file_name": "cache/pos_embedprojConv_3.const", "file_size": 256 }, - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3": { + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29017,7 +29017,7 @@ "file_name": "cache/pos_embedprojConv_4.const", "file_size": 2691072 }, - "time_text_embed.text_embedder.linear_1.weight_5_1_0": { + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29043,7 +29043,7 @@ "file_name": "cache/pos_embedprojConv_6.const", "file_size": 256 }, - "time_text_embed.text_embedder.linear_2.weight_5_1_1": { + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29069,7 +29069,7 @@ "file_name": "cache/pos_embedprojConv_8.const", "file_size": 256 }, - "encoder_hidden_states.out17_3_0_bfp.wts": { + "encoder_hidden_states.out17_4_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29082,7 +29082,7 @@ "file_name": "cache/pos_embedprojConv_9.const", "file_size": 128 }, - "onnx::MatMul_3779": { + "onnx::MatMul_3779_45_1_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29095,7 +29095,7 @@ "file_name": "cache/pos_embedprojConv_10.const", "file_size": 7274496 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29108,7 +29108,7 @@ "file_name": "cache/pos_embedprojConv_11.const", "file_size": 128 }, - "/transformer_blocks.0/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29121,7 +29121,7 @@ "file_name": "cache/pos_embedprojConv_12.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29147,7 +29147,7 @@ "file_name": "cache/pos_embedprojConv_14.const", "file_size": 5382144 }, - "/transformer_blocks.0/norm1/norm/Constant_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29160,7 +29160,7 @@ "file_name": "cache/pos_embedprojConv_15.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29225,7 +29225,7 @@ "file_name": "cache/pos_embedprojConv_20.const", "file_size": 5382144 }, - "onnx::MatMul_3800": { + "onnx::MatMul_3800_45_5_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29251,7 +29251,7 @@ "file_name": "cache/pos_embedprojConv_22.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2_context/Constant_output_0": { + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29264,7 +29264,7 @@ "file_name": "cache/pos_embedprojConv_23.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2_context/Constant_1_output_0": { + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29290,7 +29290,7 @@ "file_name": "cache/pos_embedprojConv_25.const", "file_size": 5382144 }, - "onnx::MatMul_3803": { + "onnx::MatMul_3803_45_1_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29303,7 +29303,7 @@ "file_name": "cache/pos_embedprojConv_26.const", "file_size": 10764288 }, - "onnx::MatMul_3804": { + "onnx::MatMul_3804_45_1_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29329,7 +29329,7 @@ "file_name": "cache/pos_embedprojConv_28.const", "file_size": 2691072 }, - "onnx::MatMul_3799": { + "onnx::MatMul_3799_45_5_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29355,7 +29355,7 @@ "file_name": "cache/pos_embedprojConv_30.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2/Constant_output_0": { + "/transformer_blocks.0/norm2/Constant_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29368,7 +29368,7 @@ "file_name": "cache/pos_embedprojConv_31.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2/Constant_1_output_0": { + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29394,7 +29394,7 @@ "file_name": "cache/pos_embedprojConv_33.const", "file_size": 5382144 }, - "onnx::MatMul_3801": { + "onnx::MatMul_3801_45_1_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29407,7 +29407,7 @@ "file_name": "cache/pos_embedprojConv_34.const", "file_size": 10764288 }, - "onnx::MatMul_3802": { + "onnx::MatMul_3802_45_1_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29433,7 +29433,7 @@ "file_name": "cache/pos_embedprojConv_36.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1/norm/Constant_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29446,7 +29446,7 @@ "file_name": "cache/pos_embedprojConv_37.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29472,7 +29472,7 @@ "file_name": "cache/pos_embedprojConv_39.const", "file_size": 5382144 }, - "onnx::MatMul_4074": { + "onnx::MatMul_4074_45_0_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29485,7 +29485,7 @@ "file_name": "cache/pos_embedprojConv_40.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29498,7 +29498,7 @@ "file_name": "cache/pos_embedprojConv_41.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29563,7 +29563,7 @@ "file_name": "cache/pos_embedprojConv_46.const", "file_size": 5382144 }, - "onnx::MatMul_3824": { + "onnx::MatMul_3824_45_5_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29589,7 +29589,7 @@ "file_name": "cache/pos_embedprojConv_48.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2/Constant_output_0": { + "/transformer_blocks.1/norm2/Constant_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29602,7 +29602,7 @@ "file_name": "cache/pos_embedprojConv_49.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2/Constant_1_output_0": { + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29628,7 +29628,7 @@ "file_name": "cache/pos_embedprojConv_51.const", "file_size": 5382144 }, - "onnx::MatMul_3826": { + "onnx::MatMul_3826_45_1_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29641,7 +29641,7 @@ "file_name": "cache/pos_embedprojConv_52.const", "file_size": 10764288 }, - "onnx::MatMul_3827": { + "onnx::MatMul_3827_45_1_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29667,7 +29667,7 @@ "file_name": "cache/pos_embedprojConv_54.const", "file_size": 2691072 }, - "onnx::MatMul_4075": { + "onnx::MatMul_4075_45_0_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29680,7 +29680,7 @@ "file_name": "cache/pos_embedprojConv_55.const", "file_size": 2691072 }, - "onnx::MatMul_3825": { + "onnx::MatMul_3825_45_5_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29706,7 +29706,7 @@ "file_name": "cache/pos_embedprojConv_57.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2_context/Constant_output_0": { + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29719,7 +29719,7 @@ "file_name": "cache/pos_embedprojConv_58.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2_context/Constant_1_output_0": { + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29745,7 +29745,7 @@ "file_name": "cache/pos_embedprojConv_60.const", "file_size": 5382144 }, - "onnx::MatMul_3828": { + "onnx::MatMul_3828_45_1_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29758,7 +29758,7 @@ "file_name": "cache/pos_embedprojConv_61.const", "file_size": 10764288 }, - "onnx::MatMul_3829": { + "onnx::MatMul_3829_45_1_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29784,7 +29784,7 @@ "file_name": "cache/pos_embedprojConv_63.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm1/norm/Constant_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29797,7 +29797,7 @@ "file_name": "cache/pos_embedprojConv_64.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29810,7 +29810,7 @@ "file_name": "cache/pos_embedprojConv_65.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29823,7 +29823,7 @@ "file_name": "cache/pos_embedprojConv_66.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29901,7 +29901,7 @@ "file_name": "cache/pos_embedprojConv_72.const", "file_size": 5382144 }, - "onnx::MatMul_3850": { + "onnx::MatMul_3850_45_5_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29927,7 +29927,7 @@ "file_name": "cache/pos_embedprojConv_74.const", "file_size": 2691072 }, - "onnx::MatMul_3849": { + "onnx::MatMul_3849_45_5_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -29953,7 +29953,7 @@ "file_name": "cache/pos_embedprojConv_76.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2/Constant_output_0": { + "/transformer_blocks.2/norm2/Constant_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29966,7 +29966,7 @@ "file_name": "cache/pos_embedprojConv_77.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2/Constant_1_output_0": { + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -29992,7 +29992,7 @@ "file_name": "cache/pos_embedprojConv_79.const", "file_size": 5382144 }, - "onnx::MatMul_3851": { + "onnx::MatMul_3851_45_1_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30005,7 +30005,7 @@ "file_name": "cache/pos_embedprojConv_80.const", "file_size": 10764288 }, - "onnx::MatMul_3852": { + "onnx::MatMul_3852_45_1_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30031,7 +30031,7 @@ "file_name": "cache/pos_embedprojConv_82.const", "file_size": 2691072 }, - "onnx::MatMul_4076": { + "onnx::MatMul_4076_45_0_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30044,7 +30044,7 @@ "file_name": "cache/pos_embedprojConv_83.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2_context/Constant_output_0": { + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30057,7 +30057,7 @@ "file_name": "cache/pos_embedprojConv_84.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2_context/Constant_1_output_0": { + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30083,7 +30083,7 @@ "file_name": "cache/pos_embedprojConv_86.const", "file_size": 5382144 }, - "onnx::MatMul_3853": { + "onnx::MatMul_3853_45_1_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30096,7 +30096,7 @@ "file_name": "cache/pos_embedprojConv_87.const", "file_size": 10764288 }, - "onnx::MatMul_3854": { + "onnx::MatMul_3854_45_1_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30122,7 +30122,7 @@ "file_name": "cache/pos_embedprojConv_89.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm1/norm/Constant_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30135,7 +30135,7 @@ "file_name": "cache/pos_embedprojConv_90.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30148,7 +30148,7 @@ "file_name": "cache/pos_embedprojConv_91.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30161,7 +30161,7 @@ "file_name": "cache/pos_embedprojConv_92.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30239,7 +30239,7 @@ "file_name": "cache/pos_embedprojConv_98.const", "file_size": 5382144 }, - "onnx::MatMul_3875": { + "onnx::MatMul_3875_45_5_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30265,7 +30265,7 @@ "file_name": "cache/pos_embedprojConv_100.const", "file_size": 2691072 }, - "onnx::MatMul_3874": { + "onnx::MatMul_3874_45_5_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30291,7 +30291,7 @@ "file_name": "cache/pos_embedprojConv_102.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2/Constant_output_0": { + "/transformer_blocks.3/norm2/Constant_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30304,7 +30304,7 @@ "file_name": "cache/pos_embedprojConv_103.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2/Constant_1_output_0": { + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30330,7 +30330,7 @@ "file_name": "cache/pos_embedprojConv_105.const", "file_size": 5382144 }, - "onnx::MatMul_3876": { + "onnx::MatMul_3876_45_1_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30343,7 +30343,7 @@ "file_name": "cache/pos_embedprojConv_106.const", "file_size": 10764288 }, - "onnx::MatMul_3877": { + "onnx::MatMul_3877_45_1_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30369,7 +30369,7 @@ "file_name": "cache/pos_embedprojConv_108.const", "file_size": 2691072 }, - "onnx::MatMul_4077": { + "onnx::MatMul_4077_45_0_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30382,7 +30382,7 @@ "file_name": "cache/pos_embedprojConv_109.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2_context/Constant_output_0": { + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30395,7 +30395,7 @@ "file_name": "cache/pos_embedprojConv_110.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2_context/Constant_1_output_0": { + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30421,7 +30421,7 @@ "file_name": "cache/pos_embedprojConv_112.const", "file_size": 5382144 }, - "onnx::MatMul_3878": { + "onnx::MatMul_3878_45_1_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30434,7 +30434,7 @@ "file_name": "cache/pos_embedprojConv_113.const", "file_size": 10764288 }, - "onnx::MatMul_3879": { + "onnx::MatMul_3879_45_1_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30460,7 +30460,7 @@ "file_name": "cache/pos_embedprojConv_115.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm1/norm/Constant_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30473,7 +30473,7 @@ "file_name": "cache/pos_embedprojConv_116.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30486,7 +30486,7 @@ "file_name": "cache/pos_embedprojConv_117.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30499,7 +30499,7 @@ "file_name": "cache/pos_embedprojConv_118.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30577,7 +30577,7 @@ "file_name": "cache/pos_embedprojConv_124.const", "file_size": 5382144 }, - "onnx::MatMul_3900": { + "onnx::MatMul_3900_45_5_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30603,7 +30603,7 @@ "file_name": "cache/pos_embedprojConv_126.const", "file_size": 2691072 }, - "onnx::MatMul_3899": { + "onnx::MatMul_3899_45_5_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30629,7 +30629,7 @@ "file_name": "cache/pos_embedprojConv_128.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2/Constant_output_0": { + "/transformer_blocks.4/norm2/Constant_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30642,7 +30642,7 @@ "file_name": "cache/pos_embedprojConv_129.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2/Constant_1_output_0": { + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30668,7 +30668,7 @@ "file_name": "cache/pos_embedprojConv_131.const", "file_size": 5382144 }, - "onnx::MatMul_3901": { + "onnx::MatMul_3901_45_1_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30681,7 +30681,7 @@ "file_name": "cache/pos_embedprojConv_132.const", "file_size": 10764288 }, - "onnx::MatMul_3902": { + "onnx::MatMul_3902_45_1_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30707,7 +30707,7 @@ "file_name": "cache/pos_embedprojConv_134.const", "file_size": 2691072 }, - "onnx::MatMul_4078": { + "onnx::MatMul_4078_45_0_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30720,7 +30720,7 @@ "file_name": "cache/pos_embedprojConv_135.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2_context/Constant_output_0": { + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30733,7 +30733,7 @@ "file_name": "cache/pos_embedprojConv_136.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2_context/Constant_1_output_0": { + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30759,7 +30759,7 @@ "file_name": "cache/pos_embedprojConv_138.const", "file_size": 5382144 }, - "onnx::MatMul_3903": { + "onnx::MatMul_3903_45_1_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30772,7 +30772,7 @@ "file_name": "cache/pos_embedprojConv_139.const", "file_size": 10764288 }, - "onnx::MatMul_3904": { + "onnx::MatMul_3904_45_1_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30798,7 +30798,7 @@ "file_name": "cache/pos_embedprojConv_141.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm1/norm/Constant_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30811,7 +30811,7 @@ "file_name": "cache/pos_embedprojConv_142.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30824,7 +30824,7 @@ "file_name": "cache/pos_embedprojConv_143.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30837,7 +30837,7 @@ "file_name": "cache/pos_embedprojConv_144.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30915,7 +30915,7 @@ "file_name": "cache/pos_embedprojConv_150.const", "file_size": 5382144 }, - "onnx::MatMul_3925": { + "onnx::MatMul_3925_45_5_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30941,7 +30941,7 @@ "file_name": "cache/pos_embedprojConv_152.const", "file_size": 2691072 }, - "onnx::MatMul_3924": { + "onnx::MatMul_3924_45_5_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -30967,7 +30967,7 @@ "file_name": "cache/pos_embedprojConv_154.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm2/Constant_output_0": { + "/transformer_blocks.5/norm2/Constant_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -30980,7 +30980,7 @@ "file_name": "cache/pos_embedprojConv_155.const", "file_size": 3072 }, - "/transformer_blocks.5/norm2/Constant_1_output_0": { + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31006,7 +31006,7 @@ "file_name": "cache/pos_embedprojConv_157.const", "file_size": 5382144 }, - "onnx::MatMul_3926": { + "onnx::MatMul_3926_45_1_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31019,7 +31019,7 @@ "file_name": "cache/pos_embedprojConv_158.const", "file_size": 10764288 }, - "onnx::MatMul_3927": { + "onnx::MatMul_3927_45_1_23": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31045,7 +31045,7 @@ "file_name": "cache/pos_embedprojConv_160.const", "file_size": 2691072 }, - "onnx::MatMul_4079": { + "onnx::MatMul_4079_45_0_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31058,7 +31058,7 @@ "file_name": "cache/pos_embedprojConv_161.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm2_context/Constant_output_0": { + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31071,7 +31071,7 @@ "file_name": "cache/pos_embedprojConv_162.const", "file_size": 3072 }, - "/transformer_blocks.5/norm2_context/Constant_1_output_0": { + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31097,7 +31097,7 @@ "file_name": "cache/pos_embedprojConv_164.const", "file_size": 5382144 }, - "onnx::MatMul_3928": { + "onnx::MatMul_3928_45_1_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31110,7 +31110,7 @@ "file_name": "cache/pos_embedprojConv_165.const", "file_size": 10764288 }, - "onnx::MatMul_3929": { + "onnx::MatMul_3929_45_1_24": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31136,7 +31136,7 @@ "file_name": "cache/pos_embedprojConv_167.const", "file_size": 2691072 }, - "/transformer_blocks.6/norm1/norm/Constant_output_0": { + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_24": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31149,7 +31149,7 @@ "file_name": "cache/pos_embedprojConv_168.const", "file_size": 3072 }, - "/transformer_blocks.6/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_24": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31162,7 +31162,7 @@ "file_name": "cache/pos_embedprojConv_169.const", "file_size": 3072 }, - "/transformer_blocks.6/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_25": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31175,7 +31175,7 @@ "file_name": "cache/pos_embedprojConv_170.const", "file_size": 3072 }, - "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_25": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31253,7 +31253,7 @@ "file_name": "cache/pos_embedprojConv_176.const", "file_size": 5382144 }, - "onnx::MatMul_3950": { + "onnx::MatMul_3950_45_5_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31279,7 +31279,7 @@ "file_name": "cache/pos_embedprojConv_178.const", "file_size": 2691072 }, - "onnx::MatMul_3949": { + "onnx::MatMul_3949_45_5_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31305,7 +31305,7 @@ "file_name": "cache/pos_embedprojConv_180.const", "file_size": 2691072 }, - "/transformer_blocks.6/norm2/Constant_output_0": { + "/transformer_blocks.6/norm2/Constant_output_0_14_26": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31318,7 +31318,7 @@ "file_name": "cache/pos_embedprojConv_181.const", "file_size": 3072 }, - "/transformer_blocks.6/norm2/Constant_1_output_0": { + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31344,7 +31344,7 @@ "file_name": "cache/pos_embedprojConv_183.const", "file_size": 5382144 }, - "onnx::MatMul_3951": { + "onnx::MatMul_3951_45_1_25": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31357,7 +31357,7 @@ "file_name": "cache/pos_embedprojConv_184.const", "file_size": 10764288 }, - "onnx::MatMul_3952": { + "onnx::MatMul_3952_45_1_27": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31383,7 +31383,7 @@ "file_name": "cache/pos_embedprojConv_186.const", "file_size": 2691072 }, - "onnx::MatMul_4080": { + "onnx::MatMul_4080_45_0_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31396,7 +31396,7 @@ "file_name": "cache/pos_embedprojConv_187.const", "file_size": 2691072 }, - "/transformer_blocks.6/norm2_context/Constant_output_0": { + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31409,7 +31409,7 @@ "file_name": "cache/pos_embedprojConv_188.const", "file_size": 3072 }, - "/transformer_blocks.6/norm2_context/Constant_1_output_0": { + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31435,7 +31435,7 @@ "file_name": "cache/pos_embedprojConv_190.const", "file_size": 5382144 }, - "onnx::MatMul_3953": { + "onnx::MatMul_3953_45_1_26": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31448,7 +31448,7 @@ "file_name": "cache/pos_embedprojConv_191.const", "file_size": 10764288 }, - "onnx::MatMul_3954": { + "onnx::MatMul_3954_45_1_28": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31474,7 +31474,7 @@ "file_name": "cache/pos_embedprojConv_193.const", "file_size": 2691072 }, - "/transformer_blocks.7/norm1/norm/Constant_output_0": { + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_28": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31487,7 +31487,7 @@ "file_name": "cache/pos_embedprojConv_194.const", "file_size": 3072 }, - "/transformer_blocks.7/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_28": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31500,7 +31500,7 @@ "file_name": "cache/pos_embedprojConv_195.const", "file_size": 3072 }, - "/transformer_blocks.7/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_29": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31513,7 +31513,7 @@ "file_name": "cache/pos_embedprojConv_196.const", "file_size": 3072 }, - "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_29": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31591,7 +31591,7 @@ "file_name": "cache/pos_embedprojConv_202.const", "file_size": 5382144 }, - "onnx::MatMul_3975": { + "onnx::MatMul_3975_45_5_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31617,7 +31617,7 @@ "file_name": "cache/pos_embedprojConv_204.const", "file_size": 2691072 }, - "onnx::MatMul_3974": { + "onnx::MatMul_3974_45_5_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31643,7 +31643,7 @@ "file_name": "cache/pos_embedprojConv_206.const", "file_size": 2691072 }, - "/transformer_blocks.7/norm2/Constant_output_0": { + "/transformer_blocks.7/norm2/Constant_output_0_14_30": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31656,7 +31656,7 @@ "file_name": "cache/pos_embedprojConv_207.const", "file_size": 3072 }, - "/transformer_blocks.7/norm2/Constant_1_output_0": { + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31682,7 +31682,7 @@ "file_name": "cache/pos_embedprojConv_209.const", "file_size": 5382144 }, - "onnx::MatMul_3976": { + "onnx::MatMul_3976_45_1_29": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31695,7 +31695,7 @@ "file_name": "cache/pos_embedprojConv_210.const", "file_size": 10764288 }, - "onnx::MatMul_3977": { + "onnx::MatMul_3977_45_1_31": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31721,7 +31721,7 @@ "file_name": "cache/pos_embedprojConv_212.const", "file_size": 2691072 }, - "onnx::MatMul_4081": { + "onnx::MatMul_4081_45_0_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31734,7 +31734,7 @@ "file_name": "cache/pos_embedprojConv_213.const", "file_size": 2691072 }, - "/transformer_blocks.7/norm2_context/Constant_output_0": { + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31747,7 +31747,7 @@ "file_name": "cache/pos_embedprojConv_214.const", "file_size": 3072 }, - "/transformer_blocks.7/norm2_context/Constant_1_output_0": { + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31773,7 +31773,7 @@ "file_name": "cache/pos_embedprojConv_216.const", "file_size": 5382144 }, - "onnx::MatMul_3978": { + "onnx::MatMul_3978_45_1_30": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31786,7 +31786,7 @@ "file_name": "cache/pos_embedprojConv_217.const", "file_size": 10764288 }, - "onnx::MatMul_3979": { + "onnx::MatMul_3979_45_1_32": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31812,7 +31812,7 @@ "file_name": "cache/pos_embedprojConv_219.const", "file_size": 2691072 }, - "/transformer_blocks.8/norm1/norm/Constant_output_0": { + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_32": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31825,7 +31825,7 @@ "file_name": "cache/pos_embedprojConv_220.const", "file_size": 3072 }, - "/transformer_blocks.8/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_32": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31838,7 +31838,7 @@ "file_name": "cache/pos_embedprojConv_221.const", "file_size": 3072 }, - "/transformer_blocks.8/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_33": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31851,7 +31851,7 @@ "file_name": "cache/pos_embedprojConv_222.const", "file_size": 3072 }, - "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_33": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31929,7 +31929,7 @@ "file_name": "cache/pos_embedprojConv_228.const", "file_size": 5382144 }, - "onnx::MatMul_4000": { + "onnx::MatMul_4000_45_5_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31955,7 +31955,7 @@ "file_name": "cache/pos_embedprojConv_230.const", "file_size": 2691072 }, - "onnx::MatMul_3999": { + "onnx::MatMul_3999_45_5_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -31981,7 +31981,7 @@ "file_name": "cache/pos_embedprojConv_232.const", "file_size": 2691072 }, - "/transformer_blocks.8/norm2/Constant_output_0": { + "/transformer_blocks.8/norm2/Constant_output_0_14_34": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -31994,7 +31994,7 @@ "file_name": "cache/pos_embedprojConv_233.const", "file_size": 3072 }, - "/transformer_blocks.8/norm2/Constant_1_output_0": { + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32020,7 +32020,7 @@ "file_name": "cache/pos_embedprojConv_235.const", "file_size": 5382144 }, - "onnx::MatMul_4001": { + "onnx::MatMul_4001_45_1_33": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32033,7 +32033,7 @@ "file_name": "cache/pos_embedprojConv_236.const", "file_size": 10764288 }, - "onnx::MatMul_4002": { + "onnx::MatMul_4002_45_1_35": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32059,7 +32059,7 @@ "file_name": "cache/pos_embedprojConv_238.const", "file_size": 2691072 }, - "onnx::MatMul_4082": { + "onnx::MatMul_4082_45_0_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32072,7 +32072,7 @@ "file_name": "cache/pos_embedprojConv_239.const", "file_size": 2691072 }, - "/transformer_blocks.8/norm2_context/Constant_output_0": { + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32085,7 +32085,7 @@ "file_name": "cache/pos_embedprojConv_240.const", "file_size": 3072 }, - "/transformer_blocks.8/norm2_context/Constant_1_output_0": { + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32111,7 +32111,7 @@ "file_name": "cache/pos_embedprojConv_242.const", "file_size": 5382144 }, - "onnx::MatMul_4003": { + "onnx::MatMul_4003_45_1_34": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32124,7 +32124,7 @@ "file_name": "cache/pos_embedprojConv_243.const", "file_size": 10764288 }, - "onnx::MatMul_4004": { + "onnx::MatMul_4004_45_1_36": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32150,7 +32150,7 @@ "file_name": "cache/pos_embedprojConv_245.const", "file_size": 2691072 }, - "/transformer_blocks.9/norm1/norm/Constant_output_0": { + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_36": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32163,7 +32163,7 @@ "file_name": "cache/pos_embedprojConv_246.const", "file_size": 3072 }, - "/transformer_blocks.9/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_36": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32176,7 +32176,7 @@ "file_name": "cache/pos_embedprojConv_247.const", "file_size": 3072 }, - "/transformer_blocks.9/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_37": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32189,7 +32189,7 @@ "file_name": "cache/pos_embedprojConv_248.const", "file_size": 3072 }, - "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_37": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32267,7 +32267,7 @@ "file_name": "cache/pos_embedprojConv_254.const", "file_size": 5382144 }, - "onnx::MatMul_4025": { + "onnx::MatMul_4025_45_5_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32293,7 +32293,7 @@ "file_name": "cache/pos_embedprojConv_256.const", "file_size": 2691072 }, - "onnx::MatMul_4024": { + "onnx::MatMul_4024_45_5_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32319,7 +32319,7 @@ "file_name": "cache/pos_embedprojConv_258.const", "file_size": 2691072 }, - "/transformer_blocks.9/norm2/Constant_output_0": { + "/transformer_blocks.9/norm2/Constant_output_0_14_38": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32332,7 +32332,7 @@ "file_name": "cache/pos_embedprojConv_259.const", "file_size": 3072 }, - "/transformer_blocks.9/norm2/Constant_1_output_0": { + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32358,7 +32358,7 @@ "file_name": "cache/pos_embedprojConv_261.const", "file_size": 5382144 }, - "onnx::MatMul_4026": { + "onnx::MatMul_4026_45_1_37": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32371,7 +32371,7 @@ "file_name": "cache/pos_embedprojConv_262.const", "file_size": 10764288 }, - "onnx::MatMul_4027": { + "onnx::MatMul_4027_45_1_39": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32397,7 +32397,7 @@ "file_name": "cache/pos_embedprojConv_264.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm1/norm/Constant_output_0": { + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32410,7 +32410,7 @@ "file_name": "cache/pos_embedprojConv_265.const", "file_size": 3072 }, - "/transformer_blocks.10/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32436,7 +32436,7 @@ "file_name": "cache/pos_embedprojConv_267.const", "file_size": 5382144 }, - "onnx::MatMul_4083": { + "onnx::MatMul_4083_45_0_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32449,7 +32449,7 @@ "file_name": "cache/pos_embedprojConv_268.const", "file_size": 2691072 }, - "/transformer_blocks.9/norm2_context/Constant_output_0": { + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32462,7 +32462,7 @@ "file_name": "cache/pos_embedprojConv_269.const", "file_size": 3072 }, - "/transformer_blocks.9/norm2_context/Constant_1_output_0": { + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32488,7 +32488,7 @@ "file_name": "cache/pos_embedprojConv_271.const", "file_size": 5382144 }, - "onnx::MatMul_4028": { + "onnx::MatMul_4028_45_1_38": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32501,7 +32501,7 @@ "file_name": "cache/pos_embedprojConv_272.const", "file_size": 10764288 }, - "onnx::MatMul_4029": { + "onnx::MatMul_4029_45_1_40": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32527,7 +32527,7 @@ "file_name": "cache/pos_embedprojConv_274.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32540,7 +32540,7 @@ "file_name": "cache/pos_embedprojConv_275.const", "file_size": 3072 }, - "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32605,7 +32605,7 @@ "file_name": "cache/pos_embedprojConv_280.const", "file_size": 5382144 }, - "onnx::MatMul_4049": { + "onnx::MatMul_4049_45_5_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32631,7 +32631,7 @@ "file_name": "cache/pos_embedprojConv_282.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm2/Constant_output_0": { + "/transformer_blocks.10/norm2/Constant_output_0_14_42": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32644,7 +32644,7 @@ "file_name": "cache/pos_embedprojConv_283.const", "file_size": 3072 }, - "/transformer_blocks.10/norm2/Constant_1_output_0": { + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32670,7 +32670,7 @@ "file_name": "cache/pos_embedprojConv_285.const", "file_size": 5382144 }, - "onnx::MatMul_4051": { + "onnx::MatMul_4051_45_1_41": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32683,7 +32683,7 @@ "file_name": "cache/pos_embedprojConv_286.const", "file_size": 10764288 }, - "onnx::MatMul_4052": { + "onnx::MatMul_4052_45_1_43": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32709,7 +32709,7 @@ "file_name": "cache/pos_embedprojConv_288.const", "file_size": 2691072 }, - "onnx::MatMul_4084": { + "onnx::MatMul_4084_45_0_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32722,7 +32722,7 @@ "file_name": "cache/pos_embedprojConv_289.const", "file_size": 2691072 }, - "onnx::MatMul_4050": { + "onnx::MatMul_4050_45_5_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32748,7 +32748,7 @@ "file_name": "cache/pos_embedprojConv_291.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm2_context/Constant_output_0": { + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32761,7 +32761,7 @@ "file_name": "cache/pos_embedprojConv_292.const", "file_size": 3072 }, - "/transformer_blocks.10/norm2_context/Constant_1_output_0": { + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32787,7 +32787,7 @@ "file_name": "cache/pos_embedprojConv_294.const", "file_size": 5382144 }, - "onnx::MatMul_4053": { + "onnx::MatMul_4053_45_1_42": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32800,7 +32800,7 @@ "file_name": "cache/pos_embedprojConv_295.const", "file_size": 10764288 }, - "onnx::MatMul_4054": { + "onnx::MatMul_4054_45_1_44": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32826,7 +32826,7 @@ "file_name": "cache/pos_embedprojConv_297.const", "file_size": 2691072 }, - "/transformer_blocks.11/norm1/norm/Constant_output_0": { + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_44": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32839,7 +32839,7 @@ "file_name": "cache/pos_embedprojConv_298.const", "file_size": 3072 }, - "/transformer_blocks.11/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_44": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32865,7 +32865,7 @@ "file_name": "cache/pos_embedprojConv_300.const", "file_size": 5382144 }, - "/transformer_blocks.11/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_45": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32878,7 +32878,7 @@ "file_name": "cache/pos_embedprojConv_301.const", "file_size": 3072 }, - "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_45": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32943,7 +32943,7 @@ "file_name": "cache/pos_embedprojConv_306.const", "file_size": 5382144 }, - "onnx::MatMul_4071": { + "onnx::MatMul_4071_45_5_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -32969,7 +32969,7 @@ "file_name": "cache/pos_embedprojConv_308.const", "file_size": 2691072 }, - "/transformer_blocks.11/norm2/Constant_output_0": { + "/transformer_blocks.11/norm2/Constant_output_0_14_46": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -32982,7 +32982,7 @@ "file_name": "cache/pos_embedprojConv_309.const", "file_size": 3072 }, - "/transformer_blocks.11/norm2/Constant_1_output_0": { + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -33008,7 +33008,7 @@ "file_name": "cache/pos_embedprojConv_311.const", "file_size": 5382144 }, - "onnx::MatMul_4072": { + "onnx::MatMul_4072_45_1_45": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -33021,7 +33021,7 @@ "file_name": "cache/pos_embedprojConv_312.const", "file_size": 10764288 }, - "onnx::MatMul_4073": { + "onnx::MatMul_4073_45_1_46": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -33047,7 +33047,7 @@ "file_name": "cache/pos_embedprojConv_314.const", "file_size": 2691072 }, - "onnx::MatMul_4085": { + "onnx::MatMul_4085_45_0_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -33064,84 +33064,204 @@ "dynamic_shape_subgraph": true, "dynamic_shape_list": [ { + "max_length + floor(h/2)*floor(w/2)": 1184, "floor(w/2)": 32, - "h": 64, - "max_length": 160, "w": 64, - "max_length + floor(h/2)*floor(w/2)": 1184, "batch_size": 2, + "floor(h/2)*floor(w/2)": 1024, "floor(h/2)": 32, - "floor(h/2)*floor(w/2)": 1024 + "h": 64, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1760, + "floor(w/2)": 40, + "w": 80, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1600, + "floor(h/2)": 40, + "h": 80, + "max_length": 160 }, { + "max_length + floor(h/2)*floor(w/2)": 2464, "floor(w/2)": 48, - "h": 64, - "max_length": 160, "w": 96, - "max_length + floor(h/2)*floor(w/2)": 1696, "batch_size": 2, + "floor(h/2)*floor(w/2)": 2304, + "floor(h/2)": 48, + "h": 96, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 4256, + "floor(w/2)": 64, + "w": 128, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 4096, + "floor(h/2)": 64, + "h": 128, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 928, + "floor(w/2)": 24, + "w": 48, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 768, "floor(h/2)": 32, - "floor(h/2)*floor(w/2)": 1536 + "h": 64, + "max_length": 160 }, { - "floor(w/2)": 32, + "max_length + floor(h/2)*floor(w/2)": 1360, + "floor(w/2)": 30, + "w": 60, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1200, + "floor(h/2)": 40, + "h": 80, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1888, + "floor(w/2)": 36, + "w": 72, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1728, + "floor(h/2)": 48, "h": 96, - "max_length": 160, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 3232, + "floor(w/2)": 48, + "w": 96, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 3072, + "floor(h/2)": 64, + "h": 128, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 928, + "floor(w/2)": 32, "w": 64, - "max_length + floor(h/2)*floor(w/2)": 1696, "batch_size": 2, - "floor(h/2)": 48, - "floor(h/2)*floor(w/2)": 1536 + "floor(h/2)*floor(w/2)": 768, + "floor(h/2)": 24, + "h": 48, + "max_length": 160 }, { - "floor(w/2)": 64, + "max_length + floor(h/2)*floor(w/2)": 1360, + "floor(w/2)": 40, + "w": 80, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1200, + "floor(h/2)": 30, + "h": 60, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1888, + "floor(w/2)": 48, + "w": 96, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1728, + "floor(h/2)": 36, "h": 72, - "max_length": 160, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 3232, + "floor(w/2)": 64, "w": 128, - "max_length + floor(h/2)*floor(w/2)": 2464, "batch_size": 2, - "floor(h/2)": 36, - "floor(h/2)*floor(w/2)": 2304 + "floor(h/2)*floor(w/2)": 3072, + "floor(h/2)": 48, + "h": 96, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 736, + "floor(w/2)": 18, + "w": 36, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 576, + "floor(h/2)": 32, + "h": 64, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1120, + "floor(w/2)": 24, + "w": 48, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 960, + "floor(h/2)": 40, + "h": 80, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1504, + "floor(w/2)": 28, + "w": 56, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1344, + "floor(h/2)": 48, + "h": 96, + "max_length": 160 }, { + "max_length + floor(h/2)*floor(w/2)": 2464, "floor(w/2)": 36, - "h": 128, - "max_length": 160, "w": 72, - "max_length + floor(h/2)*floor(w/2)": 2464, "batch_size": 2, + "floor(h/2)*floor(w/2)": 2304, "floor(h/2)": 64, - "floor(h/2)*floor(w/2)": 2304 + "h": 128, + "max_length": 160 }, { - "floor(w/2)": 64, - "h": 96, - "max_length": 160, - "w": 128, - "max_length + floor(h/2)*floor(w/2)": 3232, + "max_length + floor(h/2)*floor(w/2)": 736, + "floor(w/2)": 32, + "w": 64, "batch_size": 2, - "floor(h/2)": 48, - "floor(h/2)*floor(w/2)": 3072 + "floor(h/2)*floor(w/2)": 576, + "floor(h/2)": 18, + "h": 36, + "max_length": 160 }, { + "max_length + floor(h/2)*floor(w/2)": 1120, + "floor(w/2)": 40, + "w": 80, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 960, + "floor(h/2)": 24, + "h": 48, + "max_length": 160 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1504, "floor(w/2)": 48, - "h": 128, - "max_length": 160, "w": 96, - "max_length + floor(h/2)*floor(w/2)": 3232, "batch_size": 2, - "floor(h/2)": 64, - "floor(h/2)*floor(w/2)": 3072 + "floor(h/2)*floor(w/2)": 1344, + "floor(h/2)": 28, + "h": 56, + "max_length": 160 }, { + "max_length + floor(h/2)*floor(w/2)": 2464, "floor(w/2)": 64, - "h": 128, - "max_length": 160, "w": 128, - "max_length + floor(h/2)*floor(w/2)": 4256, "batch_size": 2, - "floor(h/2)": 64, - "floor(h/2)*floor(w/2)": 4096 + "floor(h/2)*floor(w/2)": 2304, + "floor(h/2)": 36, + "h": 72, + "max_length": 160 } ], "aux_info": {} diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..583a174ae3d1918163bf6759efc7fecca975e3a3 --- /dev/null +++ b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc682d7107b535cf830cb624b6356a250527b7927402514e62b80ae73788206a +size 1191988792 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state index 763f3ef60f988a236481a84d00654a8fd019eec0..04920cf90a5c8e45012babe3022ff41b9061b2e2 100644 Binary files a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state and b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 20fba9bf2a37cf94be60f19560cbd893cff85e15..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:0f9dcac6bcd7f8d0813820fa22f8545c1c24ca3ed1027ddbb58eb7570c589900 -size 88665840 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta deleted file mode 100644 index ba292ba4542c40420893398bec63ecda1b12e84c..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:f30d28d31458cf3dbef90c71e2712d7662846d75d8bc84eff63977f5aa190419 -size 13857510 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt deleted file mode 100644 index 67dd6c0eb71b2fb47fc35866c4897fa323e3ec64..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c48715d6cef2acace7b6c877356caf69503805fe5b0aeaaef0e3ef49ee820e19 -size 56226864 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta deleted file mode 100644 index e3ecadbef195a08fae422a8b435ae4e1144b7ce4..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:841a9934028d04271546d0e60d1cf3341c891b969f2d37a83c68db412092c2d8 -size 8423024 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index ff3d2d441244c7fe123ef488c9f15e683c019589..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:8bb09dad15db7844366902f9ccfb6eec1899a4aec6468534aa383b6e7558d83a -size 71362416 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta deleted file mode 100644 index 0b248c5e75dd5732c1f1056eaaa802d20304c159..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:9d7e39f58775fb06d615ab66f3b2306ffef8989b0a33a1061c80a43bca5f49a9 -size 11026033 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index c92b87826236fe6c6508a3d6a5a0d9f33f4c8e91..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:be690821dfdce89c26023325238f08ed7a612cea18f488c37aef9a683c12ae70 -size 34693488 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta deleted file mode 100644 index bd6d0e8d7252512ec9d6451fd9fe77aa26058690..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:1b6d7065cdae24f1b5a58527c3d6296309272e1e8fa81acff023725c077213e9 -size 4845068 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index f60af88d38a2f9e0873fd45de920b982b2ba8db8..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:f26a82628c3164aab21903363a8d9a76403acc9961732cfa4a2c240a8663ad88 -size 43819440 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta deleted file mode 100644 index 304e593d89ddacbe9f327c0a67891f164fc4aee2..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:e0844d4251c9234b25361830577191a93abaa7b9b19bf8f227aa5cfcfc746119 -size 6278987 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 34b23ce64a7cf3a5b890864f9188bab312922702..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:10e292adcad0248779f498189a69260fb598e3145e5de7596ef53a38fccb37bf -size 56113200 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta deleted file mode 100644 index b49bbe02318e385bc2f8283bd9c61ba72c51cd6e..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:aefa69c73a80f9216a648e6bb25652cd751c2c3de3860155281d2f870612ed5c -size 8417130 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 1dc71cd5281623215e6562d1cba4159d8ac51cea..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:94dd69c0362064e3ea19c306663363b7fcbb6c9fd408f6be7868d039451ece4d -size 71312304 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta deleted file mode 100644 index 562990b6a20f328c323296b30ae9782858629074..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:791e392709a92a2619b974e2e7fb767457adf3010f67481460bbd5aaceba2ad2 -size 11038794 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index 0a711a85af69dc16e089f71b034cf53589cecd33..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:f1312bb124eab3e41f01355ca2d8a243a5a4dced5e299884c76983637ac55fbe -size 43819440 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta deleted file mode 100644 index 9e5cffe5ec4237aa5bc23410be661dc13eff0545..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a5b90e7a36f15d7039556bf434965393f59ca1cc32982b3466264783f219c887 -size 6278987 diff --git a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super b/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-depth/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-depth/dynamic/dd/replaced.onnx b/normal/controlnet-depth/dynamic/dd/replaced.onnx index 3492b5f94e2c5c3bc26efe27d1f62e20cc1df794..3e80491112fa99b9676b70f01a2389f8db833f74 100644 --- a/normal/controlnet-depth/dynamic/dd/replaced.onnx +++ b/normal/controlnet-depth/dynamic/dd/replaced.onnx @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2c71e648861ba3cfa830a938bd679dadc4c49e4defc8b3c3f20b8bcd2c568c10 -size 458658 +oid sha256:378b78bbc26314bfb4561a1c4a10b2303941201b51d38877ba705ca5e68b4b7f +size 458764 diff --git a/normal/controlnet-pose/dynamic/dd/cache/pos_embedprojConv_meta.json b/normal/controlnet-pose/dynamic/dd/cache/pos_embedprojConv_meta.json index 0454e32ca69d8128422d3ec33e79d04c876bc538..cc0aa1a4c9096cb0c4ca4141ba10b1b9ec999030 100644 --- a/normal/controlnet-pose/dynamic/dd/cache/pos_embedprojConv_meta.json +++ b/normal/controlnet-pose/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -341,13 +341,13 @@ "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "/time_text_embed/Cast_output_0.out17_3_3" + "/time_text_embed/Cast_output_0.out17_4_3" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2" + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "attrs": { "input_shape": { @@ -405,7 +405,7 @@ "name": "/time_text_embed/timestep_embedder/act/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "const_args": [ "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1" @@ -461,10 +461,10 @@ "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3" + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4" ], "attrs": { "input_shape": { @@ -522,13 +522,13 @@ "name": "/time_text_embed/text_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "pooled_projections.out17_3_1" + "pooled_projections.out17_4_1" ], "const_args": [ - "time_text_embed.text_embedder.linear_1.weight_5_1_0" + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0" ], "out_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "attrs": { "input_shape": { @@ -586,7 +586,7 @@ "name": "/time_text_embed/text_embedder/act_1/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "const_args": [ "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0" @@ -642,10 +642,10 @@ "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" ], "const_args": [ - "time_text_embed.text_embedder.linear_2.weight_5_1_1" + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1" ], "out_args": [ - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "attrs": { "input_shape": { @@ -703,8 +703,8 @@ "name": "/time_text_embed/Add", "type": "SDAdd", "in_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "const_args": [], "out_args": [ @@ -813,16 +813,16 @@ } }, { - "name": "encoder_hidden_states.out17_3_0_SDCastBf2Bfp", + "name": "encoder_hidden_states.out17_4_0_SDCastBf2Bfp", "type": "SDCastBf2Bfp", "in_args": [ - "encoder_hidden_states.out17_3_0" + "encoder_hidden_states.out17_4_0" ], "const_args": [ - "encoder_hidden_states.out17_3_0_bfp.wts" + "encoder_hidden_states.out17_4_0_bfp.wts" ], "out_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -865,13 +865,13 @@ "name": "/context_embedder/MatMul", "type": "SDGemm_bfp", "in_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "const_args": [ - "onnx::MatMul_1943" + "onnx::MatMul_1943_45_1_0" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -926,16 +926,16 @@ } }, { - "name": "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_SDCastBfp2Bf", + "name": "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_SDCastBfp2Bf", "type": "SDCastBfp2Bf", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "const_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "attrs": { "input_shape": { @@ -978,11 +978,11 @@ "name": "/transformer_blocks.0/norm1_context/norm/LayerNormalization", "type": "SDLayerNorm_bfbfp", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0" ], "out_args": [ "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" @@ -1142,8 +1142,8 @@ "/Add_output_0.out_35_1_3" ], "const_args": [ - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1" ], "out_args": [ "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" @@ -1328,26 +1328,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -1366,17 +1364,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -1434,26 +1434,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -1472,17 +1470,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -1534,26 +1534,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -1572,17 +1570,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -1688,7 +1688,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_1964" + "onnx::MatMul_1964_45_5_1" ], "out_args": [ "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1" @@ -1783,7 +1783,7 @@ "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4_gma" @@ -1867,8 +1867,8 @@ "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0" + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3" ], "out_args": [ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" @@ -2028,7 +2028,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_1963" + "onnx::MatMul_1963_45_5_0" ], "out_args": [ "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0" @@ -2124,7 +2124,7 @@ "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" ], "const_args": [ - "onnx::MatMul_1967" + "onnx::MatMul_1967_45_1_2" ], "out_args": [ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" @@ -2194,10 +2194,10 @@ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" ], "const_args": [ - "onnx::MatMul_1968" + "onnx::MatMul_1968_45_1_4" ], "out_args": [ - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8" + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8" ], "attrs": { "input_shape": { @@ -2256,7 +2256,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ @@ -2424,8 +2424,8 @@ "/transformer_blocks.0/Add_output_0.out10_20" ], "const_args": [ - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0" + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2" ], "out_args": [ "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" @@ -2585,7 +2585,7 @@ "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_21" ], "const_args": [ - "onnx::MatMul_1965" + "onnx::MatMul_1965_45_1_1" ], "out_args": [ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" @@ -2655,10 +2655,10 @@ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" ], "const_args": [ - "onnx::MatMul_1966" + "onnx::MatMul_1966_45_1_3" ], "out_args": [ - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7" + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7" ], "attrs": { "input_shape": { @@ -2717,7 +2717,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_output_0.out10_20" ], "const_args": [ @@ -2802,8 +2802,8 @@ "/transformer_blocks.0/Add_3_output_0.out10_21" ], "const_args": [ - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4" ], "out_args": [ "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" @@ -2963,10 +2963,10 @@ "/transformer_blocks.0/Add_3_output_0.out10_21" ], "const_args": [ - "onnx::MatMul_2088" + "onnx::MatMul_2088_45_0_4" ], "out_args": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13" + "/controlnet_blocks.0/Add_output_0.out17_4_13" ], "attrs": { "input_shape": { @@ -3027,8 +3027,8 @@ "/transformer_blocks.0/Add_7_output_0.out10_1" ], "const_args": [ - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5" ], "out_args": [ "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" @@ -3213,26 +3213,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -3251,17 +3249,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -3319,26 +3319,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -3357,17 +3355,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -3419,26 +3419,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -3457,17 +3455,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -3573,7 +3573,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_1988" + "onnx::MatMul_1988_45_5_2" ], "out_args": [ "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2" @@ -3752,8 +3752,8 @@ "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0" + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6" ], "out_args": [ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" @@ -3913,7 +3913,7 @@ "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" ], "const_args": [ - "onnx::MatMul_1990" + "onnx::MatMul_1990_45_1_5" ], "out_args": [ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" @@ -3983,10 +3983,10 @@ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" ], "const_args": [ - "onnx::MatMul_1991" + "onnx::MatMul_1991_45_1_7" ], "out_args": [ - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12" + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12" ], "attrs": { "input_shape": { @@ -4045,7 +4045,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ @@ -4130,10 +4130,10 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "onnx::MatMul_2089" + "onnx::MatMul_2089_45_0_5" ], "out_args": [ - "/controlnet_blocks.1/Add_output_0.out17_3_22" + "/controlnet_blocks.1/Add_output_0.out17_4_22" ], "attrs": { "input_shape": { @@ -4194,7 +4194,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_1989" + "onnx::MatMul_1989_45_5_3" ], "out_args": [ "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3" @@ -4373,8 +4373,8 @@ "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0" + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7" ], "out_args": [ "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" @@ -4534,7 +4534,7 @@ "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" ], "const_args": [ - "onnx::MatMul_1992" + "onnx::MatMul_1992_45_1_6" ], "out_args": [ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" @@ -4604,10 +4604,10 @@ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" ], "const_args": [ - "onnx::MatMul_1993" + "onnx::MatMul_1993_45_1_8" ], "out_args": [ - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13" + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13" ], "attrs": { "input_shape": { @@ -4666,7 +4666,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ @@ -4751,8 +4751,8 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8" ], "out_args": [ "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" @@ -4912,8 +4912,8 @@ "/transformer_blocks.1/Add_7_output_0.out10_5" ], "const_args": [ - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9" ], "out_args": [ "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" @@ -5098,26 +5098,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -5136,17 +5134,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -5204,26 +5204,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -5242,17 +5240,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -5304,26 +5304,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -5342,17 +5340,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -5458,7 +5458,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_2013" + "onnx::MatMul_2013_45_5_4" ], "out_args": [ "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4" @@ -5637,8 +5637,8 @@ "/transformer_blocks.2/Add_output_0.out10_6" ], "const_args": [ - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0" + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10" ], "out_args": [ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" @@ -5798,7 +5798,7 @@ "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_7" ], "const_args": [ - "onnx::MatMul_2015" + "onnx::MatMul_2015_45_1_9" ], "out_args": [ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" @@ -5868,10 +5868,10 @@ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" ], "const_args": [ - "onnx::MatMul_2016" + "onnx::MatMul_2016_45_1_11" ], "out_args": [ - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17" + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17" ], "attrs": { "input_shape": { @@ -5930,7 +5930,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_output_0.out10_6" ], "const_args": [ @@ -6015,10 +6015,10 @@ "/transformer_blocks.2/Add_3_output_0.out10_7" ], "const_args": [ - "onnx::MatMul_2090" + "onnx::MatMul_2090_45_0_6" ], "out_args": [ - "/controlnet_blocks.2/Add_output_0.out17_3_31" + "/controlnet_blocks.2/Add_output_0.out17_4_31" ], "attrs": { "input_shape": { @@ -6079,7 +6079,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_2014" + "onnx::MatMul_2014_45_5_5" ], "out_args": [ "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5" @@ -6258,8 +6258,8 @@ "/transformer_blocks.2/Add_4_output_0.out10_8" ], "const_args": [ - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0" + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11" ], "out_args": [ "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" @@ -6419,7 +6419,7 @@ "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_9" ], "const_args": [ - "onnx::MatMul_2017" + "onnx::MatMul_2017_45_1_10" ], "out_args": [ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" @@ -6489,10 +6489,10 @@ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" ], "const_args": [ - "onnx::MatMul_2018" + "onnx::MatMul_2018_45_1_12" ], "out_args": [ - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18" + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18" ], "attrs": { "input_shape": { @@ -6551,7 +6551,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_4_output_0.out10_8" ], "const_args": [ @@ -6636,8 +6636,8 @@ "/transformer_blocks.2/Add_3_output_0.out10_7" ], "const_args": [ - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12" ], "out_args": [ "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" @@ -6797,8 +6797,8 @@ "/transformer_blocks.2/Add_7_output_0.out10_9" ], "const_args": [ - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13" ], "out_args": [ "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" @@ -6983,26 +6983,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -7021,17 +7019,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -7089,26 +7089,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -7127,17 +7125,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -7189,26 +7189,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -7227,17 +7225,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -7343,7 +7343,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_2038" + "onnx::MatMul_2038_45_5_6" ], "out_args": [ "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6" @@ -7522,8 +7522,8 @@ "/transformer_blocks.3/Add_output_0.out10_10" ], "const_args": [ - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0" + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14" ], "out_args": [ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" @@ -7683,7 +7683,7 @@ "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_11" ], "const_args": [ - "onnx::MatMul_2040" + "onnx::MatMul_2040_45_1_13" ], "out_args": [ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" @@ -7753,10 +7753,10 @@ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" ], "const_args": [ - "onnx::MatMul_2041" + "onnx::MatMul_2041_45_1_15" ], "out_args": [ - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22" + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22" ], "attrs": { "input_shape": { @@ -7815,7 +7815,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_output_0.out10_10" ], "const_args": [ @@ -7900,10 +7900,10 @@ "/transformer_blocks.3/Add_3_output_0.out10_11" ], "const_args": [ - "onnx::MatMul_2091" + "onnx::MatMul_2091_45_0_7" ], "out_args": [ - "/controlnet_blocks.3/Add_output_0.out17_3_40" + "/controlnet_blocks.3/Add_output_0.out17_4_40" ], "attrs": { "input_shape": { @@ -7964,7 +7964,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_2039" + "onnx::MatMul_2039_45_5_7" ], "out_args": [ "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7" @@ -8143,8 +8143,8 @@ "/transformer_blocks.3/Add_4_output_0.out10_12" ], "const_args": [ - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0" + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15" ], "out_args": [ "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" @@ -8304,7 +8304,7 @@ "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_13" ], "const_args": [ - "onnx::MatMul_2042" + "onnx::MatMul_2042_45_1_14" ], "out_args": [ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" @@ -8374,10 +8374,10 @@ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" ], "const_args": [ - "onnx::MatMul_2043" + "onnx::MatMul_2043_45_1_16" ], "out_args": [ - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23" + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23" ], "attrs": { "input_shape": { @@ -8436,7 +8436,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_4_output_0.out10_12" ], "const_args": [ @@ -8521,8 +8521,8 @@ "/transformer_blocks.3/Add_3_output_0.out10_11" ], "const_args": [ - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16" ], "out_args": [ "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" @@ -8682,8 +8682,8 @@ "/transformer_blocks.3/Add_7_output_0.out10_13" ], "const_args": [ - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17" ], "out_args": [ "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" @@ -8868,26 +8868,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -8906,17 +8904,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -8974,26 +8974,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -9012,17 +9010,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -9074,26 +9074,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -9112,17 +9110,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -9228,7 +9228,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_2063" + "onnx::MatMul_2063_45_5_8" ], "out_args": [ "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8" @@ -9407,8 +9407,8 @@ "/transformer_blocks.4/Add_output_0.out10_14" ], "const_args": [ - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0" + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18" ], "out_args": [ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" @@ -9568,7 +9568,7 @@ "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_15" ], "const_args": [ - "onnx::MatMul_2065" + "onnx::MatMul_2065_45_1_17" ], "out_args": [ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" @@ -9638,10 +9638,10 @@ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" ], "const_args": [ - "onnx::MatMul_2066" + "onnx::MatMul_2066_45_1_19" ], "out_args": [ - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27" + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27" ], "attrs": { "input_shape": { @@ -9700,7 +9700,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_output_0.out10_14" ], "const_args": [ @@ -9785,10 +9785,10 @@ "/transformer_blocks.4/Add_3_output_0.out10_15" ], "const_args": [ - "onnx::MatMul_2092" + "onnx::MatMul_2092_45_0_8" ], "out_args": [ - "/controlnet_blocks.4/Add_output_0.out17_3_49" + "/controlnet_blocks.4/Add_output_0.out17_4_49" ], "attrs": { "input_shape": { @@ -9849,7 +9849,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_2064" + "onnx::MatMul_2064_45_5_9" ], "out_args": [ "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9" @@ -10028,8 +10028,8 @@ "/transformer_blocks.4/Add_4_output_0.out10_16" ], "const_args": [ - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0" + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19" ], "out_args": [ "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" @@ -10189,7 +10189,7 @@ "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_17" ], "const_args": [ - "onnx::MatMul_2067" + "onnx::MatMul_2067_45_1_18" ], "out_args": [ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" @@ -10259,10 +10259,10 @@ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" ], "const_args": [ - "onnx::MatMul_2068" + "onnx::MatMul_2068_45_1_20" ], "out_args": [ - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28" + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28" ], "attrs": { "input_shape": { @@ -10321,7 +10321,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_4_output_0.out10_16" ], "const_args": [ @@ -10406,8 +10406,8 @@ "/transformer_blocks.4/Add_3_output_0.out10_15" ], "const_args": [ - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20" ], "out_args": [ "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" @@ -10567,8 +10567,8 @@ "/transformer_blocks.4/Add_7_output_0.out10_17" ], "const_args": [ - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21" ], "out_args": [ "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" @@ -10753,26 +10753,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -10791,17 +10789,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -10859,26 +10859,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -10897,17 +10895,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -10959,26 +10959,24 @@ "1" ] }, - "bias_enable": { + "weight_shape_0": { "type": "int", "value": [ - "1" + "1536", + "1536" ] }, - "input_shape_0": { - "type": "str", + "weight_shape_1": { + "type": "int", "value": [ - "batch_size", - "max_length", + "1536", "1536" ] }, - "input_shape_1": { - "type": "str", + "bias_enable": { + "type": "int", "value": [ - "batch_size", - "floor(h/2)*floor(w/2)", - "1536" + "1" ] }, "output_shape_0": { @@ -10997,17 +10995,19 @@ "1536" ] }, - "weight_shape_0": { - "type": "int", + "input_shape_0": { + "type": "str", "value": [ - "1536", + "batch_size", + "max_length", "1536" ] }, - "weight_shape_1": { - "type": "int", + "input_shape_1": { + "type": "str", "value": [ - "1536", + "batch_size", + "floor(h/2)*floor(w/2)", "1536" ] }, @@ -11113,7 +11113,7 @@ "/transformer_blocks.5/attn/Reshape_3_output_0.out22_5_bfp.out27_0_5" ], "const_args": [ - "onnx::MatMul_2085" + "onnx::MatMul_2085_45_5_10" ], "out_args": [ "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10" @@ -11292,8 +11292,8 @@ "/transformer_blocks.5/Add_output_0.out10_18" ], "const_args": [ - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0" + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22" ], "out_args": [ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" @@ -11453,7 +11453,7 @@ "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_19" ], "const_args": [ - "onnx::MatMul_2086" + "onnx::MatMul_2086_45_1_21" ], "out_args": [ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" @@ -11523,10 +11523,10 @@ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" ], "const_args": [ - "onnx::MatMul_2087" + "onnx::MatMul_2087_45_1_22" ], "out_args": [ - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31" + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31" ], "attrs": { "input_shape": { @@ -11585,7 +11585,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31", "/transformer_blocks.5/Add_output_0.out10_18" ], "const_args": [ @@ -11670,10 +11670,10 @@ "/transformer_blocks.5/Add_3_output_0.out10_19" ], "const_args": [ - "onnx::MatMul_2093" + "onnx::MatMul_2093_45_0_9" ], "out_args": [ - "/controlnet_blocks.5/Add_output_0.out17_3_57" + "/controlnet_blocks.5/Add_output_0.out17_4_57" ], "attrs": { "input_shape": { @@ -11736,21 +11736,21 @@ "hidden_states_nhwc.out5_0_0", "/pos_embed/Reshape_1_output_0.out_35_1_2", "controlnet_cond_nhwc.out5_0_1", - "/time_text_embed/Cast_output_0.out17_3_3", - "pooled_projections.out17_3_1", - "encoder_hidden_states.out17_3_0" + "/time_text_embed/Cast_output_0.out17_4_3", + "pooled_projections.out17_4_1", + "encoder_hidden_states.out17_4_0" ] }, "out": { "buffer_size": 18432, "xrt_arg_id": 1, "packed_tensors": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13", - "/controlnet_blocks.1/Add_output_0.out17_3_22", - "/controlnet_blocks.2/Add_output_0.out17_3_31", - "/controlnet_blocks.3/Add_output_0.out17_3_40", - "/controlnet_blocks.4/Add_output_0.out17_3_49", - "/controlnet_blocks.5/Add_output_0.out17_3_57" + "/controlnet_blocks.0/Add_output_0.out17_4_13", + "/controlnet_blocks.1/Add_output_0.out17_4_22", + "/controlnet_blocks.2/Add_output_0.out17_4_31", + "/controlnet_blocks.3/Add_output_0.out17_4_40", + "/controlnet_blocks.4/Add_output_0.out17_4_49", + "/controlnet_blocks.5/Add_output_0.out17_4_57" ] }, "scratch": { @@ -11761,17 +11761,17 @@ "/pos_embed/Add_2_output_0.out_35_1_2", "/pos_embed_input/Transpose_output_0.out5_0_1", "/Add_output_0.out_35_1_3", - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3", "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1", - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1", "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2", "/time_text_embed/Add_output_0.out_35_1_4", "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "encoder_hidden_states.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0", + "encoder_hidden_states.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0", "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0", "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1", @@ -11786,13 +11786,13 @@ "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1", "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_7_output_0.out10_1", "/transformer_blocks.0/Add_output_0.out10_20", "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2", "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_21", "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_3_output_0.out10_21", "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4", "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2", @@ -11807,14 +11807,14 @@ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6", "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3", "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_3_output_0.out10_3", "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", "/transformer_blocks.1/Add_4_output_0.out10_4", "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7", "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5", "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_7_output_0.out10_5", "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8", "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_6", @@ -11829,14 +11829,14 @@ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10", "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_7", "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_3_output_0.out10_7", "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5", "/transformer_blocks.2/Add_4_output_0.out10_8", "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11", "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_9", "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_7_output_0.out10_9", "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12", "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_10", @@ -11851,14 +11851,14 @@ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14", "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_11", "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_3_output_0.out10_11", "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7", "/transformer_blocks.3/Add_4_output_0.out10_12", "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15", "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_13", "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_7_output_0.out10_13", "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16", "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_14", @@ -11873,14 +11873,14 @@ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18", "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_15", "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_3_output_0.out10_15", "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9", "/transformer_blocks.4/Add_4_output_0.out10_16", "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19", "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_17", "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_7_output_0.out10_17", "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20", "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_18", @@ -11895,7 +11895,7 @@ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22", "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_19", "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31", "/transformer_blocks.5/Add_3_output_0.out10_19" ] }, @@ -11905,164 +11905,164 @@ "packed_tensors": [ "pos_embed.proj.weight", "pos_embed_input.proj.weight", - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2", + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2", "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1", - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3", - "time_text_embed.text_embedder.linear_1.weight_5_1_0", + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3", + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0", "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0", - "time_text_embed.text_embedder.linear_2.weight_5_1_1", + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1", "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2", - "encoder_hidden_states.out17_3_0_bfp.wts", - "onnx::MatMul_1943", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts", - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0", + "encoder_hidden_states.out17_4_0_bfp.wts", + "onnx::MatMul_1943_45_1_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts", + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0", - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0", + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3", "onnx::MatMul_1947_onnx::MatMul_1944", "onnx::MatMul_1948_onnx::MatMul_1945", "onnx::MatMul_1949_onnx::MatMul_1946", - "onnx::MatMul_1964", + "onnx::MatMul_1964_45_5_1", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4_gma", - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0", + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2", - "onnx::MatMul_1963", - "onnx::MatMul_1967", - "onnx::MatMul_1968", + "onnx::MatMul_1963_45_5_0", + "onnx::MatMul_1967_45_1_2", + "onnx::MatMul_1968_45_1_4", "transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma", - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0", + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1_transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0", - "onnx::MatMul_1965", - "onnx::MatMul_1966", + "onnx::MatMul_1965_45_1_1", + "onnx::MatMul_1966_45_1_3", "transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma", - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0", + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1", - "onnx::MatMul_2088", - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0", + "onnx::MatMul_2088_45_0_4", + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1", "onnx::MatMul_1972_onnx::MatMul_1969", "onnx::MatMul_1973_onnx::MatMul_1970", "onnx::MatMul_1974_onnx::MatMul_1971", - "onnx::MatMul_1988", + "onnx::MatMul_1988_45_5_2", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma", - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0", + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4", - "onnx::MatMul_1990", - "onnx::MatMul_1991", + "onnx::MatMul_1990_45_1_5", + "onnx::MatMul_1991_45_1_7", "transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma", - "onnx::MatMul_2089", - "onnx::MatMul_1989", + "onnx::MatMul_2089_45_0_5", + "onnx::MatMul_1989_45_5_3", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma", - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0", + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4", - "onnx::MatMul_1992", - "onnx::MatMul_1993", + "onnx::MatMul_1992_45_1_6", + "onnx::MatMul_1993_45_1_8", "transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma", - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1", - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1", "onnx::MatMul_1997_onnx::MatMul_1994", "onnx::MatMul_1998_onnx::MatMul_1995", "onnx::MatMul_1999_onnx::MatMul_1996", - "onnx::MatMul_2013", + "onnx::MatMul_2013_45_5_4", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma", - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0", + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4", - "onnx::MatMul_2015", - "onnx::MatMul_2016", + "onnx::MatMul_2015_45_1_9", + "onnx::MatMul_2016_45_1_11", "transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma", - "onnx::MatMul_2090", - "onnx::MatMul_2014", + "onnx::MatMul_2090_45_0_6", + "onnx::MatMul_2014_45_5_5", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma", - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0", + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4", - "onnx::MatMul_2017", - "onnx::MatMul_2018", + "onnx::MatMul_2017_45_1_10", + "onnx::MatMul_2018_45_1_12", "transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma", - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1", - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1", "onnx::MatMul_2022_onnx::MatMul_2019", "onnx::MatMul_2023_onnx::MatMul_2020", "onnx::MatMul_2024_onnx::MatMul_2021", - "onnx::MatMul_2038", + "onnx::MatMul_2038_45_5_6", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma", - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0", + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4", - "onnx::MatMul_2040", - "onnx::MatMul_2041", + "onnx::MatMul_2040_45_1_13", + "onnx::MatMul_2041_45_1_15", "transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma", - "onnx::MatMul_2091", - "onnx::MatMul_2039", + "onnx::MatMul_2091_45_0_7", + "onnx::MatMul_2039_45_5_7", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma", - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0", + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4", - "onnx::MatMul_2042", - "onnx::MatMul_2043", + "onnx::MatMul_2042_45_1_14", + "onnx::MatMul_2043_45_1_16", "transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma", - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1", - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1", "onnx::MatMul_2047_onnx::MatMul_2044", "onnx::MatMul_2048_onnx::MatMul_2045", "onnx::MatMul_2049_onnx::MatMul_2046", - "onnx::MatMul_2063", + "onnx::MatMul_2063_45_5_8", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma", - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0", + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4", - "onnx::MatMul_2065", - "onnx::MatMul_2066", + "onnx::MatMul_2065_45_1_17", + "onnx::MatMul_2066_45_1_19", "transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma", - "onnx::MatMul_2092", - "onnx::MatMul_2064", + "onnx::MatMul_2092_45_0_8", + "onnx::MatMul_2064_45_5_9", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma", - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0", + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4", - "onnx::MatMul_2067", - "onnx::MatMul_2068", + "onnx::MatMul_2067_45_1_18", + "onnx::MatMul_2068_45_1_20", "transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma", - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1", - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21", "transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1", "onnx::MatMul_2072_onnx::MatMul_2069", "onnx::MatMul_2073_onnx::MatMul_2070", "onnx::MatMul_2074_onnx::MatMul_2071", - "onnx::MatMul_2085", + "onnx::MatMul_2085_45_5_10", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma", - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0", + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4", - "onnx::MatMul_2086", - "onnx::MatMul_2087", + "onnx::MatMul_2086_45_1_21", + "onnx::MatMul_2087_45_1_22", "transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma", - "onnx::MatMul_2093" + "onnx::MatMul_2093_45_0_9" ] }, "super_instr": { @@ -12130,7 +12130,7 @@ ], "offset": 3104 }, - "/time_text_embed/Cast_output_0.out17_3_3": { + "/time_text_embed/Cast_output_0.out17_4_3": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12146,7 +12146,7 @@ ], "offset": 3136 }, - "pooled_projections.out17_3_1": { + "pooled_projections.out17_4_1": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12162,7 +12162,7 @@ ], "offset": 3648 }, - "encoder_hidden_states.out17_3_0": { + "encoder_hidden_states.out17_4_0": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12180,7 +12180,7 @@ ], "offset": 7744 }, - "/controlnet_blocks.0/Add_output_0.out17_3_13": { + "/controlnet_blocks.0/Add_output_0.out17_4_13": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12198,7 +12198,7 @@ ], "offset": 0 }, - "/controlnet_blocks.1/Add_output_0.out17_3_22": { + "/controlnet_blocks.1/Add_output_0.out17_4_22": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12216,7 +12216,7 @@ ], "offset": 3072 }, - "/controlnet_blocks.2/Add_output_0.out17_3_31": { + "/controlnet_blocks.2/Add_output_0.out17_4_31": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12234,7 +12234,7 @@ ], "offset": 6144 }, - "/controlnet_blocks.3/Add_output_0.out17_3_40": { + "/controlnet_blocks.3/Add_output_0.out17_4_40": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12252,7 +12252,7 @@ ], "offset": 9216 }, - "/controlnet_blocks.4/Add_output_0.out17_3_49": { + "/controlnet_blocks.4/Add_output_0.out17_4_49": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12270,7 +12270,7 @@ ], "offset": 12288 }, - "/controlnet_blocks.5/Add_output_0.out17_3_57": { + "/controlnet_blocks.5/Add_output_0.out17_4_57": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12360,7 +12360,7 @@ ], "offset": 9216 }, - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3": { + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12392,7 +12392,7 @@ ], "offset": 15360 }, - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4": { + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12408,7 +12408,7 @@ ], "offset": 18432 }, - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1": { + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12440,7 +12440,7 @@ ], "offset": 24576 }, - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2": { + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12488,7 +12488,7 @@ ], "offset": 33792 }, - "encoder_hidden_states.out17_3_0_bfp.out25_0": { + "encoder_hidden_states.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12506,7 +12506,7 @@ ], "offset": 36864 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12524,7 +12524,7 @@ ], "offset": 40960 }, - "/context_embedder/Add_output_0.out17_3_0": { + "/context_embedder/Add_output_0.out17_4_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12800,7 +12800,7 @@ ], "offset": 67072 }, - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8": { + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12908,7 +12908,7 @@ ], "offset": 83968 }, - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7": { + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13184,7 +13184,7 @@ ], "offset": 114688 }, - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12": { + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13310,7 +13310,7 @@ ], "offset": 133120 }, - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13": { + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13586,7 +13586,7 @@ ], "offset": 163840 }, - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17": { + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13712,7 +13712,7 @@ ], "offset": 182272 }, - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18": { + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13988,7 +13988,7 @@ ], "offset": 212992 }, - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22": { + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14114,7 +14114,7 @@ ], "offset": 231424 }, - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23": { + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14390,7 +14390,7 @@ ], "offset": 262144 }, - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27": { + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14516,7 +14516,7 @@ ], "offset": 280576 }, - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28": { + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14792,7 +14792,7 @@ ], "offset": 311296 }, - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31": { + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14854,7 +14854,7 @@ "file_name": "cache/pos_embedprojConv_1.const", "file_size": 135168 }, - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2": { + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14880,7 +14880,7 @@ "file_name": "cache/pos_embedprojConv_3.const", "file_size": 256 }, - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3": { + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14893,7 +14893,7 @@ "file_name": "cache/pos_embedprojConv_4.const", "file_size": 2691072 }, - "time_text_embed.text_embedder.linear_1.weight_5_1_0": { + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14919,7 +14919,7 @@ "file_name": "cache/pos_embedprojConv_6.const", "file_size": 256 }, - "time_text_embed.text_embedder.linear_2.weight_5_1_1": { + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14945,7 +14945,7 @@ "file_name": "cache/pos_embedprojConv_8.const", "file_size": 256 }, - "encoder_hidden_states.out17_3_0_bfp.wts": { + "encoder_hidden_states.out17_4_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14958,7 +14958,7 @@ "file_name": "cache/pos_embedprojConv_9.const", "file_size": 128 }, - "onnx::MatMul_1943": { + "onnx::MatMul_1943_45_1_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14971,7 +14971,7 @@ "file_name": "cache/pos_embedprojConv_10.const", "file_size": 7274496 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14984,7 +14984,7 @@ "file_name": "cache/pos_embedprojConv_11.const", "file_size": 128 }, - "/transformer_blocks.0/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14997,7 +14997,7 @@ "file_name": "cache/pos_embedprojConv_12.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15023,7 +15023,7 @@ "file_name": "cache/pos_embedprojConv_14.const", "file_size": 5382144 }, - "/transformer_blocks.0/norm1/norm/Constant_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15036,7 +15036,7 @@ "file_name": "cache/pos_embedprojConv_15.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15101,7 +15101,7 @@ "file_name": "cache/pos_embedprojConv_20.const", "file_size": 5382144 }, - "onnx::MatMul_1964": { + "onnx::MatMul_1964_45_5_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15127,7 +15127,7 @@ "file_name": "cache/pos_embedprojConv_22.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2_context/Constant_output_0": { + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15140,7 +15140,7 @@ "file_name": "cache/pos_embedprojConv_23.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2_context/Constant_1_output_0": { + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15166,7 +15166,7 @@ "file_name": "cache/pos_embedprojConv_25.const", "file_size": 5382144 }, - "onnx::MatMul_1963": { + "onnx::MatMul_1963_45_5_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15179,7 +15179,7 @@ "file_name": "cache/pos_embedprojConv_26.const", "file_size": 2691072 }, - "onnx::MatMul_1967": { + "onnx::MatMul_1967_45_1_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15192,7 +15192,7 @@ "file_name": "cache/pos_embedprojConv_27.const", "file_size": 10764288 }, - "onnx::MatMul_1968": { + "onnx::MatMul_1968_45_1_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15231,7 +15231,7 @@ "file_name": "cache/pos_embedprojConv_30.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2/Constant_output_0": { + "/transformer_blocks.0/norm2/Constant_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15244,7 +15244,7 @@ "file_name": "cache/pos_embedprojConv_31.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2/Constant_1_output_0": { + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15270,7 +15270,7 @@ "file_name": "cache/pos_embedprojConv_33.const", "file_size": 5382144 }, - "onnx::MatMul_1965": { + "onnx::MatMul_1965_45_1_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15283,7 +15283,7 @@ "file_name": "cache/pos_embedprojConv_34.const", "file_size": 10764288 }, - "onnx::MatMul_1966": { + "onnx::MatMul_1966_45_1_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15309,7 +15309,7 @@ "file_name": "cache/pos_embedprojConv_36.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1/norm/Constant_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15322,7 +15322,7 @@ "file_name": "cache/pos_embedprojConv_37.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15348,7 +15348,7 @@ "file_name": "cache/pos_embedprojConv_39.const", "file_size": 5382144 }, - "onnx::MatMul_2088": { + "onnx::MatMul_2088_45_0_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15361,7 +15361,7 @@ "file_name": "cache/pos_embedprojConv_40.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15374,7 +15374,7 @@ "file_name": "cache/pos_embedprojConv_41.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15439,7 +15439,7 @@ "file_name": "cache/pos_embedprojConv_46.const", "file_size": 5382144 }, - "onnx::MatMul_1988": { + "onnx::MatMul_1988_45_5_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15465,7 +15465,7 @@ "file_name": "cache/pos_embedprojConv_48.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2/Constant_output_0": { + "/transformer_blocks.1/norm2/Constant_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15478,7 +15478,7 @@ "file_name": "cache/pos_embedprojConv_49.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2/Constant_1_output_0": { + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15504,7 +15504,7 @@ "file_name": "cache/pos_embedprojConv_51.const", "file_size": 5382144 }, - "onnx::MatMul_1990": { + "onnx::MatMul_1990_45_1_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15517,7 +15517,7 @@ "file_name": "cache/pos_embedprojConv_52.const", "file_size": 10764288 }, - "onnx::MatMul_1991": { + "onnx::MatMul_1991_45_1_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15543,7 +15543,7 @@ "file_name": "cache/pos_embedprojConv_54.const", "file_size": 2691072 }, - "onnx::MatMul_2089": { + "onnx::MatMul_2089_45_0_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15556,7 +15556,7 @@ "file_name": "cache/pos_embedprojConv_55.const", "file_size": 2691072 }, - "onnx::MatMul_1989": { + "onnx::MatMul_1989_45_5_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15582,7 +15582,7 @@ "file_name": "cache/pos_embedprojConv_57.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2_context/Constant_output_0": { + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15595,7 +15595,7 @@ "file_name": "cache/pos_embedprojConv_58.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2_context/Constant_1_output_0": { + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15621,7 +15621,7 @@ "file_name": "cache/pos_embedprojConv_60.const", "file_size": 5382144 }, - "onnx::MatMul_1992": { + "onnx::MatMul_1992_45_1_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15634,7 +15634,7 @@ "file_name": "cache/pos_embedprojConv_61.const", "file_size": 10764288 }, - "onnx::MatMul_1993": { + "onnx::MatMul_1993_45_1_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15660,7 +15660,7 @@ "file_name": "cache/pos_embedprojConv_63.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm1/norm/Constant_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15673,7 +15673,7 @@ "file_name": "cache/pos_embedprojConv_64.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15699,7 +15699,7 @@ "file_name": "cache/pos_embedprojConv_66.const", "file_size": 5382144 }, - "/transformer_blocks.2/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15712,7 +15712,7 @@ "file_name": "cache/pos_embedprojConv_67.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15777,7 +15777,7 @@ "file_name": "cache/pos_embedprojConv_72.const", "file_size": 5382144 }, - "onnx::MatMul_2013": { + "onnx::MatMul_2013_45_5_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15803,7 +15803,7 @@ "file_name": "cache/pos_embedprojConv_74.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2/Constant_output_0": { + "/transformer_blocks.2/norm2/Constant_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15816,7 +15816,7 @@ "file_name": "cache/pos_embedprojConv_75.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2/Constant_1_output_0": { + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15842,7 +15842,7 @@ "file_name": "cache/pos_embedprojConv_77.const", "file_size": 5382144 }, - "onnx::MatMul_2015": { + "onnx::MatMul_2015_45_1_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15855,7 +15855,7 @@ "file_name": "cache/pos_embedprojConv_78.const", "file_size": 10764288 }, - "onnx::MatMul_2016": { + "onnx::MatMul_2016_45_1_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15881,7 +15881,7 @@ "file_name": "cache/pos_embedprojConv_80.const", "file_size": 2691072 }, - "onnx::MatMul_2090": { + "onnx::MatMul_2090_45_0_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15894,7 +15894,7 @@ "file_name": "cache/pos_embedprojConv_81.const", "file_size": 2691072 }, - "onnx::MatMul_2014": { + "onnx::MatMul_2014_45_5_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15920,7 +15920,7 @@ "file_name": "cache/pos_embedprojConv_83.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2_context/Constant_output_0": { + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15933,7 +15933,7 @@ "file_name": "cache/pos_embedprojConv_84.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2_context/Constant_1_output_0": { + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15959,7 +15959,7 @@ "file_name": "cache/pos_embedprojConv_86.const", "file_size": 5382144 }, - "onnx::MatMul_2017": { + "onnx::MatMul_2017_45_1_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15972,7 +15972,7 @@ "file_name": "cache/pos_embedprojConv_87.const", "file_size": 10764288 }, - "onnx::MatMul_2018": { + "onnx::MatMul_2018_45_1_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15998,7 +15998,7 @@ "file_name": "cache/pos_embedprojConv_89.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm1/norm/Constant_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16011,7 +16011,7 @@ "file_name": "cache/pos_embedprojConv_90.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16037,7 +16037,7 @@ "file_name": "cache/pos_embedprojConv_92.const", "file_size": 5382144 }, - "/transformer_blocks.3/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16050,7 +16050,7 @@ "file_name": "cache/pos_embedprojConv_93.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16115,7 +16115,7 @@ "file_name": "cache/pos_embedprojConv_98.const", "file_size": 5382144 }, - "onnx::MatMul_2038": { + "onnx::MatMul_2038_45_5_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16141,7 +16141,7 @@ "file_name": "cache/pos_embedprojConv_100.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2/Constant_output_0": { + "/transformer_blocks.3/norm2/Constant_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16154,7 +16154,7 @@ "file_name": "cache/pos_embedprojConv_101.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2/Constant_1_output_0": { + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16180,7 +16180,7 @@ "file_name": "cache/pos_embedprojConv_103.const", "file_size": 5382144 }, - "onnx::MatMul_2040": { + "onnx::MatMul_2040_45_1_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16193,7 +16193,7 @@ "file_name": "cache/pos_embedprojConv_104.const", "file_size": 10764288 }, - "onnx::MatMul_2041": { + "onnx::MatMul_2041_45_1_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16219,7 +16219,7 @@ "file_name": "cache/pos_embedprojConv_106.const", "file_size": 2691072 }, - "onnx::MatMul_2091": { + "onnx::MatMul_2091_45_0_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16232,7 +16232,7 @@ "file_name": "cache/pos_embedprojConv_107.const", "file_size": 2691072 }, - "onnx::MatMul_2039": { + "onnx::MatMul_2039_45_5_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16258,7 +16258,7 @@ "file_name": "cache/pos_embedprojConv_109.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2_context/Constant_output_0": { + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16271,7 +16271,7 @@ "file_name": "cache/pos_embedprojConv_110.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2_context/Constant_1_output_0": { + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16297,7 +16297,7 @@ "file_name": "cache/pos_embedprojConv_112.const", "file_size": 5382144 }, - "onnx::MatMul_2042": { + "onnx::MatMul_2042_45_1_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16310,7 +16310,7 @@ "file_name": "cache/pos_embedprojConv_113.const", "file_size": 10764288 }, - "onnx::MatMul_2043": { + "onnx::MatMul_2043_45_1_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16336,7 +16336,7 @@ "file_name": "cache/pos_embedprojConv_115.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm1/norm/Constant_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16349,7 +16349,7 @@ "file_name": "cache/pos_embedprojConv_116.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16375,7 +16375,7 @@ "file_name": "cache/pos_embedprojConv_118.const", "file_size": 5382144 }, - "/transformer_blocks.4/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16388,7 +16388,7 @@ "file_name": "cache/pos_embedprojConv_119.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16453,7 +16453,7 @@ "file_name": "cache/pos_embedprojConv_124.const", "file_size": 5382144 }, - "onnx::MatMul_2063": { + "onnx::MatMul_2063_45_5_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16479,7 +16479,7 @@ "file_name": "cache/pos_embedprojConv_126.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2/Constant_output_0": { + "/transformer_blocks.4/norm2/Constant_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16492,7 +16492,7 @@ "file_name": "cache/pos_embedprojConv_127.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2/Constant_1_output_0": { + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16518,7 +16518,7 @@ "file_name": "cache/pos_embedprojConv_129.const", "file_size": 5382144 }, - "onnx::MatMul_2065": { + "onnx::MatMul_2065_45_1_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16531,7 +16531,7 @@ "file_name": "cache/pos_embedprojConv_130.const", "file_size": 10764288 }, - "onnx::MatMul_2066": { + "onnx::MatMul_2066_45_1_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16557,7 +16557,7 @@ "file_name": "cache/pos_embedprojConv_132.const", "file_size": 2691072 }, - "onnx::MatMul_2092": { + "onnx::MatMul_2092_45_0_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16570,7 +16570,7 @@ "file_name": "cache/pos_embedprojConv_133.const", "file_size": 2691072 }, - "onnx::MatMul_2064": { + "onnx::MatMul_2064_45_5_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16596,7 +16596,7 @@ "file_name": "cache/pos_embedprojConv_135.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2_context/Constant_output_0": { + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16609,7 +16609,7 @@ "file_name": "cache/pos_embedprojConv_136.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2_context/Constant_1_output_0": { + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16635,7 +16635,7 @@ "file_name": "cache/pos_embedprojConv_138.const", "file_size": 5382144 }, - "onnx::MatMul_2067": { + "onnx::MatMul_2067_45_1_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16648,7 +16648,7 @@ "file_name": "cache/pos_embedprojConv_139.const", "file_size": 10764288 }, - "onnx::MatMul_2068": { + "onnx::MatMul_2068_45_1_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16674,7 +16674,7 @@ "file_name": "cache/pos_embedprojConv_141.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm1/norm/Constant_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16687,7 +16687,7 @@ "file_name": "cache/pos_embedprojConv_142.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16713,7 +16713,7 @@ "file_name": "cache/pos_embedprojConv_144.const", "file_size": 5382144 }, - "/transformer_blocks.5/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16726,7 +16726,7 @@ "file_name": "cache/pos_embedprojConv_145.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16791,7 +16791,7 @@ "file_name": "cache/pos_embedprojConv_150.const", "file_size": 5382144 }, - "onnx::MatMul_2085": { + "onnx::MatMul_2085_45_5_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16817,7 +16817,7 @@ "file_name": "cache/pos_embedprojConv_152.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm2/Constant_output_0": { + "/transformer_blocks.5/norm2/Constant_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16830,7 +16830,7 @@ "file_name": "cache/pos_embedprojConv_153.const", "file_size": 3072 }, - "/transformer_blocks.5/norm2/Constant_1_output_0": { + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16856,7 +16856,7 @@ "file_name": "cache/pos_embedprojConv_155.const", "file_size": 5382144 }, - "onnx::MatMul_2086": { + "onnx::MatMul_2086_45_1_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16869,7 +16869,7 @@ "file_name": "cache/pos_embedprojConv_156.const", "file_size": 10764288 }, - "onnx::MatMul_2087": { + "onnx::MatMul_2087_45_1_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16895,7 +16895,7 @@ "file_name": "cache/pos_embedprojConv_158.const", "file_size": 2691072 }, - "onnx::MatMul_2093": { + "onnx::MatMul_2093_45_0_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16912,84 +16912,204 @@ "dynamic_shape_subgraph": true, "dynamic_shape_list": [ { - "floor(h/2)": 32, "max_length + floor(h/2)*floor(w/2)": 1184, "h": 64, + "floor(w/2)": 32, "w": 64, - "floor(h/2)*floor(w/2)": 1024, + "floor(h/2)": 32, + "max_length": 160, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1024 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1760, + "h": 80, + "floor(w/2)": 40, + "w": 80, + "floor(h/2)": 40, + "max_length": 160, "batch_size": 2, + "floor(h/2)*floor(w/2)": 1600 + }, + { + "max_length + floor(h/2)*floor(w/2)": 2464, + "h": 96, + "floor(w/2)": 48, + "w": 96, + "floor(h/2)": 48, "max_length": 160, - "floor(w/2)": 32 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 2304 }, { - "floor(h/2)": 32, - "max_length + floor(h/2)*floor(w/2)": 1696, + "max_length + floor(h/2)*floor(w/2)": 4256, + "h": 128, + "floor(w/2)": 64, + "w": 128, + "floor(h/2)": 64, + "max_length": 160, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 4096 + }, + { + "max_length + floor(h/2)*floor(w/2)": 928, "h": 64, - "w": 96, - "floor(h/2)*floor(w/2)": 1536, + "floor(w/2)": 24, + "w": 48, + "floor(h/2)": 32, + "max_length": 160, "batch_size": 2, + "floor(h/2)*floor(w/2)": 768 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1360, + "h": 80, + "floor(w/2)": 30, + "w": 60, + "floor(h/2)": 40, "max_length": 160, - "floor(w/2)": 48 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1200 }, { - "floor(h/2)": 48, - "max_length + floor(h/2)*floor(w/2)": 1696, + "max_length + floor(h/2)*floor(w/2)": 1888, "h": 96, + "floor(w/2)": 36, + "w": 72, + "floor(h/2)": 48, + "max_length": 160, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1728 + }, + { + "max_length + floor(h/2)*floor(w/2)": 3232, + "h": 128, + "floor(w/2)": 48, + "w": 96, + "floor(h/2)": 64, + "max_length": 160, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 3072 + }, + { + "max_length + floor(h/2)*floor(w/2)": 928, + "h": 48, + "floor(w/2)": 32, "w": 64, - "floor(h/2)*floor(w/2)": 1536, + "floor(h/2)": 24, + "max_length": 160, "batch_size": 2, + "floor(h/2)*floor(w/2)": 768 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1360, + "h": 60, + "floor(w/2)": 40, + "w": 80, + "floor(h/2)": 30, "max_length": 160, - "floor(w/2)": 32 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1200 }, { - "floor(h/2)": 36, - "max_length + floor(h/2)*floor(w/2)": 2464, + "max_length + floor(h/2)*floor(w/2)": 1888, "h": 72, + "floor(w/2)": 48, + "w": 96, + "floor(h/2)": 36, + "max_length": 160, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1728 + }, + { + "max_length + floor(h/2)*floor(w/2)": 3232, + "h": 96, + "floor(w/2)": 64, "w": 128, - "floor(h/2)*floor(w/2)": 2304, + "floor(h/2)": 48, + "max_length": 160, "batch_size": 2, + "floor(h/2)*floor(w/2)": 3072 + }, + { + "max_length + floor(h/2)*floor(w/2)": 736, + "h": 64, + "floor(w/2)": 18, + "w": 36, + "floor(h/2)": 32, "max_length": 160, - "floor(w/2)": 64 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 576 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1120, + "h": 80, + "floor(w/2)": 24, + "w": 48, + "floor(h/2)": 40, + "max_length": 160, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 960 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1504, + "h": 96, + "floor(w/2)": 28, + "w": 56, + "floor(h/2)": 48, + "max_length": 160, + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1344 }, { - "floor(h/2)": 64, "max_length + floor(h/2)*floor(w/2)": 2464, "h": 128, + "floor(w/2)": 36, "w": 72, - "floor(h/2)*floor(w/2)": 2304, - "batch_size": 2, + "floor(h/2)": 64, "max_length": 160, - "floor(w/2)": 36 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 2304 }, { - "floor(h/2)": 48, - "max_length + floor(h/2)*floor(w/2)": 3232, - "h": 96, - "w": 128, - "floor(h/2)*floor(w/2)": 3072, + "max_length + floor(h/2)*floor(w/2)": 736, + "h": 36, + "floor(w/2)": 32, + "w": 64, + "floor(h/2)": 18, + "max_length": 160, "batch_size": 2, + "floor(h/2)*floor(w/2)": 576 + }, + { + "max_length + floor(h/2)*floor(w/2)": 1120, + "h": 48, + "floor(w/2)": 40, + "w": 80, + "floor(h/2)": 24, "max_length": 160, - "floor(w/2)": 64 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 960 }, { - "floor(h/2)": 64, - "max_length + floor(h/2)*floor(w/2)": 3232, - "h": 128, + "max_length + floor(h/2)*floor(w/2)": 1504, + "h": 56, + "floor(w/2)": 48, "w": 96, - "floor(h/2)*floor(w/2)": 3072, - "batch_size": 2, + "floor(h/2)": 28, "max_length": 160, - "floor(w/2)": 48 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 1344 }, { - "floor(h/2)": 64, - "max_length + floor(h/2)*floor(w/2)": 4256, - "h": 128, + "max_length + floor(h/2)*floor(w/2)": 2464, + "h": 72, + "floor(w/2)": 64, "w": 128, - "floor(h/2)*floor(w/2)": 4096, - "batch_size": 2, + "floor(h/2)": 36, "max_length": 160, - "floor(w/2)": 64 + "batch_size": 2, + "floor(h/2)*floor(w/2)": 2304 } ], "aux_info": {} diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..e79634d875c208049fe9e6b601ae150a56406048 --- /dev/null +++ b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebacf73a15e698fc2f13bb4e71660855a624395378c01327a9dfbcba9dd7c2e0 +size 607601907 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state index 2ecff8ad7c60a28365f4f813f06ca20f9aede365..2b63f6d7821bf63ca5012d8950ee588e26825c21 100644 Binary files a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state and b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 1cd68c968b9251ce8eac5df6973a3634f478caf6..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:3f94fdc43a513f1268b7a97d036c4745120b77760c15a6440a030b428a2e6732 -size 44376912 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta deleted file mode 100644 index 3cd8121c152431e60c425a550431767617bc88d5..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:117479f0f5549912d0cf7b0b35bbd955803b4154bac842fa00960ada55ecfed6 -size 6965484 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt deleted file mode 100644 index fbc488e06e58d587ddbe95ac304a43b9672e039c..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:e76618f00493c0f8ccc5a822fb2c9b3f053e7bdfd685486ea40dfc62ab43c2a1 -size 28207344 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta deleted file mode 100644 index cc017f9210eb6dcea0402fcf8f854d6e57cfee19..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:1b76652aa4e4c7687906b5b0d16f542472b804ff5cbe5ed1900e23bc2a0b981b -size 4252202 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index f4f9c18ff0d9440a691aa6b202e5278296ce2ffb..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:838a135e57ed85c5159646a9c75e91d1cd4c5ee6325b5d7a3f260fe847e9e310 -size 35758704 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta deleted file mode 100644 index 8dddae57bb38e493725d0611eefd45800ac1d721..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:2df1ade221625ac105bf72f2ae75035b593fe4b831aafea121ee250d1ae92277 -size 5549611 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index 65909b06a92e9f4a5475a3f83e182ef6a47125f8..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:cc415bb65aff52bf3abaab0ce294d30c180bde5e423873d736123b0450ea46d7 -size 17375184 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta deleted file mode 100644 index c34491a89c0ade70cadcaaf97af63b2e2d658a04..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c7cbdcf0741cb25358ab72ae539885ecf5e10f720078230caad99af812a7a49f -size 2452994 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index 70877fbc8bee4e4f550b87bf47c598422af3019c..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:39ee8075a880e02c8ca14caf2577db95d1d575c686c5724edafcb32da4487979 -size 21936336 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta deleted file mode 100644 index c6639692eab59df500e60ac068f1b99926e4f1ea..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:140bcf8ea8b7a838f8fbe28890a6fa698b262456a2de5cf4c2a74ea9abc2a3e1 -size 3171854 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 51ed659a665e0a73090a2e04c6cc05f9bbae1c7c..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a70cfe7ea32c5b54e7dc323b7e8a69682fd2db68d51252c79dcc528e93cde4aa -size 28093680 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta deleted file mode 100644 index a1b7cffff01a525ed44a08f8704077db94ce1019..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c901f908ebb6b5f7ec2c1961e7eb8ec24ea321ba3347af393fc4d2ba487b5ada -size 4246308 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 3269ad04ae314c9e26aa80479927e1317cd75e0d..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:3de1696ff7a340201a210a6ff2b834c8650ddca0a227dd43bb05e88fe2f434e2 -size 35708592 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta deleted file mode 100644 index 52d8e94f7727d3b247395c69083b53ccbf1d2681..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:e6d8e3584c382637684e4bd5d07536c4e4bae442020ccb745745ede47dbd68c8 -size 5562372 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index 5e6651bb20823b32e68ff822a1b8d05686bf0f5a..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ae6e94e05882935a8a881566a1d72bb162d35146c750e2be29bdf1e3bf41f1ea -size 21936336 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta deleted file mode 100644 index dbc6cbd4b9669d0315aa73d9efb578e59c154900..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:d0ff147cf4f954d45b18130253bb13d197c64daa19b3114ba36a6a5785e489c1 -size 3171854 diff --git a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super b/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-pose/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-pose/dynamic/dd/replaced.onnx b/normal/controlnet-pose/dynamic/dd/replaced.onnx index ff0b78c9f208999cb25f6a118b41575589284a9e..e84c17c4fbebe8f5d410c4029edfa36eb032145a 100644 --- a/normal/controlnet-pose/dynamic/dd/replaced.onnx +++ b/normal/controlnet-pose/dynamic/dd/replaced.onnx @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2d71a794b89f837411a5040121b729334873cf53a977ac9355cb58b9b31d7b53 -size 256940 +oid sha256:645a41d8e5dc9f68a06c1ca66f7daa3f1c7089849d466f93e2eece530b81ef5d +size 257046 diff --git a/normal/controlnet-tile/dynamic/dd/cache/pos_embedprojConv_meta.json b/normal/controlnet-tile/dynamic/dd/cache/pos_embedprojConv_meta.json index d34e99d471b16ace1dface9d9853bb6ea44f540e..0ba578eb4bcd10c361b7932d017e6fb16bcff581 100644 --- a/normal/controlnet-tile/dynamic/dd/cache/pos_embedprojConv_meta.json +++ b/normal/controlnet-tile/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -341,13 +341,13 @@ "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "/time_text_embed/Cast_output_0.out17_3_3" + "/time_text_embed/Cast_output_0.out17_4_3" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2" + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "attrs": { "input_shape": { @@ -405,7 +405,7 @@ "name": "/time_text_embed/timestep_embedder/act/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "const_args": [ "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1" @@ -461,10 +461,10 @@ "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" ], "const_args": [ - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3" + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4" ], "attrs": { "input_shape": { @@ -522,13 +522,13 @@ "name": "/time_text_embed/text_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "pooled_projections.out17_3_1" + "pooled_projections.out17_4_1" ], "const_args": [ - "time_text_embed.text_embedder.linear_1.weight_5_1_0" + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0" ], "out_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "attrs": { "input_shape": { @@ -586,7 +586,7 @@ "name": "/time_text_embed/text_embedder/act_1/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "const_args": [ "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0" @@ -642,10 +642,10 @@ "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" ], "const_args": [ - "time_text_embed.text_embedder.linear_2.weight_5_1_1" + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1" ], "out_args": [ - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "attrs": { "input_shape": { @@ -703,8 +703,8 @@ "name": "/time_text_embed/Add", "type": "SDAdd", "in_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "const_args": [], "out_args": [ @@ -813,16 +813,16 @@ } }, { - "name": "encoder_hidden_states.out17_3_0_SDCastBf2Bfp", + "name": "encoder_hidden_states.out17_4_0_SDCastBf2Bfp", "type": "SDCastBf2Bfp", "in_args": [ - "encoder_hidden_states.out17_3_0" + "encoder_hidden_states.out17_4_0" ], "const_args": [ - "encoder_hidden_states.out17_3_0_bfp.wts" + "encoder_hidden_states.out17_4_0_bfp.wts" ], "out_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -865,13 +865,13 @@ "name": "/context_embedder/MatMul", "type": "SDGemm_bfp", "in_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "const_args": [ - "onnx::MatMul_1943" + "onnx::MatMul_1943_45_1_0" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { @@ -926,16 +926,16 @@ } }, { - "name": "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_SDCastBfp2Bf", + "name": "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_SDCastBfp2Bf", "type": "SDCastBfp2Bf", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "const_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "attrs": { "input_shape": { @@ -978,11 +978,11 @@ "name": "/transformer_blocks.0/norm1_context/norm/LayerNormalization", "type": "SDLayerNorm_bfbfp", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0" ], "out_args": [ "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" @@ -1142,8 +1142,8 @@ "/Add_output_0.out_35_1_3" ], "const_args": [ - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1" ], "out_args": [ "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" @@ -1688,7 +1688,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_1964" + "onnx::MatMul_1964_45_5_1" ], "out_args": [ "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1" @@ -1783,7 +1783,7 @@ "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_2_gma" @@ -1867,8 +1867,8 @@ "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0" + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3" ], "out_args": [ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" @@ -2028,7 +2028,7 @@ "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" ], "const_args": [ - "onnx::MatMul_1967" + "onnx::MatMul_1967_45_1_2" ], "out_args": [ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" @@ -2098,10 +2098,10 @@ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" ], "const_args": [ - "onnx::MatMul_1968" + "onnx::MatMul_1968_45_1_4" ], "out_args": [ - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8" + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8" ], "attrs": { "input_shape": { @@ -2160,7 +2160,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ @@ -2245,7 +2245,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_1963" + "onnx::MatMul_1963_45_5_0" ], "out_args": [ "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0" @@ -2424,8 +2424,8 @@ "/transformer_blocks.0/Add_output_0.out10_20" ], "const_args": [ - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0" + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2" ], "out_args": [ "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" @@ -2585,7 +2585,7 @@ "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_22" ], "const_args": [ - "onnx::MatMul_1965" + "onnx::MatMul_1965_45_1_1" ], "out_args": [ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" @@ -2655,10 +2655,10 @@ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" ], "const_args": [ - "onnx::MatMul_1966" + "onnx::MatMul_1966_45_1_3" ], "out_args": [ - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7" + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7" ], "attrs": { "input_shape": { @@ -2717,7 +2717,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_output_0.out10_20" ], "const_args": [ @@ -2802,8 +2802,8 @@ "/transformer_blocks.0/Add_3_output_0.out10_21" ], "const_args": [ - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4" ], "out_args": [ "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" @@ -2963,10 +2963,10 @@ "/transformer_blocks.0/Add_3_output_0.out10_21" ], "const_args": [ - "onnx::MatMul_2088" + "onnx::MatMul_2088_45_0_4" ], "out_args": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13" + "/controlnet_blocks.0/Add_output_0.out17_4_13" ], "attrs": { "input_shape": { @@ -3027,8 +3027,8 @@ "/transformer_blocks.0/Add_7_output_0.out10_1" ], "const_args": [ - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5" ], "out_args": [ "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" @@ -3573,7 +3573,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_1988" + "onnx::MatMul_1988_45_5_2" ], "out_args": [ "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2" @@ -3752,8 +3752,8 @@ "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0" + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6" ], "out_args": [ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" @@ -3913,7 +3913,7 @@ "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" ], "const_args": [ - "onnx::MatMul_1990" + "onnx::MatMul_1990_45_1_5" ], "out_args": [ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" @@ -3983,10 +3983,10 @@ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10" ], "const_args": [ - "onnx::MatMul_1991" + "onnx::MatMul_1991_45_1_7" ], "out_args": [ - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12" + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12" ], "attrs": { "input_shape": { @@ -4045,7 +4045,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ @@ -4130,10 +4130,10 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "onnx::MatMul_2089" + "onnx::MatMul_2089_45_0_5" ], "out_args": [ - "/controlnet_blocks.1/Add_output_0.out17_3_22" + "/controlnet_blocks.1/Add_output_0.out17_4_22" ], "attrs": { "input_shape": { @@ -4194,7 +4194,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_1989" + "onnx::MatMul_1989_45_5_3" ], "out_args": [ "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3" @@ -4373,8 +4373,8 @@ "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0" + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7" ], "out_args": [ "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" @@ -4534,7 +4534,7 @@ "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" ], "const_args": [ - "onnx::MatMul_1992" + "onnx::MatMul_1992_45_1_6" ], "out_args": [ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" @@ -4604,10 +4604,10 @@ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11" ], "const_args": [ - "onnx::MatMul_1993" + "onnx::MatMul_1993_45_1_8" ], "out_args": [ - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13" + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13" ], "attrs": { "input_shape": { @@ -4666,7 +4666,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ @@ -4751,8 +4751,8 @@ "/transformer_blocks.1/Add_3_output_0.out10_3" ], "const_args": [ - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8" ], "out_args": [ "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" @@ -4912,8 +4912,8 @@ "/transformer_blocks.1/Add_7_output_0.out10_5" ], "const_args": [ - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9" ], "out_args": [ "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" @@ -5458,7 +5458,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_2013" + "onnx::MatMul_2013_45_5_4" ], "out_args": [ "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4" @@ -5637,8 +5637,8 @@ "/transformer_blocks.2/Add_output_0.out10_6" ], "const_args": [ - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0" + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10" ], "out_args": [ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" @@ -5798,7 +5798,7 @@ "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_7" ], "const_args": [ - "onnx::MatMul_2015" + "onnx::MatMul_2015_45_1_9" ], "out_args": [ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" @@ -5868,10 +5868,10 @@ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15" ], "const_args": [ - "onnx::MatMul_2016" + "onnx::MatMul_2016_45_1_11" ], "out_args": [ - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17" + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17" ], "attrs": { "input_shape": { @@ -5930,7 +5930,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_output_0.out10_6" ], "const_args": [ @@ -6015,10 +6015,10 @@ "/transformer_blocks.2/Add_3_output_0.out10_7" ], "const_args": [ - "onnx::MatMul_2090" + "onnx::MatMul_2090_45_0_6" ], "out_args": [ - "/controlnet_blocks.2/Add_output_0.out17_3_31" + "/controlnet_blocks.2/Add_output_0.out17_4_31" ], "attrs": { "input_shape": { @@ -6079,7 +6079,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_2014" + "onnx::MatMul_2014_45_5_5" ], "out_args": [ "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5" @@ -6258,8 +6258,8 @@ "/transformer_blocks.2/Add_4_output_0.out10_8" ], "const_args": [ - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0" + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11" ], "out_args": [ "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" @@ -6419,7 +6419,7 @@ "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_9" ], "const_args": [ - "onnx::MatMul_2017" + "onnx::MatMul_2017_45_1_10" ], "out_args": [ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" @@ -6489,10 +6489,10 @@ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16" ], "const_args": [ - "onnx::MatMul_2018" + "onnx::MatMul_2018_45_1_12" ], "out_args": [ - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18" + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18" ], "attrs": { "input_shape": { @@ -6551,7 +6551,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_4_output_0.out10_8" ], "const_args": [ @@ -6636,8 +6636,8 @@ "/transformer_blocks.2/Add_3_output_0.out10_7" ], "const_args": [ - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12" ], "out_args": [ "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" @@ -6797,8 +6797,8 @@ "/transformer_blocks.2/Add_7_output_0.out10_9" ], "const_args": [ - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13" ], "out_args": [ "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" @@ -7343,7 +7343,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_2038" + "onnx::MatMul_2038_45_5_6" ], "out_args": [ "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6" @@ -7522,8 +7522,8 @@ "/transformer_blocks.3/Add_output_0.out10_10" ], "const_args": [ - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0" + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14" ], "out_args": [ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" @@ -7683,7 +7683,7 @@ "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_11" ], "const_args": [ - "onnx::MatMul_2040" + "onnx::MatMul_2040_45_1_13" ], "out_args": [ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" @@ -7753,10 +7753,10 @@ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20" ], "const_args": [ - "onnx::MatMul_2041" + "onnx::MatMul_2041_45_1_15" ], "out_args": [ - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22" + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22" ], "attrs": { "input_shape": { @@ -7815,7 +7815,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_output_0.out10_10" ], "const_args": [ @@ -7900,10 +7900,10 @@ "/transformer_blocks.3/Add_3_output_0.out10_11" ], "const_args": [ - "onnx::MatMul_2091" + "onnx::MatMul_2091_45_0_7" ], "out_args": [ - "/controlnet_blocks.3/Add_output_0.out17_3_40" + "/controlnet_blocks.3/Add_output_0.out17_4_40" ], "attrs": { "input_shape": { @@ -7964,7 +7964,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_2039" + "onnx::MatMul_2039_45_5_7" ], "out_args": [ "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7" @@ -8143,8 +8143,8 @@ "/transformer_blocks.3/Add_4_output_0.out10_12" ], "const_args": [ - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0" + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15" ], "out_args": [ "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" @@ -8304,7 +8304,7 @@ "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_13" ], "const_args": [ - "onnx::MatMul_2042" + "onnx::MatMul_2042_45_1_14" ], "out_args": [ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" @@ -8374,10 +8374,10 @@ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21" ], "const_args": [ - "onnx::MatMul_2043" + "onnx::MatMul_2043_45_1_16" ], "out_args": [ - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23" + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23" ], "attrs": { "input_shape": { @@ -8436,7 +8436,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_4_output_0.out10_12" ], "const_args": [ @@ -8521,8 +8521,8 @@ "/transformer_blocks.3/Add_3_output_0.out10_11" ], "const_args": [ - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16" ], "out_args": [ "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" @@ -8682,8 +8682,8 @@ "/transformer_blocks.3/Add_7_output_0.out10_13" ], "const_args": [ - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17" ], "out_args": [ "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" @@ -9228,7 +9228,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_2063" + "onnx::MatMul_2063_45_5_8" ], "out_args": [ "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8" @@ -9407,8 +9407,8 @@ "/transformer_blocks.4/Add_output_0.out10_14" ], "const_args": [ - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0" + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18" ], "out_args": [ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" @@ -9568,7 +9568,7 @@ "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_15" ], "const_args": [ - "onnx::MatMul_2065" + "onnx::MatMul_2065_45_1_17" ], "out_args": [ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" @@ -9638,10 +9638,10 @@ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25" ], "const_args": [ - "onnx::MatMul_2066" + "onnx::MatMul_2066_45_1_19" ], "out_args": [ - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27" + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27" ], "attrs": { "input_shape": { @@ -9700,7 +9700,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_output_0.out10_14" ], "const_args": [ @@ -9785,10 +9785,10 @@ "/transformer_blocks.4/Add_3_output_0.out10_15" ], "const_args": [ - "onnx::MatMul_2092" + "onnx::MatMul_2092_45_0_8" ], "out_args": [ - "/controlnet_blocks.4/Add_output_0.out17_3_49" + "/controlnet_blocks.4/Add_output_0.out17_4_49" ], "attrs": { "input_shape": { @@ -9849,7 +9849,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_2064" + "onnx::MatMul_2064_45_5_9" ], "out_args": [ "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9" @@ -10028,8 +10028,8 @@ "/transformer_blocks.4/Add_4_output_0.out10_16" ], "const_args": [ - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0" + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19" ], "out_args": [ "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" @@ -10189,7 +10189,7 @@ "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_17" ], "const_args": [ - "onnx::MatMul_2067" + "onnx::MatMul_2067_45_1_18" ], "out_args": [ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" @@ -10259,10 +10259,10 @@ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26" ], "const_args": [ - "onnx::MatMul_2068" + "onnx::MatMul_2068_45_1_20" ], "out_args": [ - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28" + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28" ], "attrs": { "input_shape": { @@ -10321,7 +10321,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_4_output_0.out10_16" ], "const_args": [ @@ -10406,8 +10406,8 @@ "/transformer_blocks.4/Add_3_output_0.out10_15" ], "const_args": [ - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20" ], "out_args": [ "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" @@ -10567,8 +10567,8 @@ "/transformer_blocks.4/Add_7_output_0.out10_17" ], "const_args": [ - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21" ], "out_args": [ "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" @@ -11113,7 +11113,7 @@ "/transformer_blocks.5/attn/Reshape_3_output_0.out22_5_bfp.out27_0_5" ], "const_args": [ - "onnx::MatMul_2085" + "onnx::MatMul_2085_45_5_10" ], "out_args": [ "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10" @@ -11292,8 +11292,8 @@ "/transformer_blocks.5/Add_output_0.out10_18" ], "const_args": [ - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0" + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22" ], "out_args": [ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" @@ -11453,7 +11453,7 @@ "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_19" ], "const_args": [ - "onnx::MatMul_2086" + "onnx::MatMul_2086_45_1_21" ], "out_args": [ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" @@ -11523,10 +11523,10 @@ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30" ], "const_args": [ - "onnx::MatMul_2087" + "onnx::MatMul_2087_45_1_22" ], "out_args": [ - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31" + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31" ], "attrs": { "input_shape": { @@ -11585,7 +11585,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31", "/transformer_blocks.5/Add_output_0.out10_18" ], "const_args": [ @@ -11670,10 +11670,10 @@ "/transformer_blocks.5/Add_3_output_0.out10_19" ], "const_args": [ - "onnx::MatMul_2093" + "onnx::MatMul_2093_45_0_9" ], "out_args": [ - "/controlnet_blocks.5/Add_output_0.out17_3_57" + "/controlnet_blocks.5/Add_output_0.out17_4_57" ], "attrs": { "input_shape": { @@ -11736,21 +11736,21 @@ "hidden_states_nhwc.out5_0_0", "/pos_embed/Reshape_1_output_0.out_35_1_2", "controlnet_cond_nhwc.out5_0_1", - "/time_text_embed/Cast_output_0.out17_3_3", - "pooled_projections.out17_3_1", - "encoder_hidden_states.out17_3_0" + "/time_text_embed/Cast_output_0.out17_4_3", + "pooled_projections.out17_4_1", + "encoder_hidden_states.out17_4_0" ] }, "out": { "buffer_size": 18432, "xrt_arg_id": 1, "packed_tensors": [ - "/controlnet_blocks.0/Add_output_0.out17_3_13", - "/controlnet_blocks.1/Add_output_0.out17_3_22", - "/controlnet_blocks.2/Add_output_0.out17_3_31", - "/controlnet_blocks.3/Add_output_0.out17_3_40", - "/controlnet_blocks.4/Add_output_0.out17_3_49", - "/controlnet_blocks.5/Add_output_0.out17_3_57" + "/controlnet_blocks.0/Add_output_0.out17_4_13", + "/controlnet_blocks.1/Add_output_0.out17_4_22", + "/controlnet_blocks.2/Add_output_0.out17_4_31", + "/controlnet_blocks.3/Add_output_0.out17_4_40", + "/controlnet_blocks.4/Add_output_0.out17_4_49", + "/controlnet_blocks.5/Add_output_0.out17_4_57" ] }, "scratch": { @@ -11761,17 +11761,17 @@ "/pos_embed/Add_2_output_0.out_35_1_2", "/pos_embed_input/Transpose_output_0.out5_0_1", "/Add_output_0.out_35_1_3", - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3", "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1", - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1", "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2", "/time_text_embed/Add_output_0.out_35_1_4", "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "encoder_hidden_states.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0", + "encoder_hidden_states.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0", "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0", "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1", @@ -11785,14 +11785,14 @@ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3", "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1", "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_7_output_0.out10_1", "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", "/transformer_blocks.0/Add_output_0.out10_20", "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2", "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_22", "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_3_output_0.out10_21", "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_4_bfp.out15_4", "/transformer_blocks.1/norm1/Add_2_output_0.out0_0_4_bfp.out1_2", @@ -11807,14 +11807,14 @@ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6", "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3", "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_10", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_3_output_0.out10_3", "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", "/transformer_blocks.1/Add_4_output_0.out10_4", "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7", "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5", "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_11", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13", "/transformer_blocks.1/Add_7_output_0.out10_5", "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_8_bfp.out15_8", "/transformer_blocks.2/norm1/Add_2_output_0.out0_0_8_bfp.out1_6", @@ -11829,14 +11829,14 @@ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10", "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_7", "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_15", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17", "/transformer_blocks.2/Add_3_output_0.out10_7", "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5", "/transformer_blocks.2/Add_4_output_0.out10_8", "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11", "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_9", "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_16", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18", "/transformer_blocks.2/Add_7_output_0.out10_9", "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_12_bfp.out15_12", "/transformer_blocks.3/norm1/Add_2_output_0.out0_0_12_bfp.out1_10", @@ -11851,14 +11851,14 @@ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14", "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_11", "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_20", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22", "/transformer_blocks.3/Add_3_output_0.out10_11", "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7", "/transformer_blocks.3/Add_4_output_0.out10_12", "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15", "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_13", "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_21", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23", "/transformer_blocks.3/Add_7_output_0.out10_13", "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_16_bfp.out15_16", "/transformer_blocks.4/norm1/Add_2_output_0.out0_0_16_bfp.out1_14", @@ -11873,14 +11873,14 @@ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18", "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_15", "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_25", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27", "/transformer_blocks.4/Add_3_output_0.out10_15", "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9", "/transformer_blocks.4/Add_4_output_0.out10_16", "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19", "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_17", "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_26", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28", "/transformer_blocks.4/Add_7_output_0.out10_17", "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_20_bfp.out15_20", "/transformer_blocks.5/norm1/Add_2_output_0.out0_0_20_bfp.out1_18", @@ -11895,7 +11895,7 @@ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22", "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_19", "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_30", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31", "/transformer_blocks.5/Add_3_output_0.out10_19" ] }, @@ -11905,164 +11905,164 @@ "packed_tensors": [ "pos_embed.proj.weight", "pos_embed_input.proj.weight", - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2", + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2", "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1", - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3", - "time_text_embed.text_embedder.linear_1.weight_5_1_0", + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3", + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0", "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0", - "time_text_embed.text_embedder.linear_2.weight_5_1_1", + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1", "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2", - "encoder_hidden_states.out17_3_0_bfp.wts", - "onnx::MatMul_1943", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts", - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0", + "encoder_hidden_states.out17_4_0_bfp.wts", + "onnx::MatMul_1943_45_1_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts", + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_1_transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_0", - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0", + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_1_transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_0", "onnx::MatMul_1947_onnx::MatMul_1944", "onnx::MatMul_1948_onnx::MatMul_1945", "onnx::MatMul_1949_onnx::MatMul_1946", - "onnx::MatMul_1964", + "onnx::MatMul_1964_45_5_1", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_2_gma", - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0", + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_3_transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_4", - "onnx::MatMul_1967", - "onnx::MatMul_1968", + "onnx::MatMul_1967_45_1_2", + "onnx::MatMul_1968_45_1_4", "transformer_blocks.0.norm1_context.linear.weight_5_1_4_27_0_5_gma", - "onnx::MatMul_1963", + "onnx::MatMul_1963_45_5_0", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_2_gma", - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0", + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_3_transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_4", - "onnx::MatMul_1965", - "onnx::MatMul_1966", + "onnx::MatMul_1965_45_1_1", + "onnx::MatMul_1966_45_1_3", "transformer_blocks.0.norm1.linear.weight_5_1_5_27_1_5_gma", - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0", + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_0_transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_1", - "onnx::MatMul_2088", - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0", + "onnx::MatMul_2088_45_0_4", + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_0_transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_1", "onnx::MatMul_1972_onnx::MatMul_1969", "onnx::MatMul_1973_onnx::MatMul_1970", "onnx::MatMul_1974_onnx::MatMul_1971", - "onnx::MatMul_1988", + "onnx::MatMul_1988_45_5_2", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_2_gma", - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0", + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_3_transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_4", - "onnx::MatMul_1990", - "onnx::MatMul_1991", + "onnx::MatMul_1990_45_1_5", + "onnx::MatMul_1991_45_1_7", "transformer_blocks.1.norm1.linear.weight_5_1_7_27_3_5_gma", - "onnx::MatMul_2089", - "onnx::MatMul_1989", + "onnx::MatMul_2089_45_0_5", + "onnx::MatMul_1989_45_5_3", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_2_gma", - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0", + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_3_transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_4", - "onnx::MatMul_1992", - "onnx::MatMul_1993", + "onnx::MatMul_1992_45_1_6", + "onnx::MatMul_1993_45_1_8", "transformer_blocks.1.norm1_context.linear.weight_5_1_6_27_2_5_gma", - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_0_transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_1", - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_0_transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_1", "onnx::MatMul_1997_onnx::MatMul_1994", "onnx::MatMul_1998_onnx::MatMul_1995", "onnx::MatMul_1999_onnx::MatMul_1996", - "onnx::MatMul_2013", + "onnx::MatMul_2013_45_5_4", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_2_gma", - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0", + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_3_transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_4", - "onnx::MatMul_2015", - "onnx::MatMul_2016", + "onnx::MatMul_2015_45_1_9", + "onnx::MatMul_2016_45_1_11", "transformer_blocks.2.norm1.linear.weight_5_1_9_27_5_5_gma", - "onnx::MatMul_2090", - "onnx::MatMul_2014", + "onnx::MatMul_2090_45_0_6", + "onnx::MatMul_2014_45_5_5", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_2_gma", - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0", + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_3_transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_4", - "onnx::MatMul_2017", - "onnx::MatMul_2018", + "onnx::MatMul_2017_45_1_10", + "onnx::MatMul_2018_45_1_12", "transformer_blocks.2.norm1_context.linear.weight_5_1_8_27_4_5_gma", - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_0_transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_1", - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_0_transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_1", "onnx::MatMul_2022_onnx::MatMul_2019", "onnx::MatMul_2023_onnx::MatMul_2020", "onnx::MatMul_2024_onnx::MatMul_2021", - "onnx::MatMul_2038", + "onnx::MatMul_2038_45_5_6", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_2_gma", - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0", + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_3_transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_4", - "onnx::MatMul_2040", - "onnx::MatMul_2041", + "onnx::MatMul_2040_45_1_13", + "onnx::MatMul_2041_45_1_15", "transformer_blocks.3.norm1.linear.weight_5_1_11_27_7_5_gma", - "onnx::MatMul_2091", - "onnx::MatMul_2039", + "onnx::MatMul_2091_45_0_7", + "onnx::MatMul_2039_45_5_7", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_2_gma", - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0", + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_3_transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_4", - "onnx::MatMul_2042", - "onnx::MatMul_2043", + "onnx::MatMul_2042_45_1_14", + "onnx::MatMul_2043_45_1_16", "transformer_blocks.3.norm1_context.linear.weight_5_1_10_27_6_5_gma", - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_0_transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_1", - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_0_transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_1", "onnx::MatMul_2047_onnx::MatMul_2044", "onnx::MatMul_2048_onnx::MatMul_2045", "onnx::MatMul_2049_onnx::MatMul_2046", - "onnx::MatMul_2063", + "onnx::MatMul_2063_45_5_8", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_2_gma", - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0", + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_3_transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_4", - "onnx::MatMul_2065", - "onnx::MatMul_2066", + "onnx::MatMul_2065_45_1_17", + "onnx::MatMul_2066_45_1_19", "transformer_blocks.4.norm1.linear.weight_5_1_13_27_9_5_gma", - "onnx::MatMul_2092", - "onnx::MatMul_2064", + "onnx::MatMul_2092_45_0_8", + "onnx::MatMul_2064_45_5_9", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_2_gma", - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0", + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_3_transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_4", - "onnx::MatMul_2067", - "onnx::MatMul_2068", + "onnx::MatMul_2067_45_1_18", + "onnx::MatMul_2068_45_1_20", "transformer_blocks.4.norm1_context.linear.weight_5_1_12_27_8_5_gma", - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_0_transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_1", - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21", "transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_0_transformer_blocks.5.norm1_context.linear.weight_5_1_14_27_10_1", "onnx::MatMul_2072_onnx::MatMul_2069", "onnx::MatMul_2073_onnx::MatMul_2070", "onnx::MatMul_2074_onnx::MatMul_2071", - "onnx::MatMul_2085", + "onnx::MatMul_2085_45_5_10", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_2_gma", - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0", + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_3_transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_4", - "onnx::MatMul_2086", - "onnx::MatMul_2087", + "onnx::MatMul_2086_45_1_21", + "onnx::MatMul_2087_45_1_22", "transformer_blocks.5.norm1.linear.weight_5_1_15_27_11_5_gma", - "onnx::MatMul_2093" + "onnx::MatMul_2093_45_0_9" ] }, "super_instr": { @@ -12130,7 +12130,7 @@ ], "offset": 3104 }, - "/time_text_embed/Cast_output_0.out17_3_3": { + "/time_text_embed/Cast_output_0.out17_4_3": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12146,7 +12146,7 @@ ], "offset": 3136 }, - "pooled_projections.out17_3_1": { + "pooled_projections.out17_4_1": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12162,7 +12162,7 @@ ], "offset": 3648 }, - "encoder_hidden_states.out17_3_0": { + "encoder_hidden_states.out17_4_0": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -12180,7 +12180,7 @@ ], "offset": 7744 }, - "/controlnet_blocks.0/Add_output_0.out17_3_13": { + "/controlnet_blocks.0/Add_output_0.out17_4_13": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12198,7 +12198,7 @@ ], "offset": 0 }, - "/controlnet_blocks.1/Add_output_0.out17_3_22": { + "/controlnet_blocks.1/Add_output_0.out17_4_22": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12216,7 +12216,7 @@ ], "offset": 3072 }, - "/controlnet_blocks.2/Add_output_0.out17_3_31": { + "/controlnet_blocks.2/Add_output_0.out17_4_31": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12234,7 +12234,7 @@ ], "offset": 6144 }, - "/controlnet_blocks.3/Add_output_0.out17_3_40": { + "/controlnet_blocks.3/Add_output_0.out17_4_40": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12252,7 +12252,7 @@ ], "offset": 9216 }, - "/controlnet_blocks.4/Add_output_0.out17_3_49": { + "/controlnet_blocks.4/Add_output_0.out17_4_49": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12270,7 +12270,7 @@ ], "offset": 12288 }, - "/controlnet_blocks.5/Add_output_0.out17_3_57": { + "/controlnet_blocks.5/Add_output_0.out17_4_57": { "packed_buffer_label": "out", "xrt_arg_id": 1, "dtype": "bfloat16", @@ -12360,7 +12360,7 @@ ], "offset": 9216 }, - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3": { + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12392,7 +12392,7 @@ ], "offset": 15360 }, - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4": { + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12408,7 +12408,7 @@ ], "offset": 18432 }, - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1": { + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12440,7 +12440,7 @@ ], "offset": 24576 }, - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2": { + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12488,7 +12488,7 @@ ], "offset": 33792 }, - "encoder_hidden_states.out17_3_0_bfp.out25_0": { + "encoder_hidden_states.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12506,7 +12506,7 @@ ], "offset": 36864 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12524,7 +12524,7 @@ ], "offset": 40960 }, - "/context_embedder/Add_output_0.out17_3_0": { + "/context_embedder/Add_output_0.out17_4_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -12782,7 +12782,7 @@ ], "offset": 65536 }, - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8": { + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -12908,7 +12908,7 @@ ], "offset": 83968 }, - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7": { + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13184,7 +13184,7 @@ ], "offset": 114688 }, - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_20_bfp.out25_12": { + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_20_bfp.out25_12": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13310,7 +13310,7 @@ ], "offset": 133120 }, - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_21_bfp.out25_13": { + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_21_bfp.out25_13": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13586,7 +13586,7 @@ ], "offset": 163840 }, - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_29_bfp.out25_17": { + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_29_bfp.out25_17": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13712,7 +13712,7 @@ ], "offset": 182272 }, - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_30_bfp.out25_18": { + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_30_bfp.out25_18": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -13988,7 +13988,7 @@ ], "offset": 212992 }, - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_38_bfp.out25_22": { + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_38_bfp.out25_22": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14114,7 +14114,7 @@ ], "offset": 231424 }, - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_39_bfp.out25_23": { + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_39_bfp.out25_23": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14390,7 +14390,7 @@ ], "offset": 262144 }, - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_47_bfp.out25_27": { + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_47_bfp.out25_27": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14516,7 +14516,7 @@ ], "offset": 280576 }, - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_48_bfp.out25_28": { + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_48_bfp.out25_28": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14792,7 +14792,7 @@ ], "offset": 311296 }, - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_56_bfp.out25_31": { + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_56_bfp.out25_31": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -14854,7 +14854,7 @@ "file_name": "cache/pos_embedprojConv_1.const", "file_size": 135168 }, - "time_text_embed.timestep_embedder.linear_1.weight_5_1_2": { + "time_text_embed.timestep_embedder.linear_1.weight_5_1_2_45_0_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14880,7 +14880,7 @@ "file_name": "cache/pos_embedprojConv_3.const", "file_size": 256 }, - "time_text_embed.timestep_embedder.linear_2.weight_5_1_3": { + "time_text_embed.timestep_embedder.linear_2.weight_5_1_3_45_0_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14893,7 +14893,7 @@ "file_name": "cache/pos_embedprojConv_4.const", "file_size": 2691072 }, - "time_text_embed.text_embedder.linear_1.weight_5_1_0": { + "time_text_embed.text_embedder.linear_1.weight_5_1_0_45_0_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14919,7 +14919,7 @@ "file_name": "cache/pos_embedprojConv_6.const", "file_size": 256 }, - "time_text_embed.text_embedder.linear_2.weight_5_1_1": { + "time_text_embed.text_embedder.linear_2.weight_5_1_1_45_0_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14945,7 +14945,7 @@ "file_name": "cache/pos_embedprojConv_8.const", "file_size": 256 }, - "encoder_hidden_states.out17_3_0_bfp.wts": { + "encoder_hidden_states.out17_4_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14958,7 +14958,7 @@ "file_name": "cache/pos_embedprojConv_9.const", "file_size": 128 }, - "onnx::MatMul_1943": { + "onnx::MatMul_1943_45_1_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -14971,7 +14971,7 @@ "file_name": "cache/pos_embedprojConv_10.const", "file_size": 7274496 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14984,7 +14984,7 @@ "file_name": "cache/pos_embedprojConv_11.const", "file_size": 128 }, - "/transformer_blocks.0/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -14997,7 +14997,7 @@ "file_name": "cache/pos_embedprojConv_12.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15023,7 +15023,7 @@ "file_name": "cache/pos_embedprojConv_14.const", "file_size": 5382144 }, - "/transformer_blocks.0/norm1/norm/Constant_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15036,7 +15036,7 @@ "file_name": "cache/pos_embedprojConv_15.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15101,7 +15101,7 @@ "file_name": "cache/pos_embedprojConv_20.const", "file_size": 5382144 }, - "onnx::MatMul_1964": { + "onnx::MatMul_1964_45_5_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15127,7 +15127,7 @@ "file_name": "cache/pos_embedprojConv_22.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2_context/Constant_output_0": { + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15140,7 +15140,7 @@ "file_name": "cache/pos_embedprojConv_23.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2_context/Constant_1_output_0": { + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15166,7 +15166,7 @@ "file_name": "cache/pos_embedprojConv_25.const", "file_size": 5382144 }, - "onnx::MatMul_1967": { + "onnx::MatMul_1967_45_1_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15179,7 +15179,7 @@ "file_name": "cache/pos_embedprojConv_26.const", "file_size": 10764288 }, - "onnx::MatMul_1968": { + "onnx::MatMul_1968_45_1_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15205,7 +15205,7 @@ "file_name": "cache/pos_embedprojConv_28.const", "file_size": 2691072 }, - "onnx::MatMul_1963": { + "onnx::MatMul_1963_45_5_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15231,7 +15231,7 @@ "file_name": "cache/pos_embedprojConv_30.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2/Constant_output_0": { + "/transformer_blocks.0/norm2/Constant_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15244,7 +15244,7 @@ "file_name": "cache/pos_embedprojConv_31.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2/Constant_1_output_0": { + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15270,7 +15270,7 @@ "file_name": "cache/pos_embedprojConv_33.const", "file_size": 5382144 }, - "onnx::MatMul_1965": { + "onnx::MatMul_1965_45_1_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15283,7 +15283,7 @@ "file_name": "cache/pos_embedprojConv_34.const", "file_size": 10764288 }, - "onnx::MatMul_1966": { + "onnx::MatMul_1966_45_1_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15309,7 +15309,7 @@ "file_name": "cache/pos_embedprojConv_36.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1/norm/Constant_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15322,7 +15322,7 @@ "file_name": "cache/pos_embedprojConv_37.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15348,7 +15348,7 @@ "file_name": "cache/pos_embedprojConv_39.const", "file_size": 5382144 }, - "onnx::MatMul_2088": { + "onnx::MatMul_2088_45_0_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15361,7 +15361,7 @@ "file_name": "cache/pos_embedprojConv_40.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15374,7 +15374,7 @@ "file_name": "cache/pos_embedprojConv_41.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15439,7 +15439,7 @@ "file_name": "cache/pos_embedprojConv_46.const", "file_size": 5382144 }, - "onnx::MatMul_1988": { + "onnx::MatMul_1988_45_5_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15465,7 +15465,7 @@ "file_name": "cache/pos_embedprojConv_48.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2/Constant_output_0": { + "/transformer_blocks.1/norm2/Constant_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15478,7 +15478,7 @@ "file_name": "cache/pos_embedprojConv_49.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2/Constant_1_output_0": { + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15504,7 +15504,7 @@ "file_name": "cache/pos_embedprojConv_51.const", "file_size": 5382144 }, - "onnx::MatMul_1990": { + "onnx::MatMul_1990_45_1_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15517,7 +15517,7 @@ "file_name": "cache/pos_embedprojConv_52.const", "file_size": 10764288 }, - "onnx::MatMul_1991": { + "onnx::MatMul_1991_45_1_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15543,7 +15543,7 @@ "file_name": "cache/pos_embedprojConv_54.const", "file_size": 2691072 }, - "onnx::MatMul_2089": { + "onnx::MatMul_2089_45_0_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15556,7 +15556,7 @@ "file_name": "cache/pos_embedprojConv_55.const", "file_size": 2691072 }, - "onnx::MatMul_1989": { + "onnx::MatMul_1989_45_5_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15582,7 +15582,7 @@ "file_name": "cache/pos_embedprojConv_57.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2_context/Constant_output_0": { + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15595,7 +15595,7 @@ "file_name": "cache/pos_embedprojConv_58.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2_context/Constant_1_output_0": { + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15621,7 +15621,7 @@ "file_name": "cache/pos_embedprojConv_60.const", "file_size": 5382144 }, - "onnx::MatMul_1992": { + "onnx::MatMul_1992_45_1_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15634,7 +15634,7 @@ "file_name": "cache/pos_embedprojConv_61.const", "file_size": 10764288 }, - "onnx::MatMul_1993": { + "onnx::MatMul_1993_45_1_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15660,7 +15660,7 @@ "file_name": "cache/pos_embedprojConv_63.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm1/norm/Constant_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15673,7 +15673,7 @@ "file_name": "cache/pos_embedprojConv_64.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15699,7 +15699,7 @@ "file_name": "cache/pos_embedprojConv_66.const", "file_size": 5382144 }, - "/transformer_blocks.2/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15712,7 +15712,7 @@ "file_name": "cache/pos_embedprojConv_67.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15777,7 +15777,7 @@ "file_name": "cache/pos_embedprojConv_72.const", "file_size": 5382144 }, - "onnx::MatMul_2013": { + "onnx::MatMul_2013_45_5_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15803,7 +15803,7 @@ "file_name": "cache/pos_embedprojConv_74.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2/Constant_output_0": { + "/transformer_blocks.2/norm2/Constant_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15816,7 +15816,7 @@ "file_name": "cache/pos_embedprojConv_75.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2/Constant_1_output_0": { + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15842,7 +15842,7 @@ "file_name": "cache/pos_embedprojConv_77.const", "file_size": 5382144 }, - "onnx::MatMul_2015": { + "onnx::MatMul_2015_45_1_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15855,7 +15855,7 @@ "file_name": "cache/pos_embedprojConv_78.const", "file_size": 10764288 }, - "onnx::MatMul_2016": { + "onnx::MatMul_2016_45_1_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15881,7 +15881,7 @@ "file_name": "cache/pos_embedprojConv_80.const", "file_size": 2691072 }, - "onnx::MatMul_2090": { + "onnx::MatMul_2090_45_0_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15894,7 +15894,7 @@ "file_name": "cache/pos_embedprojConv_81.const", "file_size": 2691072 }, - "onnx::MatMul_2014": { + "onnx::MatMul_2014_45_5_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15920,7 +15920,7 @@ "file_name": "cache/pos_embedprojConv_83.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2_context/Constant_output_0": { + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15933,7 +15933,7 @@ "file_name": "cache/pos_embedprojConv_84.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2_context/Constant_1_output_0": { + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -15959,7 +15959,7 @@ "file_name": "cache/pos_embedprojConv_86.const", "file_size": 5382144 }, - "onnx::MatMul_2017": { + "onnx::MatMul_2017_45_1_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15972,7 +15972,7 @@ "file_name": "cache/pos_embedprojConv_87.const", "file_size": 10764288 }, - "onnx::MatMul_2018": { + "onnx::MatMul_2018_45_1_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -15998,7 +15998,7 @@ "file_name": "cache/pos_embedprojConv_89.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm1/norm/Constant_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16011,7 +16011,7 @@ "file_name": "cache/pos_embedprojConv_90.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16037,7 +16037,7 @@ "file_name": "cache/pos_embedprojConv_92.const", "file_size": 5382144 }, - "/transformer_blocks.3/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16050,7 +16050,7 @@ "file_name": "cache/pos_embedprojConv_93.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16115,7 +16115,7 @@ "file_name": "cache/pos_embedprojConv_98.const", "file_size": 5382144 }, - "onnx::MatMul_2038": { + "onnx::MatMul_2038_45_5_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16141,7 +16141,7 @@ "file_name": "cache/pos_embedprojConv_100.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2/Constant_output_0": { + "/transformer_blocks.3/norm2/Constant_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16154,7 +16154,7 @@ "file_name": "cache/pos_embedprojConv_101.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2/Constant_1_output_0": { + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16180,7 +16180,7 @@ "file_name": "cache/pos_embedprojConv_103.const", "file_size": 5382144 }, - "onnx::MatMul_2040": { + "onnx::MatMul_2040_45_1_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16193,7 +16193,7 @@ "file_name": "cache/pos_embedprojConv_104.const", "file_size": 10764288 }, - "onnx::MatMul_2041": { + "onnx::MatMul_2041_45_1_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16219,7 +16219,7 @@ "file_name": "cache/pos_embedprojConv_106.const", "file_size": 2691072 }, - "onnx::MatMul_2091": { + "onnx::MatMul_2091_45_0_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16232,7 +16232,7 @@ "file_name": "cache/pos_embedprojConv_107.const", "file_size": 2691072 }, - "onnx::MatMul_2039": { + "onnx::MatMul_2039_45_5_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16258,7 +16258,7 @@ "file_name": "cache/pos_embedprojConv_109.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2_context/Constant_output_0": { + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16271,7 +16271,7 @@ "file_name": "cache/pos_embedprojConv_110.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2_context/Constant_1_output_0": { + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16297,7 +16297,7 @@ "file_name": "cache/pos_embedprojConv_112.const", "file_size": 5382144 }, - "onnx::MatMul_2042": { + "onnx::MatMul_2042_45_1_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16310,7 +16310,7 @@ "file_name": "cache/pos_embedprojConv_113.const", "file_size": 10764288 }, - "onnx::MatMul_2043": { + "onnx::MatMul_2043_45_1_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16336,7 +16336,7 @@ "file_name": "cache/pos_embedprojConv_115.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm1/norm/Constant_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16349,7 +16349,7 @@ "file_name": "cache/pos_embedprojConv_116.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16375,7 +16375,7 @@ "file_name": "cache/pos_embedprojConv_118.const", "file_size": 5382144 }, - "/transformer_blocks.4/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16388,7 +16388,7 @@ "file_name": "cache/pos_embedprojConv_119.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16453,7 +16453,7 @@ "file_name": "cache/pos_embedprojConv_124.const", "file_size": 5382144 }, - "onnx::MatMul_2063": { + "onnx::MatMul_2063_45_5_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16479,7 +16479,7 @@ "file_name": "cache/pos_embedprojConv_126.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2/Constant_output_0": { + "/transformer_blocks.4/norm2/Constant_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16492,7 +16492,7 @@ "file_name": "cache/pos_embedprojConv_127.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2/Constant_1_output_0": { + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16518,7 +16518,7 @@ "file_name": "cache/pos_embedprojConv_129.const", "file_size": 5382144 }, - "onnx::MatMul_2065": { + "onnx::MatMul_2065_45_1_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16531,7 +16531,7 @@ "file_name": "cache/pos_embedprojConv_130.const", "file_size": 10764288 }, - "onnx::MatMul_2066": { + "onnx::MatMul_2066_45_1_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16557,7 +16557,7 @@ "file_name": "cache/pos_embedprojConv_132.const", "file_size": 2691072 }, - "onnx::MatMul_2092": { + "onnx::MatMul_2092_45_0_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16570,7 +16570,7 @@ "file_name": "cache/pos_embedprojConv_133.const", "file_size": 2691072 }, - "onnx::MatMul_2064": { + "onnx::MatMul_2064_45_5_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16596,7 +16596,7 @@ "file_name": "cache/pos_embedprojConv_135.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2_context/Constant_output_0": { + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16609,7 +16609,7 @@ "file_name": "cache/pos_embedprojConv_136.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2_context/Constant_1_output_0": { + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16635,7 +16635,7 @@ "file_name": "cache/pos_embedprojConv_138.const", "file_size": 5382144 }, - "onnx::MatMul_2067": { + "onnx::MatMul_2067_45_1_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16648,7 +16648,7 @@ "file_name": "cache/pos_embedprojConv_139.const", "file_size": 10764288 }, - "onnx::MatMul_2068": { + "onnx::MatMul_2068_45_1_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16674,7 +16674,7 @@ "file_name": "cache/pos_embedprojConv_141.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm1/norm/Constant_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16687,7 +16687,7 @@ "file_name": "cache/pos_embedprojConv_142.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16713,7 +16713,7 @@ "file_name": "cache/pos_embedprojConv_144.const", "file_size": 5382144 }, - "/transformer_blocks.5/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16726,7 +16726,7 @@ "file_name": "cache/pos_embedprojConv_145.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16791,7 +16791,7 @@ "file_name": "cache/pos_embedprojConv_150.const", "file_size": 5382144 }, - "onnx::MatMul_2085": { + "onnx::MatMul_2085_45_5_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16817,7 +16817,7 @@ "file_name": "cache/pos_embedprojConv_152.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm2/Constant_output_0": { + "/transformer_blocks.5/norm2/Constant_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16830,7 +16830,7 @@ "file_name": "cache/pos_embedprojConv_153.const", "file_size": 3072 }, - "/transformer_blocks.5/norm2/Constant_1_output_0": { + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -16856,7 +16856,7 @@ "file_name": "cache/pos_embedprojConv_155.const", "file_size": 5382144 }, - "onnx::MatMul_2086": { + "onnx::MatMul_2086_45_1_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16869,7 +16869,7 @@ "file_name": "cache/pos_embedprojConv_156.const", "file_size": 10764288 }, - "onnx::MatMul_2087": { + "onnx::MatMul_2087_45_1_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16895,7 +16895,7 @@ "file_name": "cache/pos_embedprojConv_158.const", "file_size": 2691072 }, - "onnx::MatMul_2093": { + "onnx::MatMul_2093_45_0_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -16912,84 +16912,204 @@ "dynamic_shape_subgraph": true, "dynamic_shape_list": [ { + "floor(h/2)*floor(w/2)": 1024, "batch_size": 2, - "floor(w/2)": 32, - "h": 64, "floor(h/2)": 32, + "h": 64, + "floor(w/2)": 32, "max_length": 160, - "floor(h/2)*floor(w/2)": 1024, "w": 64, "max_length + floor(h/2)*floor(w/2)": 1184 }, { + "floor(h/2)*floor(w/2)": 1600, "batch_size": 2, + "floor(h/2)": 40, + "h": 80, + "floor(w/2)": 40, + "max_length": 160, + "w": 80, + "max_length + floor(h/2)*floor(w/2)": 1760 + }, + { + "floor(h/2)*floor(w/2)": 2304, + "batch_size": 2, + "floor(h/2)": 48, + "h": 96, "floor(w/2)": 48, - "h": 64, + "max_length": 160, + "w": 96, + "max_length + floor(h/2)*floor(w/2)": 2464 + }, + { + "floor(h/2)*floor(w/2)": 4096, + "batch_size": 2, + "floor(h/2)": 64, + "h": 128, + "floor(w/2)": 64, + "max_length": 160, + "w": 128, + "max_length + floor(h/2)*floor(w/2)": 4256 + }, + { + "floor(h/2)*floor(w/2)": 768, + "batch_size": 2, "floor(h/2)": 32, + "h": 64, + "floor(w/2)": 24, + "max_length": 160, + "w": 48, + "max_length + floor(h/2)*floor(w/2)": 928 + }, + { + "floor(h/2)*floor(w/2)": 1200, + "batch_size": 2, + "floor(h/2)": 40, + "h": 80, + "floor(w/2)": 30, + "max_length": 160, + "w": 60, + "max_length + floor(h/2)*floor(w/2)": 1360 + }, + { + "floor(h/2)*floor(w/2)": 1728, + "batch_size": 2, + "floor(h/2)": 48, + "h": 96, + "floor(w/2)": 36, + "max_length": 160, + "w": 72, + "max_length + floor(h/2)*floor(w/2)": 1888 + }, + { + "floor(h/2)*floor(w/2)": 3072, + "batch_size": 2, + "floor(h/2)": 64, + "h": 128, + "floor(w/2)": 48, "max_length": 160, - "floor(h/2)*floor(w/2)": 1536, "w": 96, - "max_length + floor(h/2)*floor(w/2)": 1696 + "max_length + floor(h/2)*floor(w/2)": 3232 }, { + "floor(h/2)*floor(w/2)": 768, "batch_size": 2, + "floor(h/2)": 24, + "h": 48, "floor(w/2)": 32, - "h": 96, - "floor(h/2)": 48, "max_length": 160, - "floor(h/2)*floor(w/2)": 1536, "w": 64, - "max_length + floor(h/2)*floor(w/2)": 1696 + "max_length + floor(h/2)*floor(w/2)": 928 }, { + "floor(h/2)*floor(w/2)": 1200, + "batch_size": 2, + "floor(h/2)": 30, + "h": 60, + "floor(w/2)": 40, + "max_length": 160, + "w": 80, + "max_length + floor(h/2)*floor(w/2)": 1360 + }, + { + "floor(h/2)*floor(w/2)": 1728, "batch_size": 2, - "floor(w/2)": 64, - "h": 72, "floor(h/2)": 36, + "h": 72, + "floor(w/2)": 48, + "max_length": 160, + "w": 96, + "max_length + floor(h/2)*floor(w/2)": 1888 + }, + { + "floor(h/2)*floor(w/2)": 3072, + "batch_size": 2, + "floor(h/2)": 48, + "h": 96, + "floor(w/2)": 64, "max_length": 160, - "floor(h/2)*floor(w/2)": 2304, "w": 128, - "max_length + floor(h/2)*floor(w/2)": 2464 + "max_length + floor(h/2)*floor(w/2)": 3232 }, { + "floor(h/2)*floor(w/2)": 576, "batch_size": 2, - "floor(w/2)": 36, - "h": 128, - "floor(h/2)": 64, + "floor(h/2)": 32, + "h": 64, + "floor(w/2)": 18, + "max_length": 160, + "w": 36, + "max_length + floor(h/2)*floor(w/2)": 736 + }, + { + "floor(h/2)*floor(w/2)": 960, + "batch_size": 2, + "floor(h/2)": 40, + "h": 80, + "floor(w/2)": 24, "max_length": 160, + "w": 48, + "max_length + floor(h/2)*floor(w/2)": 1120 + }, + { + "floor(h/2)*floor(w/2)": 1344, + "batch_size": 2, + "floor(h/2)": 48, + "h": 96, + "floor(w/2)": 28, + "max_length": 160, + "w": 56, + "max_length + floor(h/2)*floor(w/2)": 1504 + }, + { "floor(h/2)*floor(w/2)": 2304, + "batch_size": 2, + "floor(h/2)": 64, + "h": 128, + "floor(w/2)": 36, + "max_length": 160, "w": 72, "max_length + floor(h/2)*floor(w/2)": 2464 }, { + "floor(h/2)*floor(w/2)": 576, "batch_size": 2, - "floor(w/2)": 64, - "h": 96, - "floor(h/2)": 48, + "floor(h/2)": 18, + "h": 36, + "floor(w/2)": 32, "max_length": 160, - "floor(h/2)*floor(w/2)": 3072, - "w": 128, - "max_length + floor(h/2)*floor(w/2)": 3232 + "w": 64, + "max_length + floor(h/2)*floor(w/2)": 736 + }, + { + "floor(h/2)*floor(w/2)": 960, + "batch_size": 2, + "floor(h/2)": 24, + "h": 48, + "floor(w/2)": 40, + "max_length": 160, + "w": 80, + "max_length + floor(h/2)*floor(w/2)": 1120 }, { + "floor(h/2)*floor(w/2)": 1344, "batch_size": 2, + "floor(h/2)": 28, + "h": 56, "floor(w/2)": 48, - "h": 128, - "floor(h/2)": 64, "max_length": 160, - "floor(h/2)*floor(w/2)": 3072, "w": 96, - "max_length + floor(h/2)*floor(w/2)": 3232 + "max_length + floor(h/2)*floor(w/2)": 1504 }, { + "floor(h/2)*floor(w/2)": 2304, "batch_size": 2, + "floor(h/2)": 36, + "h": 72, "floor(w/2)": 64, - "h": 128, - "floor(h/2)": 64, "max_length": 160, - "floor(h/2)*floor(w/2)": 4096, "w": 128, - "max_length + floor(h/2)*floor(w/2)": 4256 + "max_length + floor(h/2)*floor(w/2)": 2464 } ], "aux_info": {} diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..f41c647c8a50539bb058489287058373f4ef1ffe --- /dev/null +++ b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb4b9351216efe64021bbdaeff061e0e3143a17a3e292a21fa5b0596bafa817c +size 607963514 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state index 0b8ba11ebcdcc97b46b68c78dc83733e283152cc..4b0775cddb5db0912f5b251b8ee3ac733e1e32b0 100644 Binary files a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state and b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 210b69574716e373ad0a8592b73fe60a32ce1cc7..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:efef93241ef0440eb28a0d24aa2f8a8f1540fe14c23e34929a92d5bce3524da1 -size 44376912 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta deleted file mode 100644 index 07a9cd5d60f48780481f9759055b4a645a8d5af1..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:0612a781f5c819af469f2da0e36118d3062ec723a8cd93eb8c8792f646a301e6 -size 6962041 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt deleted file mode 100644 index fdd5fb5c3bb1b577c3eb78a76e1136b896554213..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:b06c78deb2fdfd87a0c2e9a1864c129174a2020260521730b78498d247b0f0d9 -size 28207344 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta deleted file mode 100644 index 6eef55a6139a609c540c8664c04421bfbec60d45..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a3f9f13d4b3e4a108131193a45e8c40afc962ea332da3993ceb45af094b885d4 -size 4249864 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index b2c3ef9e42de8195b97f8bb7b32f3e4e9508304b..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:857f57c452e65799b5720697f6fed7be8d9da3cc53bd7928f48a75ceff284795 -size 35758704 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta deleted file mode 100644 index 301ca6c405b5cc0bed6c7a8f431efe1b27e245bc..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:8ce5964d962622637bb2ed164a6a1d0aa5c9a24f0a8f8772db359976232f37f2 -size 5546755 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index 1235f919fe2a05bfde154c71314808a435114dee..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ace6e37cea60577884f27f24f3289d42b767d5a24eb0ef7b3d79eb08067e33e0 -size 17375184 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta deleted file mode 100644 index 18f8f75ea1e6cf1492ea440ae9aa04ec7bdd3452..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c0b185f974f4010a84c7cec52868b5ee82bbde704c61af47af915de105a51dde -size 2453779 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt deleted file mode 100644 index 4256c8e0c85ddae5aa72135e853bea12c6bd1240..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71dd744d457911972c9a6d8767cf8b88a1181b408c64ddc669f9fe0923d7b06b -size 21936336 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta deleted file mode 100644 index 89e496a29f3207f28b42c11d83ca01bdaaa5a4fd..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:ed79c2ed108b7427611a0c8f6a8cc88fae199a934330dba69b86ec56ccb53f8e -size 3170302 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 39a2f908284891ea5fbad81c9efe9cb311718585..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c97e0c7fa5105bb6626363cf1d751fe6e2dc40699cbde469947d7a5cc49cd016 -size 28093680 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta deleted file mode 100644 index ede69006e48ed1d713bc57a7601c4e588a4929e8..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:70cdac5a7185d38570a9f77c1496c40300e836f4812370ea16c8a7cf8c1603f1 -size 4243970 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt deleted file mode 100644 index 131f16723ee3695bdd65b8e21432bf1bfdd4931e..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:6e3d1c23c1909d8043d6ef90e5cb0ee2afb4eaae87352042bd35b04a232d3a5c -size 35708592 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta deleted file mode 100644 index 06b7df4a6e435aa063d8b52bc84750bd10990043..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:51f6b2047bb61a643740069d5da04134c9cd9c619d4246399c1124bb831069e8 -size 5559516 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt deleted file mode 100644 index 8a8b88d9e5876659591c395255db1b43ecd7bf10..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c45cf44190f851142166d563bd034b9c5248b7a3ef551a112ecfd6617483daf8 -size 21936336 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta deleted file mode 100644 index 44a01abcecbf2e828350c66d14ae0bdeda91fa9c..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:46ff69a842df6ee55a8aa893a0c880b20e1b6d255af2ff1dffbcea279d06a42c -size 3170302 diff --git a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super b/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/controlnet-tile/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/controlnet-tile/dynamic/dd/replaced.onnx b/normal/controlnet-tile/dynamic/dd/replaced.onnx index 889a21ba1af8c351d5dc829c8564dcbd5672bfcf..801f7e2d474f6561bafa52f394c884e9d1fcf2dd 100644 --- a/normal/controlnet-tile/dynamic/dd/replaced.onnx +++ b/normal/controlnet-tile/dynamic/dd/replaced.onnx @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9e2fcc6a1dca9347bd1f86808240526ee4d84391816e5510cc3eb4a192f827a6 -size 232631 +oid sha256:7b5467498c1e270dba680f0a261eaec2d8d950b1fed35e291c18053eefae14fd +size 232737 diff --git a/normal/tea_caching/config.json b/normal/tea_caching/config.json new file mode 100644 index 0000000000000000000000000000000000000000..62ae93eb1b8262e0d25745564f67bc3675a3c3e0 --- /dev/null +++ b/normal/tea_caching/config.json @@ -0,0 +1,26 @@ +{ + "sample_size": 128, + "patch_size": 2, + "in_channels": 16, + "num_layers": 18, + "attention_head_dim": 64, + "num_attention_heads": 18, + "joint_attention_dim": 4096, + "caption_projection_dim": 1152, + "pooled_projection_dim": 2048, + "out_channels": 16, + "pos_embed_max_size": 96, + "_use_default_values": [ + "in_channels", + "num_layers", + "caption_projection_dim", + "out_channels", + "joint_attention_dim", + "patch_size", + "pos_embed_max_size", + "num_attention_heads", + "sample_size", + "pooled_projection_dim", + "attention_head_dim" + ] +} \ No newline at end of file diff --git a/normal/tea_caching/dynamic/dd/cache/pos_embedprojConv_meta.json b/normal/tea_caching/dynamic/dd/cache/pos_embedprojConv_meta.json new file mode 100644 index 0000000000000000000000000000000000000000..9c170cae62f7ff75034b91cb41e7ef8f84ef6845 --- /dev/null +++ b/normal/tea_caching/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -0,0 +1,1529 @@ +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 4, + "state_table_updates": [], + "op_list": [ + { + "name": "/pos_embed/proj/Conv", + "type": "SDConv", + "in_args": [ + "hidden_states_nhwc.out8_0_0" + ], + "const_args": [ + "existing_model.pos_embed.proj.weight" + ], + "out_args": [ + "/pos_embed/Transpose_output_0.out8_0_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "w", + "h", + "16" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "floor(w/2)", + "floor(h/2)", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "2", + "2", + "16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/pos_embed/Add_2", + "type": "SDAdd", + "in_args": [ + "/pos_embed/Transpose_output_0.out8_0_0", + "/pos_embed/Reshape_1_output_0.out_13_1_2" + ], + "const_args": [], + "out_args": [ + "/pos_embed/Add_2_output_0.out_13_1_2" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "state_dim1", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/Cast_output_0.out11_4_2" + ], + "const_args": [ + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_15_0_2" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out11_4_2" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/act/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out11_4_2" + ], + "const_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid.weights5_1" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out5_1" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/timestep_embedder/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out5_1" + ], + "const_args": [ + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_15_0_3" + ], + "out_args": [ + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out11_4_3" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "pooled_projections.out11_4_0" + ], + "const_args": [ + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_15_0_0" + ], + "out_args": [ + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out11_4_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "2048" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2048", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/act_1/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out11_4_0" + ], + "const_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid.weights5_0" + ], + "out_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out5_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/text_embedder/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out5_0" + ], + "const_args": [ + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_15_0_1" + ], + "out_args": [ + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out11_4_1" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "1", + "1536" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_text_embed/Add", + "type": "SDAdd", + "in_args": [ + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out11_4_3", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out11_4_1" + ], + "const_args": [], + "out_args": [ + "/time_text_embed/Add_output_0.out_13_1_3" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/norm/LayerNormalization", + "type": "SDLayerNorm_bfbfp", + "in_args": [ + "/pos_embed/Add_2_output_0.out_13_1_2" + ], + "const_args": [ + "/transformer_blocks.0/norm1/norm/Constant_output_0_9_0", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_9_0" + ], + "out_args": [ + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out9_0_bfp.out10_0" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "-1" + ] + }, + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "stash_type": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/silu/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_text_embed/Add_output_0.out_13_1_3" + ], + "const_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid.weights5_2" + ], + "out_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out5_2" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/Add_2", + "type": "SDGemmGemmMulAdd_bfp", + "in_args": [ + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out5_2", + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out9_0_bfp.out10_0" + ], + "const_args": [ + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1" + ], + "out_args": [ + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0" + ], + "attrs": { + "input_shape_0": { + "type": "str", + "value": [ + "batch_4", + "1", + "1536" + ] + }, + "input_shape_1": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "weight_shape_0": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "weight_shape_1": { + "type": "int", + "value": [ + "1536", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0_SDCastBfp2Bf", + "type": "SDCastBfp2Bf", + "in_args": [ + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0" + ], + "const_args": [ + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0_bfp.wts" + ], + "out_args": [ + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "batch_4", + "state_dim1", + "1536" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfp16ebs8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 7712, + "xrt_arg_id": 0, + "packed_tensors": [ + "hidden_states_nhwc.out8_0_0", + "/pos_embed/Reshape_1_output_0.out_13_1_2", + "/time_text_embed/Cast_output_0.out11_4_2", + "pooled_projections.out11_4_0" + ] + }, + "out": { + "buffer_size": 3072, + "xrt_arg_id": 1, + "packed_tensors": [ + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0" + ] + }, + "scratch": { + "buffer_size": 33792, + "xrt_arg_id": 2, + "packed_tensors": [ + "/pos_embed/Transpose_output_0.out8_0_0", + "/pos_embed/Add_2_output_0.out_13_1_2", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out11_4_2", + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out5_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out11_4_3", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out11_4_0", + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out5_0", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out11_4_1", + "/time_text_embed/Add_output_0.out_13_1_3", + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out9_0_bfp.out10_0", + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out5_2", + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0" + ] + }, + "const": { + "buffer_size": 14943104, + "xrt_arg_id": 3, + "packed_tensors": [ + "existing_model.pos_embed.proj.weight", + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_15_0_2", + "/time_text_embed/timestep_embedder/act/Sigmoid.weights5_1", + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_15_0_3", + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_15_0_0", + "/time_text_embed/text_embedder/act_1/Sigmoid.weights5_0", + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_15_0_1", + "/transformer_blocks.0/norm1/norm/Constant_output_0_9_0", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_9_0", + "/transformer_blocks.0/norm1/silu/Sigmoid.weights5_2", + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1", + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0_bfp.wts" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "hidden_states_nhwc.out8_0_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 16 + ], + "size_in_bytes": 32, + "op_tensor_size": 32, + "dynamic_shapes": [ + "batch_4", + "w", + "h", + "False" + ], + "offset": 0 + }, + "/pos_embed/Reshape_1_output_0.out_13_1_2": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "False", + "state_dim1", + "False" + ], + "offset": 32 + }, + "/time_text_embed/Cast_output_0.out11_4_2": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 3104 + }, + "pooled_projections.out11_4_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 2048 + ], + "size_in_bytes": 4096, + "op_tensor_size": 4096, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 3616 + }, + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "state_dim1", + "False" + ], + "offset": 0 + }, + "/pos_embed/Transpose_output_0.out8_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "state_dim1", + "False" + ], + "offset": 0 + }, + "/pos_embed/Add_2_output_0.out_13_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "state_dim1", + "False" + ], + "offset": 3072 + }, + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out11_4_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 6144 + }, + "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out5_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 9216 + }, + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out11_4_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 12288 + }, + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out11_4_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 15360 + }, + "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out5_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 18432 + }, + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out11_4_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 21504 + }, + "/time_text_embed/Add_output_0.out_13_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 24576 + }, + "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out9_0_bfp.out10_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_4", + "state_dim1", + "False" + ], + "offset": 27648 + }, + "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out5_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "dynamic_shapes": [ + "batch_4", + "False" + ], + "offset": 29184 + }, + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "uint8", + "shape": [ + 1, + 1, + 1536 + ], + "size_in_bytes": 1536, + "op_tensor_size": 1536, + "dynamic_shapes": [ + "batch_4", + "state_dim1", + "False" + ], + "offset": 32256 + }, + "existing_model.pos_embed.proj.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135168 + ], + "size_in_bytes": 135168, + "op_tensor_size": 135168, + "offset": 0, + "file_name": "cache/pos_embedprojConv_0.const", + "file_size": 135168 + }, + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_15_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 448512 + ], + "size_in_bytes": 448512, + "op_tensor_size": 448512, + "offset": 135168, + "file_name": "cache/pos_embedprojConv_1.const", + "file_size": 448512 + }, + "/time_text_embed/timestep_embedder/act/Sigmoid.weights5_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 583680, + "file_name": "cache/pos_embedprojConv_2.const", + "file_size": 256 + }, + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_15_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 583936, + "file_name": "cache/pos_embedprojConv_3.const", + "file_size": 2691072 + }, + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_15_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 3588096 + ], + "size_in_bytes": 3588096, + "op_tensor_size": 3588096, + "offset": 3275008, + "file_name": "cache/pos_embedprojConv_4.const", + "file_size": 3588096 + }, + "/time_text_embed/text_embedder/act_1/Sigmoid.weights5_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 6863104, + "file_name": "cache/pos_embedprojConv_5.const", + "file_size": 256 + }, + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_15_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2691072 + ], + "size_in_bytes": 2691072, + "op_tensor_size": 2691072, + "offset": 6863360, + "file_name": "cache/pos_embedprojConv_6.const", + "file_size": 2691072 + }, + "/transformer_blocks.0/norm1/norm/Constant_output_0_9_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 9554432, + "file_name": "cache/pos_embedprojConv_7.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_9_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1536 + ], + "size_in_bytes": 3072, + "op_tensor_size": 3072, + "offset": 9557504, + "file_name": "cache/pos_embedprojConv_8.const", + "file_size": 3072 + }, + "/transformer_blocks.0/norm1/silu/Sigmoid.weights5_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 9560576, + "file_name": "cache/pos_embedprojConv_9.const", + "file_size": 256 + }, + "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "uint8", + "shape": [ + 5382144 + ], + "size_in_bytes": 5382144, + "op_tensor_size": 5382144, + "offset": 9560832, + "file_name": "cache/pos_embedprojConv_10.const", + "file_size": 5382144 + }, + "/transformer_blocks.0/norm1/Add_2_output_0.out3_0_0_bfp.out4_0_bfp.wts": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 64 + ], + "size_in_bytes": 128, + "op_tensor_size": 128, + "offset": 14942976, + "file_name": "cache/pos_embedprojConv_11.const", + "file_size": 128 + } + }, + "dynamic_shape_subgraph": true, + "dynamic_shape_list": [ + { + "h": 64, + "state_dim1": 1024, + "floor(w/2)": 32, + "batch_4": 2, + "floor(h/2)": 32, + "w": 64 + }, + { + "h": 80, + "state_dim1": 1600, + "floor(w/2)": 40, + "batch_4": 2, + "floor(h/2)": 40, + "w": 80 + }, + { + "h": 96, + "state_dim1": 2304, + "floor(w/2)": 48, + "batch_4": 2, + "floor(h/2)": 48, + "w": 96 + }, + { + "h": 128, + "state_dim1": 4096, + "floor(w/2)": 64, + "batch_4": 2, + "floor(h/2)": 64, + "w": 128 + }, + { + "h": 64, + "state_dim1": 768, + "floor(w/2)": 24, + "batch_4": 2, + "floor(h/2)": 32, + "w": 48 + }, + { + "h": 80, + "state_dim1": 1200, + "floor(w/2)": 30, + "batch_4": 2, + "floor(h/2)": 40, + "w": 60 + }, + { + "h": 96, + "state_dim1": 1728, + "floor(w/2)": 36, + "batch_4": 2, + "floor(h/2)": 48, + "w": 72 + }, + { + "h": 128, + "state_dim1": 3072, + "floor(w/2)": 48, + "batch_4": 2, + "floor(h/2)": 64, + "w": 96 + }, + { + "h": 48, + "state_dim1": 768, + "floor(w/2)": 32, + "batch_4": 2, + "floor(h/2)": 24, + "w": 64 + }, + { + "h": 60, + "state_dim1": 1200, + "floor(w/2)": 40, + "batch_4": 2, + "floor(h/2)": 30, + "w": 80 + }, + { + "h": 72, + "state_dim1": 1728, + "floor(w/2)": 48, + "batch_4": 2, + "floor(h/2)": 36, + "w": 96 + }, + { + "h": 96, + "state_dim1": 3072, + "floor(w/2)": 64, + "batch_4": 2, + "floor(h/2)": 48, + "w": 128 + }, + { + "h": 64, + "state_dim1": 576, + "floor(w/2)": 18, + "batch_4": 2, + "floor(h/2)": 32, + "w": 36 + }, + { + "h": 80, + "state_dim1": 960, + "floor(w/2)": 24, + "batch_4": 2, + "floor(h/2)": 40, + "w": 48 + }, + { + "h": 96, + "state_dim1": 1344, + "floor(w/2)": 28, + "batch_4": 2, + "floor(h/2)": 48, + "w": 56 + }, + { + "h": 128, + "state_dim1": 2304, + "floor(w/2)": 36, + "batch_4": 2, + "floor(h/2)": 64, + "w": 72 + }, + { + "h": 36, + "state_dim1": 576, + "floor(w/2)": 32, + "batch_4": 2, + "floor(h/2)": 18, + "w": 64 + }, + { + "h": 48, + "state_dim1": 960, + "floor(w/2)": 40, + "batch_4": 2, + "floor(h/2)": 24, + "w": 80 + }, + { + "h": 56, + "state_dim1": 1344, + "floor(w/2)": 48, + "batch_4": 2, + "floor(h/2)": 28, + "w": 96 + }, + { + "h": 72, + "state_dim1": 2304, + "floor(w/2)": 64, + "batch_4": 2, + "floor(h/2)": 36, + "w": 128 + } + ], + "aux_info": {} +} \ No newline at end of file diff --git a/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..c296488efa14fb91abc1477d654de407d9cd3399 --- /dev/null +++ b/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9fa8994d29cb2177d7c7d9256ea4957f5f6bb1dae105c2ae9657ceaced0e2da +size 26740309 diff --git a/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst b/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst new file mode 100644 index 0000000000000000000000000000000000000000..ea68f2ac4f219305dc6796e91314c37ce5a2bb7d --- /dev/null +++ b/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.fconst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbef93eeb473444ae8eb72c34aad855f4e220df43c6e62fb4574fd3999405156 +size 14942720 diff --git a/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state new file mode 100644 index 0000000000000000000000000000000000000000..f8ff2d9ed4045de7a1c5d36d0246192e8f8c76d1 Binary files /dev/null and b/normal/tea_caching/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/normal/tea_caching/dynamic/dd/replaced.onnx b/normal/tea_caching/dynamic/dd/replaced.onnx new file mode 100644 index 0000000000000000000000000000000000000000..28b4c45920b9969e83ecdc182e311613befebb5a --- /dev/null +++ b/normal/tea_caching/dynamic/dd/replaced.onnx @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74e8f46f68648e9a1cff101647e8ccf15b70dbe04d0f77d7379abcffc9615bf9 +size 501480 diff --git a/normal/tea_caching/dynamic/dd/replaced.onnx.data b/normal/tea_caching/dynamic/dd/replaced.onnx.data new file mode 100644 index 0000000000000000000000000000000000000000..7530b7257c754e617c16e4cfcbfa7ee310fe7811 --- /dev/null +++ b/normal/tea_caching/dynamic/dd/replaced.onnx.data @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:850d66e980fbceb846959c026f2ad96fc5fda8c0402866170acf95dfa534be13 +size 113246208 diff --git a/normal/transformer/dynamic/dd/cache/pos_embedprojConv_meta.json b/normal/transformer/dynamic/dd/cache/pos_embedprojConv_meta.json index b1906b367b04348971a2e9809cc96e3d09b2b5d7..6d67738ab0e121313bef76d99041354068d2cb6a 100644 --- a/normal/transformer/dynamic/dd/cache/pos_embedprojConv_meta.json +++ b/normal/transformer/dynamic/dd/cache/pos_embedprojConv_meta.json @@ -61,7 +61,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "w", "h", "16" @@ -70,7 +70,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "floor(w/2)", "floor(h/2)", "1536" @@ -122,7 +122,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -138,7 +138,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -174,19 +174,19 @@ "name": "/time_text_embed/timestep_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "/time_text_embed/Cast_output_0.out17_3_3" + "/time_text_embed/Cast_output_0.out17_4_3" ], "const_args": [ - "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0" + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_45_0_2" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "256" ] @@ -194,7 +194,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -238,7 +238,7 @@ "name": "/time_text_embed/timestep_embedder/act/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3" + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3" ], "const_args": [ "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1" @@ -250,14 +250,14 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, @@ -294,16 +294,16 @@ "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1" ], "const_args": [ - "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1" + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_45_0_3" ], "out_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -311,7 +311,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -355,19 +355,19 @@ "name": "/time_text_embed/text_embedder/linear_1/Gemm", "type": "SDGemm", "in_args": [ - "pooled_projections.out17_3_1" + "pooled_projections.out17_4_1" ], "const_args": [ - "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2" + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_45_0_0" ], "out_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "2048" ] @@ -375,7 +375,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -419,7 +419,7 @@ "name": "/time_text_embed/text_embedder/act_1/Sigmoid", "type": "SDSilu", "in_args": [ - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1" + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1" ], "const_args": [ "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0" @@ -431,14 +431,14 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, @@ -475,16 +475,16 @@ "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0" ], "const_args": [ - "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3" + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_45_0_1" ], "out_args": [ - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -492,7 +492,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -536,8 +536,8 @@ "name": "/time_text_embed/Add", "type": "SDAdd", "in_args": [ - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2" + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2" ], "const_args": [], "out_args": [ @@ -547,21 +547,21 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, @@ -608,14 +608,14 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "1536" ] }, @@ -646,22 +646,22 @@ } }, { - "name": "encoder_hidden_states.out17_3_0_SDCastBf2Bfp", + "name": "encoder_hidden_states.out17_4_0_SDCastBf2Bfp", "type": "SDCastBf2Bfp", "in_args": [ - "encoder_hidden_states.out17_3_0" + "encoder_hidden_states.out17_4_0" ], "const_args": [ - "encoder_hidden_states.out17_3_0_bfp.wts" + "encoder_hidden_states.out17_4_0_bfp.wts" ], "out_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "4096" ] @@ -669,7 +669,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "4096" ] @@ -698,19 +698,19 @@ "name": "/context_embedder/MatMul", "type": "SDGemm_bfp", "in_args": [ - "encoder_hidden_states.out17_3_0_bfp.out25_0" + "encoder_hidden_states.out17_4_0_bfp.out25_0" ], "const_args": [ - "onnx::MatMul_7905" + "onnx::MatMul_7905_45_1_0" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "4096" ] @@ -718,7 +718,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -759,22 +759,22 @@ } }, { - "name": "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_SDCastBfp2Bf", + "name": "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_SDCastBfp2Bf", "type": "SDCastBfp2Bf", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0" ], "const_args": [ - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts" + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts" ], "out_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -782,7 +782,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -811,11 +811,11 @@ "name": "/transformer_blocks.0/norm1_context/norm/LayerNormalization", "type": "SDLayerNorm_bfbfp", "in_args": [ - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0" ], "out_args": [ "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0" @@ -842,7 +842,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -850,7 +850,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -905,7 +905,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -913,7 +913,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -935,7 +935,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -975,8 +975,8 @@ "/pos_embed/Add_2_output_0.out_35_1_2" ], "const_args": [ - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0" + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1" ], "out_args": [ "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1" @@ -1003,7 +1003,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -1011,7 +1011,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -1066,7 +1066,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -1074,7 +1074,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -1096,7 +1096,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -1161,16 +1161,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1178,7 +1172,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -1200,7 +1194,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1208,11 +1202,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -1267,16 +1267,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1284,7 +1278,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -1306,7 +1300,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1314,11 +1308,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -1367,16 +1367,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1384,7 +1378,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -1406,7 +1400,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1414,11 +1408,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -1475,7 +1475,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -1521,7 +1521,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_7926" + "onnx::MatMul_7926_45_5_1" ], "out_args": [ "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1" @@ -1536,7 +1536,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -1562,7 +1562,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1570,7 +1570,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1616,7 +1616,7 @@ "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", "/transformer_blocks.0/attn/to_add_out/Add_output_0.out6_1_1_bfp.out7_1", - "/context_embedder/Add_output_0.out17_3_0" + "/context_embedder/Add_output_0.out17_4_0" ], "const_args": [ "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2_gma" @@ -1628,7 +1628,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -1636,7 +1636,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1644,7 +1644,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1659,7 +1659,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1700,8 +1700,8 @@ "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0" + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3" ], "out_args": [ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3" @@ -1728,7 +1728,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1736,7 +1736,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1791,7 +1791,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -1799,7 +1799,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1821,7 +1821,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1861,7 +1861,7 @@ "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1" ], "const_args": [ - "onnx::MatMul_7929" + "onnx::MatMul_7929_45_1_2" ], "out_args": [ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" @@ -1870,7 +1870,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1878,7 +1878,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -1931,16 +1931,16 @@ "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6" ], "const_args": [ - "onnx::MatMul_7930" + "onnx::MatMul_7930_45_1_4" ], "out_args": [ - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8" + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -1948,7 +1948,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -1993,7 +1993,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_4_output_0.out10_0" ], "const_args": [ @@ -2006,7 +2006,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -2014,7 +2014,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -2022,7 +2022,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -2037,7 +2037,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -2078,7 +2078,7 @@ "/transformer_blocks.0/attn/Reshape_3_output_0.out22_0_bfp.out27_0_0" ], "const_args": [ - "onnx::MatMul_7925" + "onnx::MatMul_7925_45_5_0" ], "out_args": [ "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0" @@ -2093,7 +2093,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -2119,7 +2119,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2127,7 +2127,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2185,7 +2185,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -2193,7 +2193,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2201,7 +2201,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2216,7 +2216,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2257,8 +2257,8 @@ "/transformer_blocks.0/Add_output_0.out10_92" ], "const_args": [ - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0" + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2" ], "out_args": [ "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2" @@ -2285,7 +2285,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2293,7 +2293,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2348,7 +2348,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -2356,7 +2356,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2378,7 +2378,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2418,7 +2418,7 @@ "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_95" ], "const_args": [ - "onnx::MatMul_7927" + "onnx::MatMul_7927_45_1_1" ], "out_args": [ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" @@ -2427,7 +2427,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2435,7 +2435,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -2488,16 +2488,16 @@ "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5" ], "const_args": [ - "onnx::MatMul_7928" + "onnx::MatMul_7928_45_1_3" ], "out_args": [ - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7" + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -2505,7 +2505,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2550,7 +2550,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_output_0.out10_92" ], "const_args": [ @@ -2563,7 +2563,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -2571,7 +2571,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2579,7 +2579,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2594,7 +2594,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2635,8 +2635,8 @@ "/transformer_blocks.0/Add_7_output_0.out10_1" ], "const_args": [ - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_4" ], "out_args": [ "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_4_bfp.out15_4" @@ -2663,7 +2663,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -2671,7 +2671,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -2724,7 +2724,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2732,7 +2732,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2740,7 +2740,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2779,8 +2779,8 @@ "/Add_output_0.out_35_1_4" ], "const_args": [ - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0" + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_5" ], "out_args": [ "/transformer_blocks.1/norm1/norm/LayerNormalization_output_0.out14_5_bfp.out15_5" @@ -2807,7 +2807,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2815,7 +2815,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2870,7 +2870,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -2878,7 +2878,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2900,7 +2900,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -2950,7 +2950,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -2958,7 +2958,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -2980,7 +2980,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -3045,16 +3045,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -3062,7 +3056,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3084,7 +3078,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -3092,11 +3086,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -3151,16 +3151,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -3168,7 +3162,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3190,7 +3184,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -3198,11 +3192,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -3251,16 +3251,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -3268,7 +3262,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3290,7 +3284,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -3298,11 +3292,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -3359,7 +3359,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -3405,7 +3405,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_7950" + "onnx::MatMul_7950_45_5_2" ], "out_args": [ "/transformer_blocks.1/attn/to_out.0/Add_output_0.out6_1_2_bfp.out7_2" @@ -3420,7 +3420,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -3446,7 +3446,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3454,7 +3454,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3512,7 +3512,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -3520,7 +3520,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3528,7 +3528,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3543,7 +3543,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3584,8 +3584,8 @@ "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0" + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6" ], "out_args": [ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6" @@ -3612,7 +3612,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3620,7 +3620,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3675,7 +3675,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -3683,7 +3683,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3705,7 +3705,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3745,7 +3745,7 @@ "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3" ], "const_args": [ - "onnx::MatMul_7952" + "onnx::MatMul_7952_45_1_5" ], "out_args": [ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_9" @@ -3754,7 +3754,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3762,7 +3762,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -3815,16 +3815,16 @@ "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_9" ], "const_args": [ - "onnx::MatMul_7953" + "onnx::MatMul_7953_45_1_7" ], "out_args": [ - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_19_bfp.out25_11" + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -3832,7 +3832,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3877,7 +3877,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_19_bfp.out25_11", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11", "/transformer_blocks.1/Add_output_0.out10_2" ], "const_args": [ @@ -3890,7 +3890,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -3898,7 +3898,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3906,7 +3906,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3921,7 +3921,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3970,7 +3970,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3978,7 +3978,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -3986,7 +3986,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -4025,7 +4025,7 @@ "/transformer_blocks.1/attn/Reshape_3_output_0.out22_1_bfp.out27_0_1" ], "const_args": [ - "onnx::MatMul_7951" + "onnx::MatMul_7951_45_5_3" ], "out_args": [ "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3" @@ -4040,7 +4040,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -4066,7 +4066,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4074,7 +4074,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4132,7 +4132,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -4140,7 +4140,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4148,7 +4148,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4163,7 +4163,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4204,8 +4204,8 @@ "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0" + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7" ], "out_args": [ "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7" @@ -4232,7 +4232,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4240,7 +4240,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4295,7 +4295,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -4303,7 +4303,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4325,7 +4325,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4365,7 +4365,7 @@ "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5" ], "const_args": [ - "onnx::MatMul_7954" + "onnx::MatMul_7954_45_1_6" ], "out_args": [ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_10" @@ -4374,7 +4374,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4382,7 +4382,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -4435,16 +4435,16 @@ "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_10" ], "const_args": [ - "onnx::MatMul_7955" + "onnx::MatMul_7955_45_1_8" ], "out_args": [ - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_20_bfp.out25_12" + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -4452,7 +4452,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4497,7 +4497,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_4_output_0.out10_4" ], "const_args": [ @@ -4510,7 +4510,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -4518,7 +4518,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4526,7 +4526,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4541,7 +4541,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4582,8 +4582,8 @@ "/Add_1_output_0.out_35_1_5" ], "const_args": [ - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_9" ], "out_args": [ "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_9_bfp.out15_9" @@ -4610,7 +4610,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -4618,7 +4618,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -4663,8 +4663,8 @@ "/transformer_blocks.1/Add_7_output_0.out10_5" ], "const_args": [ - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_8" ], "out_args": [ "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_8_bfp.out15_8" @@ -4691,7 +4691,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4699,7 +4699,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4754,7 +4754,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -4762,7 +4762,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -4784,7 +4784,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -4834,7 +4834,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -4842,7 +4842,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4864,7 +4864,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4929,16 +4929,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4946,7 +4940,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -4968,7 +4962,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -4976,11 +4970,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -5035,16 +5035,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5052,7 +5046,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5074,7 +5068,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5082,11 +5076,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -5135,16 +5135,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5152,7 +5146,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5174,7 +5168,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5182,11 +5176,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -5243,7 +5243,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -5289,7 +5289,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_7976" + "onnx::MatMul_7976_45_5_5" ], "out_args": [ "/transformer_blocks.2/attn/to_add_out/Add_output_0.out6_1_5_bfp.out7_5" @@ -5304,7 +5304,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -5330,7 +5330,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5338,7 +5338,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5396,7 +5396,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -5404,7 +5404,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5412,7 +5412,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5427,7 +5427,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5468,7 +5468,7 @@ "/transformer_blocks.2/attn/Reshape_3_output_0.out22_2_bfp.out27_0_2" ], "const_args": [ - "onnx::MatMul_7975" + "onnx::MatMul_7975_45_5_4" ], "out_args": [ "/transformer_blocks.2/attn/to_out.0/Add_output_0.out6_1_4_bfp.out7_4" @@ -5483,7 +5483,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -5509,7 +5509,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5517,7 +5517,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5575,7 +5575,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -5583,7 +5583,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5591,7 +5591,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5606,7 +5606,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5647,8 +5647,8 @@ "/transformer_blocks.2/Add_output_0.out10_46" ], "const_args": [ - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0" + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10" ], "out_args": [ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10" @@ -5675,7 +5675,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5683,7 +5683,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5738,7 +5738,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -5746,7 +5746,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5768,7 +5768,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5808,7 +5808,7 @@ "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47" ], "const_args": [ - "onnx::MatMul_7977" + "onnx::MatMul_7977_45_1_9" ], "out_args": [ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_13" @@ -5817,7 +5817,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5825,7 +5825,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -5878,16 +5878,16 @@ "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_13" ], "const_args": [ - "onnx::MatMul_7978" + "onnx::MatMul_7978_45_1_11" ], "out_args": [ - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_27_bfp.out25_15" + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -5895,7 +5895,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -5942,8 +5942,8 @@ "/transformer_blocks.2/Add_4_output_0.out10_48" ], "const_args": [ - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0" + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11" ], "out_args": [ "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11" @@ -5970,7 +5970,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -5978,7 +5978,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6033,7 +6033,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -6041,7 +6041,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6063,7 +6063,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6103,7 +6103,7 @@ "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49" ], "const_args": [ - "onnx::MatMul_7979" + "onnx::MatMul_7979_45_1_10" ], "out_args": [ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_14" @@ -6112,7 +6112,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6120,7 +6120,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -6173,16 +6173,16 @@ "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_14" ], "const_args": [ - "onnx::MatMul_7980" + "onnx::MatMul_7980_45_1_12" ], "out_args": [ - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_28_bfp.out25_16" + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -6190,7 +6190,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6235,7 +6235,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_28_bfp.out25_16", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16", "/transformer_blocks.2/Add_4_output_0.out10_48" ], "const_args": [ @@ -6248,7 +6248,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -6256,7 +6256,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6264,7 +6264,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6279,7 +6279,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6320,8 +6320,8 @@ "/transformer_blocks.2/Add_7_output_0.out10_49" ], "const_args": [ - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_12" ], "out_args": [ "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_12_bfp.out15_12" @@ -6348,7 +6348,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6356,7 +6356,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6399,7 +6399,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_27_bfp.out25_15", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15", "/transformer_blocks.2/Add_output_0.out10_46" ], "const_args": [ @@ -6412,7 +6412,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -6420,7 +6420,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6428,7 +6428,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6443,7 +6443,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6492,7 +6492,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6500,7 +6500,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6508,7 +6508,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6547,8 +6547,8 @@ "/Add_2_output_0.out_35_1_6" ], "const_args": [ - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0" + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_13" ], "out_args": [ "/transformer_blocks.3/norm1/norm/LayerNormalization_output_0.out14_13_bfp.out15_13" @@ -6575,7 +6575,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6583,7 +6583,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6638,7 +6638,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -6646,7 +6646,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6668,7 +6668,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6718,7 +6718,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -6726,7 +6726,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6748,7 +6748,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6813,16 +6813,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6830,7 +6824,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6852,7 +6846,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6860,11 +6854,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -6919,16 +6919,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6936,7 +6930,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -6958,7 +6952,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -6966,11 +6960,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -7019,16 +7019,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -7036,7 +7030,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7058,7 +7052,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -7066,11 +7060,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -7127,7 +7127,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -7173,7 +7173,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_8001" + "onnx::MatMul_8001_45_5_7" ], "out_args": [ "/transformer_blocks.3/attn/to_add_out/Add_output_0.out6_1_7_bfp.out7_7" @@ -7188,7 +7188,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -7214,7 +7214,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -7222,7 +7222,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -7280,7 +7280,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -7288,7 +7288,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -7296,7 +7296,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -7311,7 +7311,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -7352,7 +7352,7 @@ "/transformer_blocks.3/attn/Reshape_3_output_0.out22_3_bfp.out27_0_3" ], "const_args": [ - "onnx::MatMul_8000" + "onnx::MatMul_8000_45_5_6" ], "out_args": [ "/transformer_blocks.3/attn/to_out.0/Add_output_0.out6_1_6_bfp.out7_6" @@ -7367,7 +7367,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -7393,7 +7393,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7401,7 +7401,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7459,7 +7459,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -7467,7 +7467,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7475,7 +7475,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7490,7 +7490,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7531,8 +7531,8 @@ "/transformer_blocks.3/Add_output_0.out10_64" ], "const_args": [ - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0" + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14" ], "out_args": [ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14" @@ -7559,7 +7559,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7567,7 +7567,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7622,7 +7622,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -7630,7 +7630,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7652,7 +7652,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7692,7 +7692,7 @@ "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_66" ], "const_args": [ - "onnx::MatMul_8002" + "onnx::MatMul_8002_45_1_13" ], "out_args": [ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_17" @@ -7701,7 +7701,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7709,7 +7709,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -7762,16 +7762,16 @@ "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_17" ], "const_args": [ - "onnx::MatMul_8003" + "onnx::MatMul_8003_45_1_15" ], "out_args": [ - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_35_bfp.out25_19" + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -7779,7 +7779,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7824,7 +7824,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_35_bfp.out25_19", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19", "/transformer_blocks.3/Add_output_0.out10_64" ], "const_args": [ @@ -7837,7 +7837,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -7845,7 +7845,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7853,7 +7853,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7868,7 +7868,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7917,7 +7917,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7925,7 +7925,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7933,7 +7933,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -7972,8 +7972,8 @@ "/transformer_blocks.3/Add_4_output_0.out10_66" ], "const_args": [ - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0" + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15" ], "out_args": [ "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15" @@ -8000,7 +8000,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8008,7 +8008,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8063,7 +8063,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -8071,7 +8071,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8093,7 +8093,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8133,7 +8133,7 @@ "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_68" ], "const_args": [ - "onnx::MatMul_8004" + "onnx::MatMul_8004_45_1_14" ], "out_args": [ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_18" @@ -8142,7 +8142,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8150,7 +8150,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -8203,16 +8203,16 @@ "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_18" ], "const_args": [ - "onnx::MatMul_8005" + "onnx::MatMul_8005_45_1_16" ], "out_args": [ - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_36_bfp.out25_20" + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -8220,7 +8220,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8265,7 +8265,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_36_bfp.out25_20", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20", "/transformer_blocks.3/Add_4_output_0.out10_66" ], "const_args": [ @@ -8278,7 +8278,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -8286,7 +8286,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8294,7 +8294,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8309,7 +8309,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8350,8 +8350,8 @@ "/Add_3_output_0.out_35_1_7" ], "const_args": [ - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_17" ], "out_args": [ "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_17_bfp.out15_17" @@ -8378,7 +8378,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -8386,7 +8386,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -8431,8 +8431,8 @@ "/transformer_blocks.3/Add_7_output_0.out10_67" ], "const_args": [ - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_16" ], "out_args": [ "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_16_bfp.out15_16" @@ -8459,7 +8459,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8467,7 +8467,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8522,7 +8522,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -8530,7 +8530,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -8552,7 +8552,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -8602,7 +8602,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -8610,7 +8610,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8632,7 +8632,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8697,16 +8697,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8714,7 +8708,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -8736,7 +8730,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8744,11 +8738,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -8803,16 +8803,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8820,7 +8814,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -8842,7 +8836,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8850,11 +8844,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -8903,16 +8903,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8920,7 +8914,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -8942,7 +8936,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -8950,11 +8944,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -9011,7 +9011,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -9057,7 +9057,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_8026" + "onnx::MatMul_8026_45_5_9" ], "out_args": [ "/transformer_blocks.4/attn/to_add_out/Add_output_0.out6_1_9_bfp.out7_9" @@ -9072,7 +9072,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -9098,7 +9098,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9106,7 +9106,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9164,7 +9164,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -9172,7 +9172,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9180,7 +9180,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9195,7 +9195,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9236,7 +9236,7 @@ "/transformer_blocks.4/attn/Reshape_3_output_0.out22_4_bfp.out27_0_4" ], "const_args": [ - "onnx::MatMul_8025" + "onnx::MatMul_8025_45_5_8" ], "out_args": [ "/transformer_blocks.4/attn/to_out.0/Add_output_0.out6_1_8_bfp.out7_8" @@ -9251,7 +9251,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -9277,7 +9277,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9285,7 +9285,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9343,7 +9343,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -9351,7 +9351,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9359,7 +9359,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9374,7 +9374,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9415,8 +9415,8 @@ "/transformer_blocks.4/Add_output_0.out10_68" ], "const_args": [ - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0" + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18" ], "out_args": [ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18" @@ -9443,7 +9443,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9451,7 +9451,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9506,7 +9506,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -9514,7 +9514,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9536,7 +9536,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9576,7 +9576,7 @@ "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_70" ], "const_args": [ - "onnx::MatMul_8027" + "onnx::MatMul_8027_45_1_17" ], "out_args": [ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_21" @@ -9585,7 +9585,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9593,7 +9593,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -9646,16 +9646,16 @@ "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_21" ], "const_args": [ - "onnx::MatMul_8028" + "onnx::MatMul_8028_45_1_19" ], "out_args": [ - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_43_bfp.out25_23" + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -9663,7 +9663,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -9710,8 +9710,8 @@ "/transformer_blocks.4/Add_4_output_0.out10_70" ], "const_args": [ - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0" + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19" ], "out_args": [ "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19" @@ -9738,7 +9738,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9746,7 +9746,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9801,7 +9801,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -9809,7 +9809,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9831,7 +9831,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9871,7 +9871,7 @@ "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_72" ], "const_args": [ - "onnx::MatMul_8029" + "onnx::MatMul_8029_45_1_18" ], "out_args": [ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_22" @@ -9880,7 +9880,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -9888,7 +9888,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -9941,16 +9941,16 @@ "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_22" ], "const_args": [ - "onnx::MatMul_8030" + "onnx::MatMul_8030_45_1_20" ], "out_args": [ - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_44_bfp.out25_24" + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -9958,7 +9958,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10003,7 +10003,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_44_bfp.out25_24", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24", "/transformer_blocks.4/Add_4_output_0.out10_70" ], "const_args": [ @@ -10016,7 +10016,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -10024,7 +10024,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10032,7 +10032,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10047,7 +10047,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10088,8 +10088,8 @@ "/transformer_blocks.4/Add_7_output_0.out10_71" ], "const_args": [ - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_20" ], "out_args": [ "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_20_bfp.out15_20" @@ -10116,7 +10116,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10124,7 +10124,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10167,7 +10167,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_43_bfp.out25_23", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23", "/transformer_blocks.4/Add_output_0.out10_68" ], "const_args": [ @@ -10180,7 +10180,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -10188,7 +10188,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10196,7 +10196,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10211,7 +10211,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10260,7 +10260,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10268,7 +10268,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10276,7 +10276,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10315,8 +10315,8 @@ "/Add_4_output_0.out_35_1_8" ], "const_args": [ - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0" + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_21" ], "out_args": [ "/transformer_blocks.5/norm1/norm/LayerNormalization_output_0.out14_21_bfp.out15_21" @@ -10343,7 +10343,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10351,7 +10351,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10406,7 +10406,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -10414,7 +10414,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10436,7 +10436,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10486,7 +10486,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -10494,7 +10494,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10516,7 +10516,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10581,16 +10581,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10598,7 +10592,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10620,7 +10614,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10628,11 +10622,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -10687,16 +10687,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10704,7 +10698,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10726,7 +10720,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10734,11 +10728,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -10787,16 +10787,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10804,7 +10798,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -10826,7 +10820,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10834,11 +10828,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -10895,7 +10895,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -10941,7 +10941,7 @@ "/transformer_blocks.5/attn/Reshape_3_output_0.out22_5_bfp.out27_0_5" ], "const_args": [ - "onnx::MatMul_8051" + "onnx::MatMul_8051_45_5_11" ], "out_args": [ "/transformer_blocks.5/attn/to_add_out/Add_output_0.out6_1_11_bfp.out7_11" @@ -10956,7 +10956,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -10982,7 +10982,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -10990,7 +10990,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11048,7 +11048,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -11056,7 +11056,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11064,7 +11064,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11079,7 +11079,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11120,7 +11120,7 @@ "/transformer_blocks.5/attn/Reshape_3_output_0.out22_5_bfp.out27_0_5" ], "const_args": [ - "onnx::MatMul_8050" + "onnx::MatMul_8050_45_5_10" ], "out_args": [ "/transformer_blocks.5/attn/to_out.0/Add_output_0.out6_1_10_bfp.out7_10" @@ -11135,7 +11135,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -11161,7 +11161,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11169,7 +11169,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11227,7 +11227,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -11235,7 +11235,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11243,7 +11243,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11258,7 +11258,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11299,8 +11299,8 @@ "/transformer_blocks.5/Add_output_0.out10_72" ], "const_args": [ - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0" + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22" ], "out_args": [ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22" @@ -11327,7 +11327,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11335,7 +11335,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11390,7 +11390,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -11398,7 +11398,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11420,7 +11420,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11460,7 +11460,7 @@ "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_74" ], "const_args": [ - "onnx::MatMul_8052" + "onnx::MatMul_8052_45_1_21" ], "out_args": [ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_25" @@ -11469,7 +11469,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11477,7 +11477,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -11530,16 +11530,16 @@ "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_25" ], "const_args": [ - "onnx::MatMul_8053" + "onnx::MatMul_8053_45_1_23" ], "out_args": [ - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_51_bfp.out25_27" + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -11547,7 +11547,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11592,7 +11592,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_51_bfp.out25_27", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27", "/transformer_blocks.5/Add_output_0.out10_72" ], "const_args": [ @@ -11605,7 +11605,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -11613,7 +11613,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11621,7 +11621,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11636,7 +11636,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11685,7 +11685,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11693,7 +11693,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11701,7 +11701,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -11740,8 +11740,8 @@ "/transformer_blocks.5/Add_4_output_0.out10_74" ], "const_args": [ - "/transformer_blocks.5/norm2_context/Constant_output_0", - "/transformer_blocks.5/norm2_context/Constant_1_output_0" + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23" ], "out_args": [ "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23" @@ -11768,7 +11768,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11776,7 +11776,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11831,7 +11831,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -11839,7 +11839,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11861,7 +11861,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11901,7 +11901,7 @@ "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_76" ], "const_args": [ - "onnx::MatMul_8054" + "onnx::MatMul_8054_45_1_22" ], "out_args": [ "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_26" @@ -11910,7 +11910,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -11918,7 +11918,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -11971,16 +11971,16 @@ "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_26" ], "const_args": [ - "onnx::MatMul_8055" + "onnx::MatMul_8055_45_1_24" ], "out_args": [ - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_52_bfp.out25_28" + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -11988,7 +11988,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12033,7 +12033,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_52_bfp.out25_28", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28", "/transformer_blocks.5/Add_4_output_0.out10_74" ], "const_args": [ @@ -12046,7 +12046,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -12054,7 +12054,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12062,7 +12062,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12077,7 +12077,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12118,8 +12118,8 @@ "/Add_5_output_0.out_35_1_9" ], "const_args": [ - "/transformer_blocks.6/norm1/norm/Constant_output_0", - "/transformer_blocks.6/norm1/norm/Constant_1_output_0" + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_25" ], "out_args": [ "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_25_bfp.out15_25" @@ -12146,7 +12146,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -12154,7 +12154,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -12199,8 +12199,8 @@ "/transformer_blocks.5/Add_7_output_0.out10_75" ], "const_args": [ - "/transformer_blocks.6/norm1_context/norm/Constant_output_0", - "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_24" ], "out_args": [ "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_24_bfp.out15_24" @@ -12227,7 +12227,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12235,7 +12235,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12290,7 +12290,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -12298,7 +12298,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -12320,7 +12320,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -12370,7 +12370,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -12378,7 +12378,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12400,7 +12400,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12465,16 +12465,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12482,7 +12476,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -12504,7 +12498,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12512,11 +12506,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -12571,16 +12571,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12588,7 +12582,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -12610,7 +12604,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12618,11 +12612,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -12671,16 +12671,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12688,7 +12682,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -12710,7 +12704,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12718,11 +12712,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -12779,7 +12779,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -12825,7 +12825,7 @@ "/transformer_blocks.6/attn/Reshape_3_output_0.out22_6_bfp.out27_0_6" ], "const_args": [ - "onnx::MatMul_8076" + "onnx::MatMul_8076_45_5_13" ], "out_args": [ "/transformer_blocks.6/attn/to_add_out/Add_output_0.out6_1_13_bfp.out7_13" @@ -12840,7 +12840,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -12866,7 +12866,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12874,7 +12874,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12932,7 +12932,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -12940,7 +12940,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12948,7 +12948,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -12963,7 +12963,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13004,7 +13004,7 @@ "/transformer_blocks.6/attn/Reshape_3_output_0.out22_6_bfp.out27_0_6" ], "const_args": [ - "onnx::MatMul_8075" + "onnx::MatMul_8075_45_5_12" ], "out_args": [ "/transformer_blocks.6/attn/to_out.0/Add_output_0.out6_1_12_bfp.out7_12" @@ -13019,7 +13019,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -13045,7 +13045,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13053,7 +13053,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13111,7 +13111,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -13119,7 +13119,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13127,7 +13127,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13142,7 +13142,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13183,8 +13183,8 @@ "/transformer_blocks.6/Add_output_0.out10_76" ], "const_args": [ - "/transformer_blocks.6/norm2/Constant_output_0", - "/transformer_blocks.6/norm2/Constant_1_output_0" + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26" ], "out_args": [ "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26" @@ -13211,7 +13211,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13219,7 +13219,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13274,7 +13274,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -13282,7 +13282,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13304,7 +13304,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13344,7 +13344,7 @@ "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_78" ], "const_args": [ - "onnx::MatMul_8077" + "onnx::MatMul_8077_45_1_25" ], "out_args": [ "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_29" @@ -13353,7 +13353,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13361,7 +13361,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -13414,16 +13414,16 @@ "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_29" ], "const_args": [ - "onnx::MatMul_8078" + "onnx::MatMul_8078_45_1_27" ], "out_args": [ - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_59_bfp.out25_31" + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -13431,7 +13431,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13478,8 +13478,8 @@ "/transformer_blocks.6/Add_4_output_0.out10_78" ], "const_args": [ - "/transformer_blocks.6/norm2_context/Constant_output_0", - "/transformer_blocks.6/norm2_context/Constant_1_output_0" + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27" ], "out_args": [ "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27" @@ -13506,7 +13506,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13514,7 +13514,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13569,7 +13569,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -13577,7 +13577,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13599,7 +13599,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13639,7 +13639,7 @@ "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_80" ], "const_args": [ - "onnx::MatMul_8079" + "onnx::MatMul_8079_45_1_26" ], "out_args": [ "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_30" @@ -13648,7 +13648,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13656,7 +13656,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -13709,16 +13709,16 @@ "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_30" ], "const_args": [ - "onnx::MatMul_8080" + "onnx::MatMul_8080_45_1_28" ], "out_args": [ - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_60_bfp.out25_32" + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -13726,7 +13726,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13771,7 +13771,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_60_bfp.out25_32", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32", "/transformer_blocks.6/Add_4_output_0.out10_78" ], "const_args": [ @@ -13784,7 +13784,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -13792,7 +13792,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13800,7 +13800,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13815,7 +13815,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13856,8 +13856,8 @@ "/transformer_blocks.6/Add_7_output_0.out10_79" ], "const_args": [ - "/transformer_blocks.7/norm1_context/norm/Constant_output_0", - "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_28" ], "out_args": [ "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_28_bfp.out15_28" @@ -13884,7 +13884,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13892,7 +13892,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -13935,7 +13935,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_59_bfp.out25_31", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31", "/transformer_blocks.6/Add_output_0.out10_76" ], "const_args": [ @@ -13948,7 +13948,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -13956,7 +13956,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13964,7 +13964,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -13979,7 +13979,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14028,7 +14028,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14036,7 +14036,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14044,7 +14044,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14083,8 +14083,8 @@ "/Add_6_output_0.out_35_1_10" ], "const_args": [ - "/transformer_blocks.7/norm1/norm/Constant_output_0", - "/transformer_blocks.7/norm1/norm/Constant_1_output_0" + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_29" ], "out_args": [ "/transformer_blocks.7/norm1/norm/LayerNormalization_output_0.out14_29_bfp.out15_29" @@ -14111,7 +14111,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14119,7 +14119,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14174,7 +14174,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -14182,7 +14182,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14204,7 +14204,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14254,7 +14254,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -14262,7 +14262,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14284,7 +14284,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14349,16 +14349,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14366,7 +14360,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14388,7 +14382,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14396,11 +14390,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -14455,16 +14455,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14472,7 +14466,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14494,7 +14488,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14502,11 +14496,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -14555,16 +14555,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14572,7 +14566,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14594,7 +14588,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14602,11 +14596,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -14663,7 +14663,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -14709,7 +14709,7 @@ "/transformer_blocks.7/attn/Reshape_3_output_0.out22_7_bfp.out27_0_7" ], "const_args": [ - "onnx::MatMul_8101" + "onnx::MatMul_8101_45_5_15" ], "out_args": [ "/transformer_blocks.7/attn/to_add_out/Add_output_0.out6_1_15_bfp.out7_15" @@ -14724,7 +14724,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -14750,7 +14750,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14758,7 +14758,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14816,7 +14816,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -14824,7 +14824,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14832,7 +14832,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14847,7 +14847,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -14888,7 +14888,7 @@ "/transformer_blocks.7/attn/Reshape_3_output_0.out22_7_bfp.out27_0_7" ], "const_args": [ - "onnx::MatMul_8100" + "onnx::MatMul_8100_45_5_14" ], "out_args": [ "/transformer_blocks.7/attn/to_out.0/Add_output_0.out6_1_14_bfp.out7_14" @@ -14903,7 +14903,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -14929,7 +14929,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14937,7 +14937,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -14995,7 +14995,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -15003,7 +15003,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15011,7 +15011,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15026,7 +15026,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15067,8 +15067,8 @@ "/transformer_blocks.7/Add_output_0.out10_80" ], "const_args": [ - "/transformer_blocks.7/norm2/Constant_output_0", - "/transformer_blocks.7/norm2/Constant_1_output_0" + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30" ], "out_args": [ "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30" @@ -15095,7 +15095,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15103,7 +15103,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15158,7 +15158,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -15166,7 +15166,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15188,7 +15188,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15228,7 +15228,7 @@ "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_82" ], "const_args": [ - "onnx::MatMul_8102" + "onnx::MatMul_8102_45_1_29" ], "out_args": [ "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_33" @@ -15237,7 +15237,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15245,7 +15245,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -15298,16 +15298,16 @@ "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_33" ], "const_args": [ - "onnx::MatMul_8103" + "onnx::MatMul_8103_45_1_31" ], "out_args": [ - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_67_bfp.out25_35" + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -15315,7 +15315,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15360,7 +15360,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_67_bfp.out25_35", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35", "/transformer_blocks.7/Add_output_0.out10_80" ], "const_args": [ @@ -15373,7 +15373,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -15381,7 +15381,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15389,7 +15389,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15404,7 +15404,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15453,7 +15453,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15461,7 +15461,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15469,7 +15469,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15508,8 +15508,8 @@ "/transformer_blocks.7/Add_4_output_0.out10_82" ], "const_args": [ - "/transformer_blocks.7/norm2_context/Constant_output_0", - "/transformer_blocks.7/norm2_context/Constant_1_output_0" + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31" ], "out_args": [ "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31" @@ -15536,7 +15536,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15544,7 +15544,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15599,7 +15599,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -15607,7 +15607,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15629,7 +15629,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15669,7 +15669,7 @@ "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_84" ], "const_args": [ - "onnx::MatMul_8104" + "onnx::MatMul_8104_45_1_30" ], "out_args": [ "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_34" @@ -15678,7 +15678,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15686,7 +15686,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -15739,16 +15739,16 @@ "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_34" ], "const_args": [ - "onnx::MatMul_8105" + "onnx::MatMul_8105_45_1_32" ], "out_args": [ - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_68_bfp.out25_36" + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -15756,7 +15756,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15801,7 +15801,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_68_bfp.out25_36", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36", "/transformer_blocks.7/Add_4_output_0.out10_82" ], "const_args": [ @@ -15814,7 +15814,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -15822,7 +15822,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15830,7 +15830,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15845,7 +15845,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -15886,8 +15886,8 @@ "/Add_7_output_0.out_35_1_11" ], "const_args": [ - "/transformer_blocks.8/norm1/norm/Constant_output_0", - "/transformer_blocks.8/norm1/norm/Constant_1_output_0" + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_33" ], "out_args": [ "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_33_bfp.out15_33" @@ -15914,7 +15914,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15922,7 +15922,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -15967,8 +15967,8 @@ "/transformer_blocks.7/Add_7_output_0.out10_83" ], "const_args": [ - "/transformer_blocks.8/norm1_context/norm/Constant_output_0", - "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_32" ], "out_args": [ "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_32_bfp.out15_32" @@ -15995,7 +15995,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16003,7 +16003,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16058,7 +16058,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -16066,7 +16066,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16088,7 +16088,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16138,7 +16138,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -16146,7 +16146,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16168,7 +16168,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16233,16 +16233,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16250,7 +16244,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16272,7 +16266,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16280,11 +16274,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -16339,16 +16339,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16356,7 +16350,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16378,7 +16372,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16386,11 +16380,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -16439,16 +16439,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16456,7 +16450,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16478,7 +16472,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16486,11 +16480,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -16547,7 +16547,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -16593,7 +16593,7 @@ "/transformer_blocks.8/attn/Reshape_3_output_0.out22_8_bfp.out27_0_8" ], "const_args": [ - "onnx::MatMul_8126" + "onnx::MatMul_8126_45_5_17" ], "out_args": [ "/transformer_blocks.8/attn/to_add_out/Add_output_0.out6_1_17_bfp.out7_17" @@ -16608,7 +16608,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -16634,7 +16634,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16642,7 +16642,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16700,7 +16700,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -16708,7 +16708,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16716,7 +16716,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16731,7 +16731,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -16772,7 +16772,7 @@ "/transformer_blocks.8/attn/Reshape_3_output_0.out22_8_bfp.out27_0_8" ], "const_args": [ - "onnx::MatMul_8125" + "onnx::MatMul_8125_45_5_16" ], "out_args": [ "/transformer_blocks.8/attn/to_out.0/Add_output_0.out6_1_16_bfp.out7_16" @@ -16787,7 +16787,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -16813,7 +16813,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16821,7 +16821,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16879,7 +16879,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -16887,7 +16887,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16895,7 +16895,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16910,7 +16910,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16951,8 +16951,8 @@ "/transformer_blocks.8/Add_output_0.out10_84" ], "const_args": [ - "/transformer_blocks.8/norm2/Constant_output_0", - "/transformer_blocks.8/norm2/Constant_1_output_0" + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34" ], "out_args": [ "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34" @@ -16979,7 +16979,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -16987,7 +16987,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17042,7 +17042,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -17050,7 +17050,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17072,7 +17072,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17112,7 +17112,7 @@ "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_86" ], "const_args": [ - "onnx::MatMul_8127" + "onnx::MatMul_8127_45_1_33" ], "out_args": [ "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_37" @@ -17121,7 +17121,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17129,7 +17129,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -17182,16 +17182,16 @@ "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_37" ], "const_args": [ - "onnx::MatMul_8128" + "onnx::MatMul_8128_45_1_35" ], "out_args": [ - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_75_bfp.out25_39" + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -17199,7 +17199,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17246,8 +17246,8 @@ "/transformer_blocks.8/Add_4_output_0.out10_86" ], "const_args": [ - "/transformer_blocks.8/norm2_context/Constant_output_0", - "/transformer_blocks.8/norm2_context/Constant_1_output_0" + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35" ], "out_args": [ "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35" @@ -17274,7 +17274,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17282,7 +17282,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17337,7 +17337,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -17345,7 +17345,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17367,7 +17367,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17407,7 +17407,7 @@ "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_88" ], "const_args": [ - "onnx::MatMul_8129" + "onnx::MatMul_8129_45_1_34" ], "out_args": [ "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_38" @@ -17416,7 +17416,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17424,7 +17424,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -17477,16 +17477,16 @@ "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_38" ], "const_args": [ - "onnx::MatMul_8130" + "onnx::MatMul_8130_45_1_36" ], "out_args": [ - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_76_bfp.out25_40" + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -17494,7 +17494,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17539,7 +17539,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_76_bfp.out25_40", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40", "/transformer_blocks.8/Add_4_output_0.out10_86" ], "const_args": [ @@ -17552,7 +17552,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -17560,7 +17560,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17568,7 +17568,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17583,7 +17583,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17624,8 +17624,8 @@ "/transformer_blocks.8/Add_7_output_0.out10_87" ], "const_args": [ - "/transformer_blocks.9/norm1_context/norm/Constant_output_0", - "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_36" ], "out_args": [ "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_36_bfp.out15_36" @@ -17652,7 +17652,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17660,7 +17660,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -17703,7 +17703,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_75_bfp.out25_39", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39", "/transformer_blocks.8/Add_output_0.out10_84" ], "const_args": [ @@ -17716,7 +17716,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -17724,7 +17724,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17732,7 +17732,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17747,7 +17747,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17796,7 +17796,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17804,7 +17804,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17812,7 +17812,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17851,8 +17851,8 @@ "/Add_8_output_0.out_35_1_12" ], "const_args": [ - "/transformer_blocks.9/norm1/norm/Constant_output_0", - "/transformer_blocks.9/norm1/norm/Constant_1_output_0" + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_37" ], "out_args": [ "/transformer_blocks.9/norm1/norm/LayerNormalization_output_0.out14_37_bfp.out15_37" @@ -17879,7 +17879,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17887,7 +17887,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17942,7 +17942,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -17950,7 +17950,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -17972,7 +17972,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18022,7 +18022,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -18030,7 +18030,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18052,7 +18052,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18117,16 +18117,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18134,7 +18128,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18156,7 +18150,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18164,11 +18158,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -18223,16 +18223,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18240,7 +18234,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18262,7 +18256,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18270,11 +18264,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -18323,16 +18323,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18340,7 +18334,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18362,7 +18356,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18370,11 +18364,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -18431,7 +18431,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -18477,7 +18477,7 @@ "/transformer_blocks.9/attn/Reshape_3_output_0.out22_9_bfp.out27_0_9" ], "const_args": [ - "onnx::MatMul_8151" + "onnx::MatMul_8151_45_5_19" ], "out_args": [ "/transformer_blocks.9/attn/to_add_out/Add_output_0.out6_1_19_bfp.out7_19" @@ -18492,7 +18492,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -18518,7 +18518,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18526,7 +18526,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18584,7 +18584,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -18592,7 +18592,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18600,7 +18600,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18615,7 +18615,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -18656,7 +18656,7 @@ "/transformer_blocks.9/attn/Reshape_3_output_0.out22_9_bfp.out27_0_9" ], "const_args": [ - "onnx::MatMul_8150" + "onnx::MatMul_8150_45_5_18" ], "out_args": [ "/transformer_blocks.9/attn/to_out.0/Add_output_0.out6_1_18_bfp.out7_18" @@ -18671,7 +18671,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -18697,7 +18697,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18705,7 +18705,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18763,7 +18763,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -18771,7 +18771,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18779,7 +18779,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18794,7 +18794,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18835,8 +18835,8 @@ "/transformer_blocks.9/Add_output_0.out10_88" ], "const_args": [ - "/transformer_blocks.9/norm2/Constant_output_0", - "/transformer_blocks.9/norm2/Constant_1_output_0" + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38" ], "out_args": [ "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38" @@ -18863,7 +18863,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18871,7 +18871,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18926,7 +18926,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -18934,7 +18934,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18956,7 +18956,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -18996,7 +18996,7 @@ "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_90" ], "const_args": [ - "onnx::MatMul_8152" + "onnx::MatMul_8152_45_1_37" ], "out_args": [ "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_41" @@ -19005,7 +19005,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19013,7 +19013,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -19066,16 +19066,16 @@ "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_41" ], "const_args": [ - "onnx::MatMul_8153" + "onnx::MatMul_8153_45_1_39" ], "out_args": [ - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_83_bfp.out25_43" + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -19083,7 +19083,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19128,7 +19128,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_83_bfp.out25_43", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43", "/transformer_blocks.9/Add_output_0.out10_88" ], "const_args": [ @@ -19141,7 +19141,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -19149,7 +19149,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19157,7 +19157,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19172,7 +19172,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19221,7 +19221,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19229,7 +19229,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19237,7 +19237,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19276,8 +19276,8 @@ "/Add_9_output_0.out_35_1_13" ], "const_args": [ - "/transformer_blocks.10/norm1/norm/Constant_output_0", - "/transformer_blocks.10/norm1/norm/Constant_1_output_0" + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40" ], "out_args": [ "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40" @@ -19304,7 +19304,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19312,7 +19312,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19367,7 +19367,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -19375,7 +19375,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19397,7 +19397,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -19437,8 +19437,8 @@ "/transformer_blocks.9/Add_4_output_0.out10_90" ], "const_args": [ - "/transformer_blocks.9/norm2_context/Constant_output_0", - "/transformer_blocks.9/norm2_context/Constant_1_output_0" + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39" ], "out_args": [ "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39" @@ -19465,7 +19465,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19473,7 +19473,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19528,7 +19528,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -19536,7 +19536,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19558,7 +19558,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19598,7 +19598,7 @@ "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_92" ], "const_args": [ - "onnx::MatMul_8154" + "onnx::MatMul_8154_45_1_38" ], "out_args": [ "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_42" @@ -19607,7 +19607,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19615,7 +19615,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -19668,16 +19668,16 @@ "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_42" ], "const_args": [ - "onnx::MatMul_8155" + "onnx::MatMul_8155_45_1_40" ], "out_args": [ - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_44" + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -19685,7 +19685,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19730,7 +19730,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_44", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44", "/transformer_blocks.9/Add_4_output_0.out10_90" ], "const_args": [ @@ -19743,7 +19743,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -19751,7 +19751,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19759,7 +19759,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19774,7 +19774,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19815,8 +19815,8 @@ "/transformer_blocks.9/Add_7_output_0.out10_91" ], "const_args": [ - "/transformer_blocks.10/norm1_context/norm/Constant_output_0", - "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41" ], "out_args": [ "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41" @@ -19843,7 +19843,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19851,7 +19851,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19906,7 +19906,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -19914,7 +19914,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -19936,7 +19936,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20001,16 +20001,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20018,7 +20012,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20040,7 +20034,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20048,11 +20042,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -20107,16 +20107,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20124,7 +20118,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20146,7 +20140,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20154,11 +20148,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -20207,16 +20207,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20224,7 +20218,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20246,7 +20240,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20254,11 +20248,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -20315,7 +20315,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -20361,7 +20361,7 @@ "/transformer_blocks.10/attn/Reshape_3_output_0.out22_10_bfp.out27_0_10" ], "const_args": [ - "onnx::MatMul_8175" + "onnx::MatMul_8175_45_5_20" ], "out_args": [ "/transformer_blocks.10/attn/to_out.0/Add_output_0.out6_1_20_bfp.out7_20" @@ -20376,7 +20376,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -20402,7 +20402,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20410,7 +20410,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20468,7 +20468,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -20476,7 +20476,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20484,7 +20484,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20499,7 +20499,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20540,8 +20540,8 @@ "/transformer_blocks.10/Add_output_0.out10_6" ], "const_args": [ - "/transformer_blocks.10/norm2/Constant_output_0", - "/transformer_blocks.10/norm2/Constant_1_output_0" + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42" ], "out_args": [ "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42" @@ -20568,7 +20568,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20576,7 +20576,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20631,7 +20631,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -20639,7 +20639,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20661,7 +20661,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20701,7 +20701,7 @@ "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7" ], "const_args": [ - "onnx::MatMul_8177" + "onnx::MatMul_8177_45_1_41" ], "out_args": [ "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_45" @@ -20710,7 +20710,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20718,7 +20718,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -20771,16 +20771,16 @@ "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_45" ], "const_args": [ - "onnx::MatMul_8178" + "onnx::MatMul_8178_45_1_43" ], "out_args": [ - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_91_bfp.out25_47" + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -20788,7 +20788,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20833,7 +20833,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_91_bfp.out25_47", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47", "/transformer_blocks.10/Add_output_0.out10_6" ], "const_args": [ @@ -20846,7 +20846,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -20854,7 +20854,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20862,7 +20862,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20877,7 +20877,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -20918,7 +20918,7 @@ "/transformer_blocks.10/attn/Reshape_3_output_0.out22_10_bfp.out27_0_10" ], "const_args": [ - "onnx::MatMul_8176" + "onnx::MatMul_8176_45_5_21" ], "out_args": [ "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21" @@ -20933,7 +20933,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -20959,7 +20959,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -20967,7 +20967,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21025,7 +21025,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -21033,7 +21033,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21041,7 +21041,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21056,7 +21056,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21097,8 +21097,8 @@ "/transformer_blocks.10/Add_4_output_0.out10_8" ], "const_args": [ - "/transformer_blocks.10/norm2_context/Constant_output_0", - "/transformer_blocks.10/norm2_context/Constant_1_output_0" + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43" ], "out_args": [ "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43" @@ -21125,7 +21125,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21133,7 +21133,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21188,7 +21188,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -21196,7 +21196,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21218,7 +21218,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21258,7 +21258,7 @@ "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9" ], "const_args": [ - "onnx::MatMul_8179" + "onnx::MatMul_8179_45_1_42" ], "out_args": [ "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_46" @@ -21267,7 +21267,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21275,7 +21275,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -21328,16 +21328,16 @@ "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_46" ], "const_args": [ - "onnx::MatMul_8180" + "onnx::MatMul_8180_45_1_44" ], "out_args": [ - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_92_bfp.out25_48" + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -21345,7 +21345,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21390,7 +21390,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_92_bfp.out25_48", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48", "/transformer_blocks.10/Add_4_output_0.out10_8" ], "const_args": [ @@ -21403,7 +21403,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -21411,7 +21411,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21419,7 +21419,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21434,7 +21434,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21475,8 +21475,8 @@ "/transformer_blocks.10/Add_7_output_0.out10_9" ], "const_args": [ - "/transformer_blocks.11/norm1_context/norm/Constant_output_0", - "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_44" ], "out_args": [ "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_44_bfp.out15_44" @@ -21503,7 +21503,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21511,7 +21511,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21564,7 +21564,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21572,7 +21572,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21580,7 +21580,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21619,8 +21619,8 @@ "/Add_10_output_0.out_35_1_14" ], "const_args": [ - "/transformer_blocks.11/norm1/norm/Constant_output_0", - "/transformer_blocks.11/norm1/norm/Constant_1_output_0" + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_45" ], "out_args": [ "/transformer_blocks.11/norm1/norm/LayerNormalization_output_0.out14_45_bfp.out15_45" @@ -21647,7 +21647,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21655,7 +21655,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21710,7 +21710,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -21718,7 +21718,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21740,7 +21740,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21790,7 +21790,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -21798,7 +21798,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21820,7 +21820,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21885,16 +21885,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21902,7 +21896,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -21924,7 +21918,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -21932,11 +21926,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -21991,16 +21991,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -22008,7 +22002,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22030,7 +22024,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -22038,11 +22032,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -22091,16 +22091,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -22108,7 +22102,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22130,7 +22124,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -22138,11 +22132,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -22199,7 +22199,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -22245,7 +22245,7 @@ "/transformer_blocks.11/attn/Reshape_3_output_0.out22_11_bfp.out27_0_11" ], "const_args": [ - "onnx::MatMul_8200" + "onnx::MatMul_8200_45_5_22" ], "out_args": [ "/transformer_blocks.11/attn/to_out.0/Add_output_0.out6_1_22_bfp.out7_22" @@ -22260,7 +22260,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -22286,7 +22286,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22294,7 +22294,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22352,7 +22352,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -22360,7 +22360,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22368,7 +22368,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22383,7 +22383,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22424,8 +22424,8 @@ "/transformer_blocks.11/Add_output_0.out10_10" ], "const_args": [ - "/transformer_blocks.11/norm2/Constant_output_0", - "/transformer_blocks.11/norm2/Constant_1_output_0" + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46" ], "out_args": [ "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46" @@ -22452,7 +22452,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22460,7 +22460,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22515,7 +22515,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -22523,7 +22523,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22545,7 +22545,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22585,7 +22585,7 @@ "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11" ], "const_args": [ - "onnx::MatMul_8202" + "onnx::MatMul_8202_45_1_45" ], "out_args": [ "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_49" @@ -22594,7 +22594,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22602,7 +22602,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -22655,16 +22655,16 @@ "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_49" ], "const_args": [ - "onnx::MatMul_8203" + "onnx::MatMul_8203_45_1_47" ], "out_args": [ - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_99_bfp.out25_51" + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -22672,7 +22672,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22717,7 +22717,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_99_bfp.out25_51", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51", "/transformer_blocks.11/Add_output_0.out10_10" ], "const_args": [ @@ -22730,7 +22730,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -22738,7 +22738,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22746,7 +22746,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22761,7 +22761,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22810,7 +22810,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22818,7 +22818,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22826,7 +22826,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -22865,7 +22865,7 @@ "/transformer_blocks.11/attn/Reshape_3_output_0.out22_11_bfp.out27_0_11" ], "const_args": [ - "onnx::MatMul_8201" + "onnx::MatMul_8201_45_5_23" ], "out_args": [ "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23" @@ -22880,7 +22880,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -22906,7 +22906,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -22914,7 +22914,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -22972,7 +22972,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -22980,7 +22980,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -22988,7 +22988,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23003,7 +23003,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23044,8 +23044,8 @@ "/transformer_blocks.11/Add_4_output_0.out10_12" ], "const_args": [ - "/transformer_blocks.11/norm2_context/Constant_output_0", - "/transformer_blocks.11/norm2_context/Constant_1_output_0" + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47", + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47" ], "out_args": [ "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47" @@ -23072,7 +23072,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23080,7 +23080,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23135,7 +23135,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -23143,7 +23143,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23165,7 +23165,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23205,7 +23205,7 @@ "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13" ], "const_args": [ - "onnx::MatMul_8204" + "onnx::MatMul_8204_45_1_46" ], "out_args": [ "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_50" @@ -23214,7 +23214,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23222,7 +23222,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -23275,16 +23275,16 @@ "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_50" ], "const_args": [ - "onnx::MatMul_8205" + "onnx::MatMul_8205_45_1_48" ], "out_args": [ - "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_3_100_bfp.out25_52" + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -23292,7 +23292,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23337,7 +23337,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_3_100_bfp.out25_52", + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52", "/transformer_blocks.11/Add_4_output_0.out10_12" ], "const_args": [ @@ -23350,7 +23350,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -23358,7 +23358,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23366,7 +23366,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23381,7 +23381,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23422,8 +23422,8 @@ "/Add_11_output_0.out_35_1_15" ], "const_args": [ - "/transformer_blocks.12/norm1/norm/Constant_output_0", - "/transformer_blocks.12/norm1/norm/Constant_1_output_0" + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_49", + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_49" ], "out_args": [ "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_49_bfp.out15_49" @@ -23450,7 +23450,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -23458,7 +23458,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -23513,7 +23513,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -23521,7 +23521,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -23543,7 +23543,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -23583,8 +23583,8 @@ "/transformer_blocks.11/Add_7_output_0.out10_13" ], "const_args": [ - "/transformer_blocks.12/norm1_context/norm/Constant_output_0", - "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_48", + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_48" ], "out_args": [ "/transformer_blocks.12/norm1_context/norm/LayerNormalization_output_0.out14_48_bfp.out15_48" @@ -23611,7 +23611,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23619,7 +23619,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23674,7 +23674,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -23682,7 +23682,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23704,7 +23704,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23769,16 +23769,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23786,7 +23780,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -23808,7 +23802,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23816,11 +23810,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -23875,16 +23875,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23892,7 +23886,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -23914,7 +23908,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23922,11 +23916,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -23975,16 +23975,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -23992,7 +23986,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24014,7 +24008,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24022,11 +24016,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -24083,7 +24083,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -24129,7 +24129,7 @@ "/transformer_blocks.12/attn/Reshape_3_output_0.out22_12_bfp.out27_0_12" ], "const_args": [ - "onnx::MatMul_8225" + "onnx::MatMul_8225_45_5_24" ], "out_args": [ "/transformer_blocks.12/attn/to_out.0/Add_output_0.out6_1_24_bfp.out7_24" @@ -24144,7 +24144,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -24170,7 +24170,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24178,7 +24178,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24236,7 +24236,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -24244,7 +24244,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24252,7 +24252,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24267,7 +24267,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24308,8 +24308,8 @@ "/transformer_blocks.12/Add_output_0.out10_14" ], "const_args": [ - "/transformer_blocks.12/norm2/Constant_output_0", - "/transformer_blocks.12/norm2/Constant_1_output_0" + "/transformer_blocks.12/norm2/Constant_output_0_14_50", + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50" ], "out_args": [ "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50" @@ -24336,7 +24336,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24344,7 +24344,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24399,7 +24399,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -24407,7 +24407,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24429,7 +24429,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24469,7 +24469,7 @@ "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15" ], "const_args": [ - "onnx::MatMul_8227" + "onnx::MatMul_8227_45_1_49" ], "out_args": [ "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_53" @@ -24478,7 +24478,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24486,7 +24486,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -24539,16 +24539,16 @@ "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_53" ], "const_args": [ - "onnx::MatMul_8228" + "onnx::MatMul_8228_45_1_51" ], "out_args": [ - "/transformer_blocks.12/ff/net.2/Add_output_0.out17_3_107_bfp.out25_55" + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -24556,7 +24556,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24601,7 +24601,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.12/ff/net.2/Add_output_0.out17_3_107_bfp.out25_55", + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55", "/transformer_blocks.12/Add_output_0.out10_14" ], "const_args": [ @@ -24614,7 +24614,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -24622,7 +24622,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24630,7 +24630,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24645,7 +24645,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -24686,7 +24686,7 @@ "/transformer_blocks.12/attn/Reshape_3_output_0.out22_12_bfp.out27_0_12" ], "const_args": [ - "onnx::MatMul_8226" + "onnx::MatMul_8226_45_5_25" ], "out_args": [ "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25" @@ -24701,7 +24701,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -24727,7 +24727,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24735,7 +24735,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24793,7 +24793,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -24801,7 +24801,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24809,7 +24809,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24824,7 +24824,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24865,8 +24865,8 @@ "/transformer_blocks.12/Add_4_output_0.out10_16" ], "const_args": [ - "/transformer_blocks.12/norm2_context/Constant_output_0", - "/transformer_blocks.12/norm2_context/Constant_1_output_0" + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51", + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51" ], "out_args": [ "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51" @@ -24893,7 +24893,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24901,7 +24901,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24956,7 +24956,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -24964,7 +24964,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -24986,7 +24986,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25026,7 +25026,7 @@ "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17" ], "const_args": [ - "onnx::MatMul_8229" + "onnx::MatMul_8229_45_1_50" ], "out_args": [ "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_54" @@ -25035,7 +25035,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25043,7 +25043,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -25096,16 +25096,16 @@ "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_54" ], "const_args": [ - "onnx::MatMul_8230" + "onnx::MatMul_8230_45_1_52" ], "out_args": [ - "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_3_108_bfp.out25_56" + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -25113,7 +25113,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25158,7 +25158,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_3_108_bfp.out25_56", + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56", "/transformer_blocks.12/Add_4_output_0.out10_16" ], "const_args": [ @@ -25171,7 +25171,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -25179,7 +25179,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25187,7 +25187,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25202,7 +25202,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25243,8 +25243,8 @@ "/transformer_blocks.12/Add_7_output_0.out10_17" ], "const_args": [ - "/transformer_blocks.13/norm1_context/norm/Constant_output_0", - "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_52", + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_52" ], "out_args": [ "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_52_bfp.out15_52" @@ -25271,7 +25271,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25279,7 +25279,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25332,7 +25332,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25340,7 +25340,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25348,7 +25348,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25387,8 +25387,8 @@ "/Add_12_output_0.out_35_1_16" ], "const_args": [ - "/transformer_blocks.13/norm1/norm/Constant_output_0", - "/transformer_blocks.13/norm1/norm/Constant_1_output_0" + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_53", + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_53" ], "out_args": [ "/transformer_blocks.13/norm1/norm/LayerNormalization_output_0.out14_53_bfp.out15_53" @@ -25415,7 +25415,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25423,7 +25423,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25478,7 +25478,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -25486,7 +25486,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25508,7 +25508,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25558,7 +25558,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -25566,7 +25566,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25588,7 +25588,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25653,16 +25653,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25670,7 +25664,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25692,7 +25686,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25700,11 +25694,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -25759,16 +25759,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25776,7 +25770,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25798,7 +25792,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25806,11 +25800,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -25859,16 +25859,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25876,7 +25870,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -25898,7 +25892,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -25906,11 +25900,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -25967,7 +25967,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -26013,7 +26013,7 @@ "/transformer_blocks.13/attn/Reshape_3_output_0.out22_13_bfp.out27_0_13" ], "const_args": [ - "onnx::MatMul_8250" + "onnx::MatMul_8250_45_5_26" ], "out_args": [ "/transformer_blocks.13/attn/to_out.0/Add_output_0.out6_1_26_bfp.out7_26" @@ -26028,7 +26028,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -26054,7 +26054,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26062,7 +26062,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26120,7 +26120,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -26128,7 +26128,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26136,7 +26136,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26151,7 +26151,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26192,8 +26192,8 @@ "/transformer_blocks.13/Add_output_0.out10_18" ], "const_args": [ - "/transformer_blocks.13/norm2/Constant_output_0", - "/transformer_blocks.13/norm2/Constant_1_output_0" + "/transformer_blocks.13/norm2/Constant_output_0_14_54", + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54" ], "out_args": [ "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54" @@ -26220,7 +26220,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26228,7 +26228,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26283,7 +26283,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -26291,7 +26291,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26313,7 +26313,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26353,7 +26353,7 @@ "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19" ], "const_args": [ - "onnx::MatMul_8252" + "onnx::MatMul_8252_45_1_53" ], "out_args": [ "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_57" @@ -26362,7 +26362,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26370,7 +26370,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -26423,16 +26423,16 @@ "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_57" ], "const_args": [ - "onnx::MatMul_8253" + "onnx::MatMul_8253_45_1_55" ], "out_args": [ - "/transformer_blocks.13/ff/net.2/Add_output_0.out17_3_115_bfp.out25_59" + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -26440,7 +26440,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26485,7 +26485,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.13/ff/net.2/Add_output_0.out17_3_115_bfp.out25_59", + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59", "/transformer_blocks.13/Add_output_0.out10_18" ], "const_args": [ @@ -26498,7 +26498,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -26506,7 +26506,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26514,7 +26514,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26529,7 +26529,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26578,7 +26578,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26586,7 +26586,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26594,7 +26594,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -26633,7 +26633,7 @@ "/transformer_blocks.13/attn/Reshape_3_output_0.out22_13_bfp.out27_0_13" ], "const_args": [ - "onnx::MatMul_8251" + "onnx::MatMul_8251_45_5_27" ], "out_args": [ "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27" @@ -26648,7 +26648,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -26674,7 +26674,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26682,7 +26682,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26740,7 +26740,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -26748,7 +26748,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26756,7 +26756,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26771,7 +26771,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26812,8 +26812,8 @@ "/transformer_blocks.13/Add_4_output_0.out10_20" ], "const_args": [ - "/transformer_blocks.13/norm2_context/Constant_output_0", - "/transformer_blocks.13/norm2_context/Constant_1_output_0" + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55", + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55" ], "out_args": [ "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55" @@ -26840,7 +26840,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26848,7 +26848,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26903,7 +26903,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -26911,7 +26911,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26933,7 +26933,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26973,7 +26973,7 @@ "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21" ], "const_args": [ - "onnx::MatMul_8254" + "onnx::MatMul_8254_45_1_54" ], "out_args": [ "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_58" @@ -26982,7 +26982,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -26990,7 +26990,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -27043,16 +27043,16 @@ "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_58" ], "const_args": [ - "onnx::MatMul_8255" + "onnx::MatMul_8255_45_1_56" ], "out_args": [ - "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_3_116_bfp.out25_60" + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -27060,7 +27060,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27105,7 +27105,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_3_116_bfp.out25_60", + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60", "/transformer_blocks.13/Add_4_output_0.out10_20" ], "const_args": [ @@ -27118,7 +27118,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -27126,7 +27126,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27134,7 +27134,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27149,7 +27149,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27190,8 +27190,8 @@ "/Add_13_output_0.out_35_1_17" ], "const_args": [ - "/transformer_blocks.14/norm1/norm/Constant_output_0", - "/transformer_blocks.14/norm1/norm/Constant_1_output_0" + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_57", + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_57" ], "out_args": [ "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_57_bfp.out15_57" @@ -27218,7 +27218,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27226,7 +27226,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27281,7 +27281,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -27289,7 +27289,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27311,7 +27311,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27351,8 +27351,8 @@ "/transformer_blocks.13/Add_7_output_0.out10_21" ], "const_args": [ - "/transformer_blocks.14/norm1_context/norm/Constant_output_0", - "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_56", + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_56" ], "out_args": [ "/transformer_blocks.14/norm1_context/norm/LayerNormalization_output_0.out14_56_bfp.out15_56" @@ -27379,7 +27379,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27387,7 +27387,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27442,7 +27442,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -27450,7 +27450,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27472,7 +27472,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27537,16 +27537,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27554,7 +27548,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27576,7 +27570,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27584,11 +27578,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -27643,16 +27643,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27660,7 +27654,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27682,7 +27676,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27690,11 +27684,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -27743,16 +27743,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27760,7 +27754,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27782,7 +27776,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -27790,11 +27784,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -27851,7 +27851,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -27897,7 +27897,7 @@ "/transformer_blocks.14/attn/Reshape_3_output_0.out22_14_bfp.out27_0_14" ], "const_args": [ - "onnx::MatMul_8275" + "onnx::MatMul_8275_45_5_28" ], "out_args": [ "/transformer_blocks.14/attn/to_out.0/Add_output_0.out6_1_28_bfp.out7_28" @@ -27912,7 +27912,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -27938,7 +27938,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -27946,7 +27946,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28004,7 +28004,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -28012,7 +28012,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28020,7 +28020,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28035,7 +28035,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28076,8 +28076,8 @@ "/transformer_blocks.14/Add_output_0.out10_22" ], "const_args": [ - "/transformer_blocks.14/norm2/Constant_output_0", - "/transformer_blocks.14/norm2/Constant_1_output_0" + "/transformer_blocks.14/norm2/Constant_output_0_14_58", + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58" ], "out_args": [ "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58" @@ -28104,7 +28104,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28112,7 +28112,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28167,7 +28167,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -28175,7 +28175,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28197,7 +28197,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28237,7 +28237,7 @@ "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23" ], "const_args": [ - "onnx::MatMul_8277" + "onnx::MatMul_8277_45_1_57" ], "out_args": [ "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_61" @@ -28246,7 +28246,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28254,7 +28254,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -28307,16 +28307,16 @@ "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_61" ], "const_args": [ - "onnx::MatMul_8278" + "onnx::MatMul_8278_45_1_59" ], "out_args": [ - "/transformer_blocks.14/ff/net.2/Add_output_0.out17_3_123_bfp.out25_63" + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -28324,7 +28324,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28369,7 +28369,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.14/ff/net.2/Add_output_0.out17_3_123_bfp.out25_63", + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63", "/transformer_blocks.14/Add_output_0.out10_22" ], "const_args": [ @@ -28382,7 +28382,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -28390,7 +28390,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28398,7 +28398,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28413,7 +28413,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -28454,7 +28454,7 @@ "/transformer_blocks.14/attn/Reshape_3_output_0.out22_14_bfp.out27_0_14" ], "const_args": [ - "onnx::MatMul_8276" + "onnx::MatMul_8276_45_5_29" ], "out_args": [ "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29" @@ -28469,7 +28469,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -28495,7 +28495,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28503,7 +28503,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28561,7 +28561,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -28569,7 +28569,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28577,7 +28577,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28592,7 +28592,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28633,8 +28633,8 @@ "/transformer_blocks.14/Add_4_output_0.out10_24" ], "const_args": [ - "/transformer_blocks.14/norm2_context/Constant_output_0", - "/transformer_blocks.14/norm2_context/Constant_1_output_0" + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59", + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59" ], "out_args": [ "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59" @@ -28661,7 +28661,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28669,7 +28669,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28724,7 +28724,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -28732,7 +28732,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28754,7 +28754,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28794,7 +28794,7 @@ "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25" ], "const_args": [ - "onnx::MatMul_8279" + "onnx::MatMul_8279_45_1_58" ], "out_args": [ "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_62" @@ -28803,7 +28803,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28811,7 +28811,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -28864,16 +28864,16 @@ "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_62" ], "const_args": [ - "onnx::MatMul_8280" + "onnx::MatMul_8280_45_1_60" ], "out_args": [ - "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_3_124_bfp.out25_64" + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -28881,7 +28881,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28926,7 +28926,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_3_124_bfp.out25_64", + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64", "/transformer_blocks.14/Add_4_output_0.out10_24" ], "const_args": [ @@ -28939,7 +28939,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -28947,7 +28947,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28955,7 +28955,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -28970,7 +28970,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29011,8 +29011,8 @@ "/transformer_blocks.14/Add_7_output_0.out10_25" ], "const_args": [ - "/transformer_blocks.15/norm1_context/norm/Constant_output_0", - "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_60", + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_60" ], "out_args": [ "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_60_bfp.out15_60" @@ -29039,7 +29039,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29047,7 +29047,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29100,7 +29100,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29108,7 +29108,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29116,7 +29116,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29155,8 +29155,8 @@ "/Add_14_output_0.out_35_1_18" ], "const_args": [ - "/transformer_blocks.15/norm1/norm/Constant_output_0", - "/transformer_blocks.15/norm1/norm/Constant_1_output_0" + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_61", + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_61" ], "out_args": [ "/transformer_blocks.15/norm1/norm/LayerNormalization_output_0.out14_61_bfp.out15_61" @@ -29183,7 +29183,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29191,7 +29191,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29246,7 +29246,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -29254,7 +29254,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29276,7 +29276,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29326,7 +29326,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -29334,7 +29334,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29356,7 +29356,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29421,16 +29421,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29438,7 +29432,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29460,7 +29454,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29468,11 +29462,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -29527,16 +29527,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29544,7 +29538,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29566,7 +29560,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29574,11 +29568,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -29627,16 +29627,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29644,7 +29638,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29666,7 +29660,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -29674,11 +29668,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -29735,7 +29735,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -29781,7 +29781,7 @@ "/transformer_blocks.15/attn/Reshape_3_output_0.out22_15_bfp.out27_0_15" ], "const_args": [ - "onnx::MatMul_8300" + "onnx::MatMul_8300_45_5_30" ], "out_args": [ "/transformer_blocks.15/attn/to_out.0/Add_output_0.out6_1_30_bfp.out7_30" @@ -29796,7 +29796,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -29822,7 +29822,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29830,7 +29830,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29888,7 +29888,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -29896,7 +29896,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29904,7 +29904,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29919,7 +29919,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29960,8 +29960,8 @@ "/transformer_blocks.15/Add_output_0.out10_26" ], "const_args": [ - "/transformer_blocks.15/norm2/Constant_output_0", - "/transformer_blocks.15/norm2/Constant_1_output_0" + "/transformer_blocks.15/norm2/Constant_output_0_14_62", + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62" ], "out_args": [ "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62" @@ -29988,7 +29988,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -29996,7 +29996,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30051,7 +30051,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -30059,7 +30059,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30081,7 +30081,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30121,7 +30121,7 @@ "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27" ], "const_args": [ - "onnx::MatMul_8302" + "onnx::MatMul_8302_45_1_61" ], "out_args": [ "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_65" @@ -30130,7 +30130,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30138,7 +30138,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -30191,16 +30191,16 @@ "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_65" ], "const_args": [ - "onnx::MatMul_8303" + "onnx::MatMul_8303_45_1_63" ], "out_args": [ - "/transformer_blocks.15/ff/net.2/Add_output_0.out17_3_131_bfp.out25_67" + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -30208,7 +30208,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30253,7 +30253,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.15/ff/net.2/Add_output_0.out17_3_131_bfp.out25_67", + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67", "/transformer_blocks.15/Add_output_0.out10_26" ], "const_args": [ @@ -30266,7 +30266,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -30274,7 +30274,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30282,7 +30282,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30297,7 +30297,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30346,7 +30346,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30354,7 +30354,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30362,7 +30362,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30401,7 +30401,7 @@ "/transformer_blocks.15/attn/Reshape_3_output_0.out22_15_bfp.out27_0_15" ], "const_args": [ - "onnx::MatMul_8301" + "onnx::MatMul_8301_45_5_31" ], "out_args": [ "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31" @@ -30416,7 +30416,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -30442,7 +30442,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30450,7 +30450,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30508,7 +30508,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -30516,7 +30516,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30524,7 +30524,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30539,7 +30539,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30580,8 +30580,8 @@ "/transformer_blocks.15/Add_4_output_0.out10_28" ], "const_args": [ - "/transformer_blocks.15/norm2_context/Constant_output_0", - "/transformer_blocks.15/norm2_context/Constant_1_output_0" + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63", + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63" ], "out_args": [ "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63" @@ -30608,7 +30608,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30616,7 +30616,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30671,7 +30671,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -30679,7 +30679,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30701,7 +30701,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30741,7 +30741,7 @@ "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29" ], "const_args": [ - "onnx::MatMul_8304" + "onnx::MatMul_8304_45_1_62" ], "out_args": [ "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_66" @@ -30750,7 +30750,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30758,7 +30758,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -30811,16 +30811,16 @@ "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_66" ], "const_args": [ - "onnx::MatMul_8305" + "onnx::MatMul_8305_45_1_64" ], "out_args": [ - "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_3_132_bfp.out25_68" + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -30828,7 +30828,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30873,7 +30873,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_3_132_bfp.out25_68", + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68", "/transformer_blocks.15/Add_4_output_0.out10_28" ], "const_args": [ @@ -30886,7 +30886,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -30894,7 +30894,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30902,7 +30902,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30917,7 +30917,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -30958,8 +30958,8 @@ "/Add_15_output_0.out_35_1_19" ], "const_args": [ - "/transformer_blocks.16/norm1/norm/Constant_output_0", - "/transformer_blocks.16/norm1/norm/Constant_1_output_0" + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_65", + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_65" ], "out_args": [ "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_65_bfp.out15_65" @@ -30986,7 +30986,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -30994,7 +30994,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31049,7 +31049,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -31057,7 +31057,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31079,7 +31079,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31119,8 +31119,8 @@ "/transformer_blocks.15/Add_7_output_0.out10_29" ], "const_args": [ - "/transformer_blocks.16/norm1_context/norm/Constant_output_0", - "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_64", + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_64" ], "out_args": [ "/transformer_blocks.16/norm1_context/norm/LayerNormalization_output_0.out14_64_bfp.out15_64" @@ -31147,7 +31147,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31155,7 +31155,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31210,7 +31210,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -31218,7 +31218,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31240,7 +31240,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31305,16 +31305,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31322,7 +31316,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31344,7 +31338,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31352,11 +31346,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -31411,16 +31411,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31428,7 +31422,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31450,7 +31444,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31458,11 +31452,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -31511,16 +31511,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31528,7 +31522,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31550,7 +31544,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -31558,11 +31552,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -31619,7 +31619,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -31665,7 +31665,7 @@ "/transformer_blocks.16/attn/Reshape_3_output_0.out22_16_bfp.out27_0_16" ], "const_args": [ - "onnx::MatMul_8325" + "onnx::MatMul_8325_45_5_32" ], "out_args": [ "/transformer_blocks.16/attn/to_out.0/Add_output_0.out6_1_32_bfp.out7_32" @@ -31680,7 +31680,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -31706,7 +31706,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31714,7 +31714,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31772,7 +31772,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -31780,7 +31780,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31788,7 +31788,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31803,7 +31803,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31844,8 +31844,8 @@ "/transformer_blocks.16/Add_output_0.out10_30" ], "const_args": [ - "/transformer_blocks.16/norm2/Constant_output_0", - "/transformer_blocks.16/norm2/Constant_1_output_0" + "/transformer_blocks.16/norm2/Constant_output_0_14_66", + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66" ], "out_args": [ "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66" @@ -31872,7 +31872,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31880,7 +31880,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31935,7 +31935,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -31943,7 +31943,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -31965,7 +31965,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32005,7 +32005,7 @@ "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31" ], "const_args": [ - "onnx::MatMul_8327" + "onnx::MatMul_8327_45_1_65" ], "out_args": [ "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_69" @@ -32014,7 +32014,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32022,7 +32022,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -32075,16 +32075,16 @@ "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_69" ], "const_args": [ - "onnx::MatMul_8328" + "onnx::MatMul_8328_45_1_67" ], "out_args": [ - "/transformer_blocks.16/ff/net.2/Add_output_0.out17_3_139_bfp.out25_71" + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -32092,7 +32092,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32137,7 +32137,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.16/ff/net.2/Add_output_0.out17_3_139_bfp.out25_71", + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71", "/transformer_blocks.16/Add_output_0.out10_30" ], "const_args": [ @@ -32150,7 +32150,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -32158,7 +32158,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32166,7 +32166,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32181,7 +32181,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32222,7 +32222,7 @@ "/transformer_blocks.16/attn/Reshape_3_output_0.out22_16_bfp.out27_0_16" ], "const_args": [ - "onnx::MatMul_8326" + "onnx::MatMul_8326_45_5_33" ], "out_args": [ "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33" @@ -32237,7 +32237,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -32263,7 +32263,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32271,7 +32271,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32329,7 +32329,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -32337,7 +32337,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32345,7 +32345,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32360,7 +32360,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32401,8 +32401,8 @@ "/transformer_blocks.16/Add_4_output_0.out10_32" ], "const_args": [ - "/transformer_blocks.16/norm2_context/Constant_output_0", - "/transformer_blocks.16/norm2_context/Constant_1_output_0" + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67", + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67" ], "out_args": [ "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67" @@ -32429,7 +32429,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32437,7 +32437,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32492,7 +32492,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -32500,7 +32500,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32522,7 +32522,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32562,7 +32562,7 @@ "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33" ], "const_args": [ - "onnx::MatMul_8329" + "onnx::MatMul_8329_45_1_66" ], "out_args": [ "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_70" @@ -32571,7 +32571,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32579,7 +32579,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -32632,16 +32632,16 @@ "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_70" ], "const_args": [ - "onnx::MatMul_8330" + "onnx::MatMul_8330_45_1_68" ], "out_args": [ - "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_3_140_bfp.out25_72" + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -32649,7 +32649,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32694,7 +32694,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_3_140_bfp.out25_72", + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72", "/transformer_blocks.16/Add_4_output_0.out10_32" ], "const_args": [ @@ -32707,7 +32707,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -32715,7 +32715,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32723,7 +32723,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32738,7 +32738,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32779,8 +32779,8 @@ "/transformer_blocks.16/Add_7_output_0.out10_33" ], "const_args": [ - "/transformer_blocks.17/norm1_context/norm/Constant_output_0", - "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_68", + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_68" ], "out_args": [ "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_68_bfp.out15_68" @@ -32807,7 +32807,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32815,7 +32815,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -32868,7 +32868,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32876,7 +32876,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32884,7 +32884,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32923,8 +32923,8 @@ "/Add_16_output_0.out_35_1_20" ], "const_args": [ - "/transformer_blocks.17/norm1/norm/Constant_output_0", - "/transformer_blocks.17/norm1/norm/Constant_1_output_0" + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_69", + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_69" ], "out_args": [ "/transformer_blocks.17/norm1/norm/LayerNormalization_output_0.out14_69_bfp.out15_69" @@ -32951,7 +32951,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -32959,7 +32959,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33014,7 +33014,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -33022,7 +33022,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33044,7 +33044,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33094,7 +33094,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -33102,7 +33102,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33124,7 +33124,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33189,16 +33189,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33206,7 +33200,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33228,7 +33222,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33236,11 +33230,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -33295,16 +33295,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33312,7 +33306,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33334,7 +33328,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33342,11 +33336,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -33395,16 +33395,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33412,7 +33406,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33434,7 +33428,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -33442,11 +33436,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -33503,7 +33503,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -33549,7 +33549,7 @@ "/transformer_blocks.17/attn/Reshape_3_output_0.out22_17_bfp.out27_0_17" ], "const_args": [ - "onnx::MatMul_8350" + "onnx::MatMul_8350_45_5_34" ], "out_args": [ "/transformer_blocks.17/attn/to_out.0/Add_output_0.out6_1_34_bfp.out7_34" @@ -33564,7 +33564,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -33590,7 +33590,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33598,7 +33598,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33656,7 +33656,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -33664,7 +33664,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33672,7 +33672,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33687,7 +33687,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33728,8 +33728,8 @@ "/transformer_blocks.17/Add_output_0.out10_34" ], "const_args": [ - "/transformer_blocks.17/norm2/Constant_output_0", - "/transformer_blocks.17/norm2/Constant_1_output_0" + "/transformer_blocks.17/norm2/Constant_output_0_14_70", + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70" ], "out_args": [ "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70" @@ -33756,7 +33756,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33764,7 +33764,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33819,7 +33819,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -33827,7 +33827,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33849,7 +33849,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33889,7 +33889,7 @@ "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35" ], "const_args": [ - "onnx::MatMul_8352" + "onnx::MatMul_8352_45_1_69" ], "out_args": [ "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_73" @@ -33898,7 +33898,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -33906,7 +33906,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -33959,16 +33959,16 @@ "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_73" ], "const_args": [ - "onnx::MatMul_8353" + "onnx::MatMul_8353_45_1_71" ], "out_args": [ - "/transformer_blocks.17/ff/net.2/Add_output_0.out17_3_147_bfp.out25_75" + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -33976,7 +33976,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34021,7 +34021,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.17/ff/net.2/Add_output_0.out17_3_147_bfp.out25_75", + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75", "/transformer_blocks.17/Add_output_0.out10_34" ], "const_args": [ @@ -34034,7 +34034,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -34042,7 +34042,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34050,7 +34050,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34065,7 +34065,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34114,7 +34114,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34122,7 +34122,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34130,7 +34130,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34169,7 +34169,7 @@ "/transformer_blocks.17/attn/Reshape_3_output_0.out22_17_bfp.out27_0_17" ], "const_args": [ - "onnx::MatMul_8351" + "onnx::MatMul_8351_45_5_35" ], "out_args": [ "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35" @@ -34184,7 +34184,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -34210,7 +34210,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34218,7 +34218,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34276,7 +34276,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -34284,7 +34284,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34292,7 +34292,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34307,7 +34307,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34348,8 +34348,8 @@ "/transformer_blocks.17/Add_4_output_0.out10_36" ], "const_args": [ - "/transformer_blocks.17/norm2_context/Constant_output_0", - "/transformer_blocks.17/norm2_context/Constant_1_output_0" + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71", + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71" ], "out_args": [ "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71" @@ -34376,7 +34376,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34384,7 +34384,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34439,7 +34439,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -34447,7 +34447,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34469,7 +34469,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34509,7 +34509,7 @@ "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37" ], "const_args": [ - "onnx::MatMul_8354" + "onnx::MatMul_8354_45_1_70" ], "out_args": [ "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_74" @@ -34518,7 +34518,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34526,7 +34526,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -34579,16 +34579,16 @@ "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_74" ], "const_args": [ - "onnx::MatMul_8355" + "onnx::MatMul_8355_45_1_72" ], "out_args": [ - "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_3_148_bfp.out25_76" + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -34596,7 +34596,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34641,7 +34641,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_3_148_bfp.out25_76", + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76", "/transformer_blocks.17/Add_4_output_0.out10_36" ], "const_args": [ @@ -34654,7 +34654,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -34662,7 +34662,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34670,7 +34670,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34685,7 +34685,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34726,8 +34726,8 @@ "/Add_17_output_0.out_35_1_21" ], "const_args": [ - "/transformer_blocks.18/norm1/norm/Constant_output_0", - "/transformer_blocks.18/norm1/norm/Constant_1_output_0" + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_73", + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_73" ], "out_args": [ "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_73_bfp.out15_73" @@ -34754,7 +34754,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34762,7 +34762,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34817,7 +34817,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -34825,7 +34825,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34847,7 +34847,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -34887,8 +34887,8 @@ "/transformer_blocks.17/Add_7_output_0.out10_37" ], "const_args": [ - "/transformer_blocks.18/norm1_context/norm/Constant_output_0", - "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_72", + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_72" ], "out_args": [ "/transformer_blocks.18/norm1_context/norm/LayerNormalization_output_0.out14_72_bfp.out15_72" @@ -34915,7 +34915,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34923,7 +34923,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -34978,7 +34978,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -34986,7 +34986,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35008,7 +35008,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35073,16 +35073,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35090,7 +35084,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35112,7 +35106,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35120,11 +35114,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -35179,16 +35179,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35196,7 +35190,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35218,7 +35212,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35226,11 +35220,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -35279,16 +35279,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35296,7 +35290,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35318,7 +35312,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -35326,11 +35320,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -35387,7 +35387,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -35433,7 +35433,7 @@ "/transformer_blocks.18/attn/Reshape_3_output_0.out22_18_bfp.out27_0_18" ], "const_args": [ - "onnx::MatMul_8375" + "onnx::MatMul_8375_45_5_36" ], "out_args": [ "/transformer_blocks.18/attn/to_out.0/Add_output_0.out6_1_36_bfp.out7_36" @@ -35448,7 +35448,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -35474,7 +35474,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35482,7 +35482,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35540,7 +35540,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -35548,7 +35548,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35556,7 +35556,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35571,7 +35571,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35612,8 +35612,8 @@ "/transformer_blocks.18/Add_output_0.out10_38" ], "const_args": [ - "/transformer_blocks.18/norm2/Constant_output_0", - "/transformer_blocks.18/norm2/Constant_1_output_0" + "/transformer_blocks.18/norm2/Constant_output_0_14_74", + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74" ], "out_args": [ "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74" @@ -35640,7 +35640,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35648,7 +35648,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35703,7 +35703,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -35711,7 +35711,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35733,7 +35733,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35773,7 +35773,7 @@ "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39" ], "const_args": [ - "onnx::MatMul_8377" + "onnx::MatMul_8377_45_1_73" ], "out_args": [ "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_77" @@ -35782,7 +35782,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35790,7 +35790,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -35843,16 +35843,16 @@ "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_77" ], "const_args": [ - "onnx::MatMul_8378" + "onnx::MatMul_8378_45_1_75" ], "out_args": [ - "/transformer_blocks.18/ff/net.2/Add_output_0.out17_3_155_bfp.out25_79" + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -35860,7 +35860,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35905,7 +35905,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.18/ff/net.2/Add_output_0.out17_3_155_bfp.out25_79", + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79", "/transformer_blocks.18/Add_output_0.out10_38" ], "const_args": [ @@ -35918,7 +35918,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -35926,7 +35926,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35934,7 +35934,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35949,7 +35949,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -35990,7 +35990,7 @@ "/transformer_blocks.18/attn/Reshape_3_output_0.out22_18_bfp.out27_0_18" ], "const_args": [ - "onnx::MatMul_8376" + "onnx::MatMul_8376_45_5_37" ], "out_args": [ "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37" @@ -36005,7 +36005,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -36031,7 +36031,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36039,7 +36039,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36097,7 +36097,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -36105,7 +36105,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36113,7 +36113,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36128,7 +36128,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36169,8 +36169,8 @@ "/transformer_blocks.18/Add_4_output_0.out10_40" ], "const_args": [ - "/transformer_blocks.18/norm2_context/Constant_output_0", - "/transformer_blocks.18/norm2_context/Constant_1_output_0" + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75", + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75" ], "out_args": [ "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75" @@ -36197,7 +36197,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36205,7 +36205,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36260,7 +36260,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -36268,7 +36268,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36290,7 +36290,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36330,7 +36330,7 @@ "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41" ], "const_args": [ - "onnx::MatMul_8379" + "onnx::MatMul_8379_45_1_74" ], "out_args": [ "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_78" @@ -36339,7 +36339,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36347,7 +36347,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -36400,16 +36400,16 @@ "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_78" ], "const_args": [ - "onnx::MatMul_8380" + "onnx::MatMul_8380_45_1_76" ], "out_args": [ - "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_3_156_bfp.out25_80" + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -36417,7 +36417,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36462,7 +36462,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_3_156_bfp.out25_80", + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80", "/transformer_blocks.18/Add_4_output_0.out10_40" ], "const_args": [ @@ -36475,7 +36475,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -36483,7 +36483,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36491,7 +36491,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36506,7 +36506,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36547,8 +36547,8 @@ "/transformer_blocks.18/Add_7_output_0.out10_41" ], "const_args": [ - "/transformer_blocks.19/norm1_context/norm/Constant_output_0", - "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_76", + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_76" ], "out_args": [ "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_76_bfp.out15_76" @@ -36575,7 +36575,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36583,7 +36583,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36636,7 +36636,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36644,7 +36644,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36652,7 +36652,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36691,8 +36691,8 @@ "/Add_18_output_0.out_35_1_22" ], "const_args": [ - "/transformer_blocks.19/norm1/norm/Constant_output_0", - "/transformer_blocks.19/norm1/norm/Constant_1_output_0" + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_77", + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_77" ], "out_args": [ "/transformer_blocks.19/norm1/norm/LayerNormalization_output_0.out14_77_bfp.out15_77" @@ -36719,7 +36719,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36727,7 +36727,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36782,7 +36782,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -36790,7 +36790,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36812,7 +36812,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36862,7 +36862,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -36870,7 +36870,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36892,7 +36892,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36957,16 +36957,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -36974,7 +36968,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -36996,7 +36990,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -37004,11 +36998,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -37063,16 +37063,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -37080,7 +37074,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37102,7 +37096,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -37110,11 +37104,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -37163,16 +37163,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -37180,7 +37174,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37202,7 +37196,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -37210,11 +37204,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -37271,7 +37271,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -37317,7 +37317,7 @@ "/transformer_blocks.19/attn/Reshape_3_output_0.out22_19_bfp.out27_0_19" ], "const_args": [ - "onnx::MatMul_8400" + "onnx::MatMul_8400_45_5_38" ], "out_args": [ "/transformer_blocks.19/attn/to_out.0/Add_output_0.out6_1_38_bfp.out7_38" @@ -37332,7 +37332,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -37358,7 +37358,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37366,7 +37366,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37424,7 +37424,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -37432,7 +37432,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37440,7 +37440,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37455,7 +37455,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37496,8 +37496,8 @@ "/transformer_blocks.19/Add_output_0.out10_42" ], "const_args": [ - "/transformer_blocks.19/norm2/Constant_output_0", - "/transformer_blocks.19/norm2/Constant_1_output_0" + "/transformer_blocks.19/norm2/Constant_output_0_14_78", + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78" ], "out_args": [ "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78" @@ -37524,7 +37524,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37532,7 +37532,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37587,7 +37587,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -37595,7 +37595,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37617,7 +37617,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37657,7 +37657,7 @@ "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43" ], "const_args": [ - "onnx::MatMul_8402" + "onnx::MatMul_8402_45_1_77" ], "out_args": [ "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_81" @@ -37666,7 +37666,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37674,7 +37674,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -37727,16 +37727,16 @@ "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_81" ], "const_args": [ - "onnx::MatMul_8403" + "onnx::MatMul_8403_45_1_79" ], "out_args": [ - "/transformer_blocks.19/ff/net.2/Add_output_0.out17_3_163_bfp.out25_83" + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -37744,7 +37744,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37789,7 +37789,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.19/ff/net.2/Add_output_0.out17_3_163_bfp.out25_83", + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83", "/transformer_blocks.19/Add_output_0.out10_42" ], "const_args": [ @@ -37802,7 +37802,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -37810,7 +37810,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37818,7 +37818,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37833,7 +37833,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37882,7 +37882,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37890,7 +37890,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37898,7 +37898,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -37937,7 +37937,7 @@ "/transformer_blocks.19/attn/Reshape_3_output_0.out22_19_bfp.out27_0_19" ], "const_args": [ - "onnx::MatMul_8401" + "onnx::MatMul_8401_45_5_39" ], "out_args": [ "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39" @@ -37952,7 +37952,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -37978,7 +37978,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -37986,7 +37986,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38044,7 +38044,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -38052,7 +38052,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38060,7 +38060,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38075,7 +38075,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38116,8 +38116,8 @@ "/transformer_blocks.19/Add_4_output_0.out10_44" ], "const_args": [ - "/transformer_blocks.19/norm2_context/Constant_output_0", - "/transformer_blocks.19/norm2_context/Constant_1_output_0" + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79", + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79" ], "out_args": [ "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79" @@ -38144,7 +38144,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38152,7 +38152,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38207,7 +38207,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -38215,7 +38215,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38237,7 +38237,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38277,7 +38277,7 @@ "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45" ], "const_args": [ - "onnx::MatMul_8404" + "onnx::MatMul_8404_45_1_78" ], "out_args": [ "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_82" @@ -38286,7 +38286,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38294,7 +38294,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -38347,16 +38347,16 @@ "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_82" ], "const_args": [ - "onnx::MatMul_8405" + "onnx::MatMul_8405_45_1_80" ], "out_args": [ - "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_3_164_bfp.out25_84" + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -38364,7 +38364,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38409,7 +38409,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_3_164_bfp.out25_84", + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84", "/transformer_blocks.19/Add_4_output_0.out10_44" ], "const_args": [ @@ -38422,7 +38422,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -38430,7 +38430,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38438,7 +38438,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38453,7 +38453,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38494,8 +38494,8 @@ "/Add_19_output_0.out_35_1_23" ], "const_args": [ - "/transformer_blocks.20/norm1/norm/Constant_output_0", - "/transformer_blocks.20/norm1/norm/Constant_1_output_0" + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_81", + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_81" ], "out_args": [ "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_81_bfp.out15_81" @@ -38522,7 +38522,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -38530,7 +38530,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -38575,8 +38575,8 @@ "/transformer_blocks.19/Add_7_output_0.out10_45" ], "const_args": [ - "/transformer_blocks.20/norm1_context/norm/Constant_output_0", - "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_80", + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_80" ], "out_args": [ "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_80_bfp.out15_80" @@ -38603,7 +38603,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38611,7 +38611,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38666,7 +38666,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -38674,7 +38674,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -38696,7 +38696,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -38746,7 +38746,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -38754,7 +38754,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38776,7 +38776,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38841,16 +38841,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38858,7 +38852,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -38880,7 +38874,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38888,11 +38882,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -38947,16 +38947,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38964,7 +38958,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -38986,7 +38980,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -38994,11 +38988,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -39047,16 +39047,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39064,7 +39058,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39086,7 +39080,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39094,11 +39088,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -39155,7 +39155,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -39201,7 +39201,7 @@ "/transformer_blocks.20/attn/Reshape_3_output_0.out22_20_bfp.out27_0_20" ], "const_args": [ - "onnx::MatMul_8426" + "onnx::MatMul_8426_45_5_41" ], "out_args": [ "/transformer_blocks.20/attn/to_add_out/Add_output_0.out6_1_41_bfp.out7_41" @@ -39216,7 +39216,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -39242,7 +39242,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39250,7 +39250,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39297,7 +39297,7 @@ "/transformer_blocks.20/attn/Reshape_3_output_0.out22_20_bfp.out27_0_20" ], "const_args": [ - "onnx::MatMul_8425" + "onnx::MatMul_8425_45_5_40" ], "out_args": [ "/transformer_blocks.20/attn/to_out.0/Add_output_0.out6_1_40_bfp.out7_40" @@ -39312,7 +39312,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -39338,7 +39338,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39346,7 +39346,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39404,7 +39404,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -39412,7 +39412,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39420,7 +39420,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39435,7 +39435,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39476,8 +39476,8 @@ "/transformer_blocks.20/Add_output_0.out10_50" ], "const_args": [ - "/transformer_blocks.20/norm2/Constant_output_0", - "/transformer_blocks.20/norm2/Constant_1_output_0" + "/transformer_blocks.20/norm2/Constant_output_0_14_82", + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82" ], "out_args": [ "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82" @@ -39504,7 +39504,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39512,7 +39512,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39567,7 +39567,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -39575,7 +39575,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39597,7 +39597,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39637,7 +39637,7 @@ "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51" ], "const_args": [ - "onnx::MatMul_8427" + "onnx::MatMul_8427_45_1_81" ], "out_args": [ "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_85" @@ -39646,7 +39646,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39654,7 +39654,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -39707,16 +39707,16 @@ "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_85" ], "const_args": [ - "onnx::MatMul_8428" + "onnx::MatMul_8428_45_1_83" ], "out_args": [ - "/transformer_blocks.20/ff/net.2/Add_output_0.out17_3_171_bfp.out25_87" + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -39724,7 +39724,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39769,7 +39769,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.20/ff/net.2/Add_output_0.out17_3_171_bfp.out25_87", + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87", "/transformer_blocks.20/Add_output_0.out10_50" ], "const_args": [ @@ -39782,7 +39782,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -39790,7 +39790,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39798,7 +39798,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39813,7 +39813,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -39865,7 +39865,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -39873,7 +39873,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39881,7 +39881,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39896,7 +39896,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39937,8 +39937,8 @@ "/transformer_blocks.20/Add_4_output_0.out10_52" ], "const_args": [ - "/transformer_blocks.20/norm2_context/Constant_output_0", - "/transformer_blocks.20/norm2_context/Constant_1_output_0" + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83", + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83" ], "out_args": [ "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83" @@ -39965,7 +39965,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -39973,7 +39973,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40028,7 +40028,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -40036,7 +40036,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40058,7 +40058,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40098,7 +40098,7 @@ "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53" ], "const_args": [ - "onnx::MatMul_8429" + "onnx::MatMul_8429_45_1_82" ], "out_args": [ "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_86" @@ -40107,7 +40107,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40115,7 +40115,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -40168,16 +40168,16 @@ "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_86" ], "const_args": [ - "onnx::MatMul_8430" + "onnx::MatMul_8430_45_1_84" ], "out_args": [ - "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_3_172_bfp.out25_88" + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -40185,7 +40185,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40230,7 +40230,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_3_172_bfp.out25_88", + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88", "/transformer_blocks.20/Add_4_output_0.out10_52" ], "const_args": [ @@ -40243,7 +40243,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -40251,7 +40251,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40259,7 +40259,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40274,7 +40274,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40315,8 +40315,8 @@ "/transformer_blocks.20/Add_7_output_0.out10_53" ], "const_args": [ - "/transformer_blocks.21/norm1_context/norm/Constant_output_0", - "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_84", + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_84" ], "out_args": [ "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_84_bfp.out15_84" @@ -40343,7 +40343,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40351,7 +40351,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40404,7 +40404,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40412,7 +40412,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40420,7 +40420,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40459,8 +40459,8 @@ "/Add_20_output_0.out_35_1_24" ], "const_args": [ - "/transformer_blocks.21/norm1/norm/Constant_output_0", - "/transformer_blocks.21/norm1/norm/Constant_1_output_0" + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_85", + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_85" ], "out_args": [ "/transformer_blocks.21/norm1/norm/LayerNormalization_output_0.out14_85_bfp.out15_85" @@ -40487,7 +40487,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40495,7 +40495,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40550,7 +40550,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -40558,7 +40558,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40580,7 +40580,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40630,7 +40630,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -40638,7 +40638,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40660,7 +40660,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40725,16 +40725,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40742,7 +40736,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40764,7 +40758,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40772,11 +40766,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -40831,16 +40831,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40848,7 +40842,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40870,7 +40864,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40878,11 +40872,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -40931,16 +40931,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40948,7 +40942,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -40970,7 +40964,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -40978,11 +40972,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -41039,7 +41039,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -41085,7 +41085,7 @@ "/transformer_blocks.21/attn/Reshape_3_output_0.out22_21_bfp.out27_0_21" ], "const_args": [ - "onnx::MatMul_8450" + "onnx::MatMul_8450_45_5_42" ], "out_args": [ "/transformer_blocks.21/attn/to_out.0/Add_output_0.out6_1_42_bfp.out7_42" @@ -41100,7 +41100,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -41126,7 +41126,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41134,7 +41134,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41192,7 +41192,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -41200,7 +41200,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41208,7 +41208,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41223,7 +41223,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41264,8 +41264,8 @@ "/transformer_blocks.21/Add_output_0.out10_54" ], "const_args": [ - "/transformer_blocks.21/norm2/Constant_output_0", - "/transformer_blocks.21/norm2/Constant_1_output_0" + "/transformer_blocks.21/norm2/Constant_output_0_14_86", + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86" ], "out_args": [ "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86" @@ -41292,7 +41292,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41300,7 +41300,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41355,7 +41355,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -41363,7 +41363,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41385,7 +41385,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41425,7 +41425,7 @@ "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55" ], "const_args": [ - "onnx::MatMul_8452" + "onnx::MatMul_8452_45_1_85" ], "out_args": [ "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_89" @@ -41434,7 +41434,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41442,7 +41442,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -41495,16 +41495,16 @@ "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_89" ], "const_args": [ - "onnx::MatMul_8453" + "onnx::MatMul_8453_45_1_87" ], "out_args": [ - "/transformer_blocks.21/ff/net.2/Add_output_0.out17_3_179_bfp.out25_91" + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -41512,7 +41512,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41557,7 +41557,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.21/ff/net.2/Add_output_0.out17_3_179_bfp.out25_91", + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91", "/transformer_blocks.21/Add_output_0.out10_54" ], "const_args": [ @@ -41570,7 +41570,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -41578,7 +41578,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41586,7 +41586,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41601,7 +41601,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41650,7 +41650,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41658,7 +41658,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41666,7 +41666,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -41705,7 +41705,7 @@ "/transformer_blocks.21/attn/Reshape_3_output_0.out22_21_bfp.out27_0_21" ], "const_args": [ - "onnx::MatMul_8451" + "onnx::MatMul_8451_45_5_43" ], "out_args": [ "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43" @@ -41720,7 +41720,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -41746,7 +41746,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -41754,7 +41754,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -41812,7 +41812,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -41820,7 +41820,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -41828,7 +41828,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -41843,7 +41843,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -41884,8 +41884,8 @@ "/transformer_blocks.21/Add_4_output_0.out10_56" ], "const_args": [ - "/transformer_blocks.21/norm2_context/Constant_output_0", - "/transformer_blocks.21/norm2_context/Constant_1_output_0" + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87", + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87" ], "out_args": [ "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87" @@ -41912,7 +41912,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -41920,7 +41920,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -41975,7 +41975,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -41983,7 +41983,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42005,7 +42005,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42045,7 +42045,7 @@ "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57" ], "const_args": [ - "onnx::MatMul_8454" + "onnx::MatMul_8454_45_1_86" ], "out_args": [ "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_90" @@ -42054,7 +42054,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42062,7 +42062,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -42115,16 +42115,16 @@ "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_90" ], "const_args": [ - "onnx::MatMul_8455" + "onnx::MatMul_8455_45_1_88" ], "out_args": [ - "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_3_180_bfp.out25_92" + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -42132,7 +42132,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42177,7 +42177,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_3_180_bfp.out25_92", + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92", "/transformer_blocks.21/Add_4_output_0.out10_56" ], "const_args": [ @@ -42190,7 +42190,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -42198,7 +42198,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42206,7 +42206,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42221,7 +42221,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42262,8 +42262,8 @@ "/Add_21_output_0.out_35_1_25" ], "const_args": [ - "/transformer_blocks.22/norm1/norm/Constant_output_0", - "/transformer_blocks.22/norm1/norm/Constant_1_output_0" + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_89", + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_89" ], "out_args": [ "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_89_bfp.out15_89" @@ -42290,7 +42290,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -42298,7 +42298,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -42353,7 +42353,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -42361,7 +42361,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -42383,7 +42383,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -42423,8 +42423,8 @@ "/transformer_blocks.21/Add_7_output_0.out10_57" ], "const_args": [ - "/transformer_blocks.22/norm1_context/norm/Constant_output_0", - "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_88", + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_88" ], "out_args": [ "/transformer_blocks.22/norm1_context/norm/LayerNormalization_output_0.out14_88_bfp.out15_88" @@ -42451,7 +42451,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42459,7 +42459,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42514,7 +42514,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -42522,7 +42522,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42544,7 +42544,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42609,16 +42609,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42626,7 +42620,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -42648,7 +42642,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42656,11 +42650,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -42715,16 +42715,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42732,7 +42726,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -42754,7 +42748,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42762,11 +42756,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -42815,16 +42815,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42832,7 +42826,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -42854,7 +42848,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -42862,11 +42856,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -42923,7 +42923,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -42969,7 +42969,7 @@ "/transformer_blocks.22/attn/Reshape_3_output_0.out22_22_bfp.out27_0_22" ], "const_args": [ - "onnx::MatMul_8475" + "onnx::MatMul_8475_45_5_44" ], "out_args": [ "/transformer_blocks.22/attn/to_out.0/Add_output_0.out6_1_44_bfp.out7_44" @@ -42984,7 +42984,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -43010,7 +43010,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43018,7 +43018,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43076,7 +43076,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -43084,7 +43084,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43092,7 +43092,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43107,7 +43107,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43148,8 +43148,8 @@ "/transformer_blocks.22/Add_output_0.out10_58" ], "const_args": [ - "/transformer_blocks.22/norm2/Constant_output_0", - "/transformer_blocks.22/norm2/Constant_1_output_0" + "/transformer_blocks.22/norm2/Constant_output_0_14_90", + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90" ], "out_args": [ "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90" @@ -43176,7 +43176,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43184,7 +43184,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43239,7 +43239,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -43247,7 +43247,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43269,7 +43269,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43309,7 +43309,7 @@ "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59" ], "const_args": [ - "onnx::MatMul_8477" + "onnx::MatMul_8477_45_1_89" ], "out_args": [ "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_93" @@ -43318,7 +43318,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43326,7 +43326,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -43379,16 +43379,16 @@ "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_93" ], "const_args": [ - "onnx::MatMul_8478" + "onnx::MatMul_8478_45_1_91" ], "out_args": [ - "/transformer_blocks.22/ff/net.2/Add_output_0.out17_3_187_bfp.out25_95" + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -43396,7 +43396,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43441,7 +43441,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.22/ff/net.2/Add_output_0.out17_3_187_bfp.out25_95", + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95", "/transformer_blocks.22/Add_output_0.out10_58" ], "const_args": [ @@ -43454,7 +43454,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -43462,7 +43462,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43470,7 +43470,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43485,7 +43485,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -43526,7 +43526,7 @@ "/transformer_blocks.22/attn/Reshape_3_output_0.out22_22_bfp.out27_0_22" ], "const_args": [ - "onnx::MatMul_8476" + "onnx::MatMul_8476_45_5_45" ], "out_args": [ "/transformer_blocks.22/attn/to_add_out/Add_output_0.out6_1_45_bfp.out7_45" @@ -43541,7 +43541,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -43567,7 +43567,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43575,7 +43575,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43633,7 +43633,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -43641,7 +43641,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43649,7 +43649,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43664,7 +43664,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43705,8 +43705,8 @@ "/transformer_blocks.22/Add_4_output_0.out10_60" ], "const_args": [ - "/transformer_blocks.22/norm2_context/Constant_output_0", - "/transformer_blocks.22/norm2_context/Constant_1_output_0" + "/transformer_blocks.22/norm2_context/Constant_output_0_14_91", + "/transformer_blocks.22/norm2_context/Constant_1_output_0_14_91" ], "out_args": [ "/transformer_blocks.22/norm2_context/LayerNormalization_output_0.out14_91_bfp.out15_91" @@ -43733,7 +43733,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43741,7 +43741,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43796,7 +43796,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -43804,7 +43804,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43826,7 +43826,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43866,7 +43866,7 @@ "/transformer_blocks.22/Add_6_output_0.out0_0_91_bfp.out1_61" ], "const_args": [ - "onnx::MatMul_8479" + "onnx::MatMul_8479_45_1_90" ], "out_args": [ "/transformer_blocks.22/ff_context/net.0/Mul_5_output_0.out17_2_45_bfp.out25_94" @@ -43875,7 +43875,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43883,7 +43883,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -43936,16 +43936,16 @@ "/transformer_blocks.22/ff_context/net.0/Mul_5_output_0.out17_2_45_bfp.out25_94" ], "const_args": [ - "onnx::MatMul_8480" + "onnx::MatMul_8480_45_1_92" ], "out_args": [ - "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_3_188_bfp.out25_96" + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "6144" ] @@ -43953,7 +43953,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -43998,7 +43998,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_3_188_bfp.out25_96", + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96", "/transformer_blocks.22/Add_4_output_0.out10_60" ], "const_args": [ @@ -44011,7 +44011,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -44019,7 +44019,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44027,7 +44027,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44042,7 +44042,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44083,8 +44083,8 @@ "/transformer_blocks.22/Add_7_output_0.out10_61" ], "const_args": [ - "/transformer_blocks.23/norm1_context/norm/Constant_output_0", - "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0" + "/transformer_blocks.23/norm1_context/norm/Constant_output_0_14_92", + "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0_14_92" ], "out_args": [ "/transformer_blocks.23/norm1_context/norm/LayerNormalization_output_0.out14_92_bfp.out15_92" @@ -44111,7 +44111,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44119,7 +44119,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44172,7 +44172,7 @@ "a_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44180,7 +44180,7 @@ "b_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44188,7 +44188,7 @@ "c_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44227,8 +44227,8 @@ "/Add_22_output_0.out_35_1_26" ], "const_args": [ - "/transformer_blocks.23/norm1/norm/Constant_output_0", - "/transformer_blocks.23/norm1/norm/Constant_1_output_0" + "/transformer_blocks.23/norm1/norm/Constant_output_0_14_93", + "/transformer_blocks.23/norm1/norm/Constant_1_output_0_14_93" ], "out_args": [ "/transformer_blocks.23/norm1/norm/LayerNormalization_output_0.out14_93_bfp.out15_93" @@ -44255,7 +44255,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44263,7 +44263,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44318,7 +44318,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -44326,7 +44326,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44348,7 +44348,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44398,7 +44398,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -44406,7 +44406,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44428,7 +44428,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44493,16 +44493,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44510,7 +44504,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44532,7 +44526,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44540,11 +44534,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -44599,16 +44599,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44616,7 +44610,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44638,7 +44632,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44646,11 +44640,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -44699,16 +44699,10 @@ "1" ] }, - "bias_enable": { - "type": "int", - "value": [ - "1" - ] - }, "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44716,7 +44710,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44738,7 +44732,7 @@ "output_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length", "1536" ] @@ -44746,11 +44740,17 @@ "output_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, "in_dtypes": { "type": "str", "value": [ @@ -44807,7 +44807,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "64", "max_length + state_dim1" @@ -44853,7 +44853,7 @@ "/transformer_blocks.23/attn/Reshape_3_output_0.out22_23_bfp.out27_0_23" ], "const_args": [ - "onnx::MatMul_8497" + "onnx::MatMul_8497_45_5_46" ], "out_args": [ "/transformer_blocks.23/attn/to_out.0/Add_output_0.out6_1_46_bfp.out7_46" @@ -44868,7 +44868,7 @@ "slice_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "max_length + state_dim1", "1536" ] @@ -44894,7 +44894,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44902,7 +44902,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44960,7 +44960,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -44968,7 +44968,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44976,7 +44976,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -44991,7 +44991,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45032,8 +45032,8 @@ "/transformer_blocks.23/Add_output_0.out10_62" ], "const_args": [ - "/transformer_blocks.23/norm2/Constant_output_0", - "/transformer_blocks.23/norm2/Constant_1_output_0" + "/transformer_blocks.23/norm2/Constant_output_0_14_94", + "/transformer_blocks.23/norm2/Constant_1_output_0_14_94" ], "out_args": [ "/transformer_blocks.23/norm2/LayerNormalization_output_0.out14_94_bfp.out15_94" @@ -45060,7 +45060,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45068,7 +45068,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45123,7 +45123,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -45131,7 +45131,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45153,7 +45153,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45193,7 +45193,7 @@ "/transformer_blocks.23/Add_2_output_0.out0_0_94_bfp.out1_63" ], "const_args": [ - "onnx::MatMul_8498" + "onnx::MatMul_8498_45_1_93" ], "out_args": [ "/transformer_blocks.23/ff/net.0/Mul_5_output_0.out17_2_46_bfp.out25_97" @@ -45202,7 +45202,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45210,7 +45210,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -45263,16 +45263,16 @@ "/transformer_blocks.23/ff/net.0/Mul_5_output_0.out17_2_46_bfp.out25_97" ], "const_args": [ - "onnx::MatMul_8499" + "onnx::MatMul_8499_45_1_94" ], "out_args": [ - "/transformer_blocks.23/ff/net.2/Add_output_0.out17_3_195_bfp.out25_98" + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98" ], "attrs": { "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "6144" ] @@ -45280,7 +45280,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45325,7 +45325,7 @@ "type": "SDGemmMulAdd_bfpbfbf", "in_args": [ "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "/transformer_blocks.23/ff/net.2/Add_output_0.out17_3_195_bfp.out25_98", + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98", "/transformer_blocks.23/Add_output_0.out10_62" ], "const_args": [ @@ -45338,7 +45338,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -45346,7 +45346,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45354,7 +45354,7 @@ "input_shape_2": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45369,7 +45369,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45410,8 +45410,8 @@ "/transformer_blocks.23/Add_3_output_0.out10_63" ], "const_args": [ - "/norm_out/norm/Constant_output_0", - "/norm_out/norm/Constant_1_output_0" + "/norm_out/norm/Constant_output_0_14_95", + "/norm_out/norm/Constant_1_output_0_14_95" ], "out_args": [ "/norm_out/norm/LayerNormalization_output_0.out14_95_bfp.out15_95" @@ -45438,7 +45438,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45446,7 +45446,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45501,7 +45501,7 @@ "input_shape_0": { "type": "str", "value": [ - "batch_size", + "batch_4", "1", "1536" ] @@ -45509,7 +45509,7 @@ "input_shape_1": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45531,7 +45531,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45580,7 +45580,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45588,7 +45588,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45620,7 +45620,7 @@ "/norm_out/Add_2_output_0.out0_0_95" ], "const_args": [ - "onnx::MatMul_8500" + "onnx::MatMul_8500_45_0_4" ], "out_args": [ "/Reshape_output_0.out17_0_0" @@ -45629,7 +45629,7 @@ "input_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "1536" ] @@ -45637,7 +45637,7 @@ "output_shape": { "type": "str", "value": [ - "batch_size", + "batch_4", "state_dim1", "64" ] @@ -45680,14 +45680,14 @@ ], "fused_tensors": { "in": { - "buffer_size": 86560, + "buffer_size": 52768, "xrt_arg_id": 0, "packed_tensors": [ "hidden_states_nhwc.out5_0_0", "/pos_embed/Reshape_1_output_0.out_35_1_2", - "/time_text_embed/Cast_output_0.out17_3_3", - "pooled_projections.out17_3_1", - "encoder_hidden_states.out17_3_0", + "/time_text_embed/Cast_output_0.out17_4_3", + "pooled_projections.out17_4_1", + "encoder_hidden_states.out17_4_0", "block_controlnet_hidden_states_0.out_35_1_4", "block_controlnet_hidden_states_1.out_35_1_6", "block_controlnet_hidden_states_2.out_35_1_8", @@ -45715,17 +45715,17 @@ "packed_tensors": [ "/pos_embed/Transpose_output_0.out5_0_0", "/pos_embed/Add_2_output_0.out_35_1_2", - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3", + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3", "/time_text_embed/timestep_embedder/act/Sigmoid_output_0.out2_1", - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4", - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1", + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4", + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1", "/time_text_embed/text_embedder/act_1/Sigmoid_output_0.out2_0", - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2", + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2", "/time_text_embed/Add_output_0.out_35_1_3", "/transformer_blocks.0/norm1/silu/Sigmoid_output_0.out2_2", - "encoder_hidden_states.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0", - "/context_embedder/Add_output_0.out17_3_0", + "encoder_hidden_states.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0", + "/context_embedder/Add_output_0.out17_4_0", "/transformer_blocks.0/norm1_context/norm/LayerNormalization_output_0.out14_0_bfp.out15_0", "/transformer_blocks.0/norm1_context/Add_2_output_0.out0_0_0_bfp.out1_0", "/transformer_blocks.0/norm1/norm/LayerNormalization_output_0.out14_1_bfp.out15_1", @@ -45739,14 +45739,14 @@ "/transformer_blocks.0/norm2_context/LayerNormalization_output_0.out14_3_bfp.out15_3", "/transformer_blocks.0/Add_6_output_0.out0_0_3_bfp.out1_1", "/transformer_blocks.0/ff_context/net.0/Mul_5_output_0.out17_2_1_bfp.out25_6", - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8", + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8", "/transformer_blocks.0/Add_7_output_0.out10_1", "/transformer_blocks.0/attn/to_out.0/Add_output_0.out6_1_0_bfp.out7_0", "/transformer_blocks.0/Add_output_0.out10_92", "/transformer_blocks.0/norm2/LayerNormalization_output_0.out14_2_bfp.out15_2", "/transformer_blocks.0/Add_2_output_0.out0_0_2_bfp.out1_95", "/transformer_blocks.0/ff/net.0/Mul_5_output_0.out17_2_0_bfp.out25_5", - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7", + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7", "/transformer_blocks.0/Add_3_output_0.out10_93", "/transformer_blocks.1/norm1_context/norm/LayerNormalization_output_0.out14_4_bfp.out15_4", "/Add_output_0.out_35_1_4", @@ -45762,7 +45762,7 @@ "/transformer_blocks.1/norm2/LayerNormalization_output_0.out14_6_bfp.out15_6", "/transformer_blocks.1/Add_2_output_0.out0_0_6_bfp.out1_3", "/transformer_blocks.1/ff/net.0/Mul_5_output_0.out17_2_2_bfp.out25_9", - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_19_bfp.out25_11", + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11", "/transformer_blocks.1/Add_3_output_0.out10_3", "/Add_1_output_0.out_35_1_5", "/transformer_blocks.1/attn/to_add_out/Add_output_0.out6_1_3_bfp.out7_3", @@ -45770,7 +45770,7 @@ "/transformer_blocks.1/norm2_context/LayerNormalization_output_0.out14_7_bfp.out15_7", "/transformer_blocks.1/Add_6_output_0.out0_0_7_bfp.out1_5", "/transformer_blocks.1/ff_context/net.0/Mul_5_output_0.out17_2_3_bfp.out25_10", - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_20_bfp.out25_12", + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12", "/transformer_blocks.1/Add_7_output_0.out10_5", "/transformer_blocks.2/norm1/norm/LayerNormalization_output_0.out14_9_bfp.out15_9", "/transformer_blocks.2/norm1_context/norm/LayerNormalization_output_0.out14_8_bfp.out15_8", @@ -45787,11 +45787,11 @@ "/transformer_blocks.2/norm2/LayerNormalization_output_0.out14_10_bfp.out15_10", "/transformer_blocks.2/Add_2_output_0.out0_0_10_bfp.out1_47", "/transformer_blocks.2/ff/net.0/Mul_5_output_0.out17_2_4_bfp.out25_13", - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_27_bfp.out25_15", + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15", "/transformer_blocks.2/norm2_context/LayerNormalization_output_0.out14_11_bfp.out15_11", "/transformer_blocks.2/Add_6_output_0.out0_0_11_bfp.out1_49", "/transformer_blocks.2/ff_context/net.0/Mul_5_output_0.out17_2_5_bfp.out25_14", - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_28_bfp.out25_16", + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16", "/transformer_blocks.2/Add_7_output_0.out10_49", "/transformer_blocks.3/norm1_context/norm/LayerNormalization_output_0.out14_12_bfp.out15_12", "/transformer_blocks.2/Add_3_output_0.out10_47", @@ -45810,13 +45810,13 @@ "/transformer_blocks.3/norm2/LayerNormalization_output_0.out14_14_bfp.out15_14", "/transformer_blocks.3/Add_2_output_0.out0_0_14_bfp.out1_66", "/transformer_blocks.3/ff/net.0/Mul_5_output_0.out17_2_6_bfp.out25_17", - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_35_bfp.out25_19", + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19", "/transformer_blocks.3/Add_3_output_0.out10_65", "/Add_3_output_0.out_35_1_7", "/transformer_blocks.3/norm2_context/LayerNormalization_output_0.out14_15_bfp.out15_15", "/transformer_blocks.3/Add_6_output_0.out0_0_15_bfp.out1_68", "/transformer_blocks.3/ff_context/net.0/Mul_5_output_0.out17_2_7_bfp.out25_18", - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_36_bfp.out25_20", + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20", "/transformer_blocks.3/Add_7_output_0.out10_67", "/transformer_blocks.4/norm1/norm/LayerNormalization_output_0.out14_17_bfp.out15_17", "/transformer_blocks.4/norm1_context/norm/LayerNormalization_output_0.out14_16_bfp.out15_16", @@ -45833,11 +45833,11 @@ "/transformer_blocks.4/norm2/LayerNormalization_output_0.out14_18_bfp.out15_18", "/transformer_blocks.4/Add_2_output_0.out0_0_18_bfp.out1_70", "/transformer_blocks.4/ff/net.0/Mul_5_output_0.out17_2_8_bfp.out25_21", - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_43_bfp.out25_23", + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23", "/transformer_blocks.4/norm2_context/LayerNormalization_output_0.out14_19_bfp.out15_19", "/transformer_blocks.4/Add_6_output_0.out0_0_19_bfp.out1_72", "/transformer_blocks.4/ff_context/net.0/Mul_5_output_0.out17_2_9_bfp.out25_22", - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_44_bfp.out25_24", + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24", "/transformer_blocks.4/Add_7_output_0.out10_71", "/transformer_blocks.5/norm1_context/norm/LayerNormalization_output_0.out14_20_bfp.out15_20", "/transformer_blocks.4/Add_3_output_0.out10_69", @@ -45856,13 +45856,13 @@ "/transformer_blocks.5/norm2/LayerNormalization_output_0.out14_22_bfp.out15_22", "/transformer_blocks.5/Add_2_output_0.out0_0_22_bfp.out1_74", "/transformer_blocks.5/ff/net.0/Mul_5_output_0.out17_2_10_bfp.out25_25", - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_51_bfp.out25_27", + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27", "/transformer_blocks.5/Add_3_output_0.out10_73", "/Add_5_output_0.out_35_1_9", "/transformer_blocks.5/norm2_context/LayerNormalization_output_0.out14_23_bfp.out15_23", "/transformer_blocks.5/Add_6_output_0.out0_0_23_bfp.out1_76", "/transformer_blocks.5/ff_context/net.0/Mul_5_output_0.out17_2_11_bfp.out25_26", - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_52_bfp.out25_28", + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28", "/transformer_blocks.5/Add_7_output_0.out10_75", "/transformer_blocks.6/norm1/norm/LayerNormalization_output_0.out14_25_bfp.out15_25", "/transformer_blocks.6/norm1_context/norm/LayerNormalization_output_0.out14_24_bfp.out15_24", @@ -45879,11 +45879,11 @@ "/transformer_blocks.6/norm2/LayerNormalization_output_0.out14_26_bfp.out15_26", "/transformer_blocks.6/Add_2_output_0.out0_0_26_bfp.out1_78", "/transformer_blocks.6/ff/net.0/Mul_5_output_0.out17_2_12_bfp.out25_29", - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_59_bfp.out25_31", + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31", "/transformer_blocks.6/norm2_context/LayerNormalization_output_0.out14_27_bfp.out15_27", "/transformer_blocks.6/Add_6_output_0.out0_0_27_bfp.out1_80", "/transformer_blocks.6/ff_context/net.0/Mul_5_output_0.out17_2_13_bfp.out25_30", - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_60_bfp.out25_32", + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32", "/transformer_blocks.6/Add_7_output_0.out10_79", "/transformer_blocks.7/norm1_context/norm/LayerNormalization_output_0.out14_28_bfp.out15_28", "/transformer_blocks.6/Add_3_output_0.out10_77", @@ -45902,13 +45902,13 @@ "/transformer_blocks.7/norm2/LayerNormalization_output_0.out14_30_bfp.out15_30", "/transformer_blocks.7/Add_2_output_0.out0_0_30_bfp.out1_82", "/transformer_blocks.7/ff/net.0/Mul_5_output_0.out17_2_14_bfp.out25_33", - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_67_bfp.out25_35", + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35", "/transformer_blocks.7/Add_3_output_0.out10_81", "/Add_7_output_0.out_35_1_11", "/transformer_blocks.7/norm2_context/LayerNormalization_output_0.out14_31_bfp.out15_31", "/transformer_blocks.7/Add_6_output_0.out0_0_31_bfp.out1_84", "/transformer_blocks.7/ff_context/net.0/Mul_5_output_0.out17_2_15_bfp.out25_34", - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_68_bfp.out25_36", + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36", "/transformer_blocks.7/Add_7_output_0.out10_83", "/transformer_blocks.8/norm1/norm/LayerNormalization_output_0.out14_33_bfp.out15_33", "/transformer_blocks.8/norm1_context/norm/LayerNormalization_output_0.out14_32_bfp.out15_32", @@ -45925,11 +45925,11 @@ "/transformer_blocks.8/norm2/LayerNormalization_output_0.out14_34_bfp.out15_34", "/transformer_blocks.8/Add_2_output_0.out0_0_34_bfp.out1_86", "/transformer_blocks.8/ff/net.0/Mul_5_output_0.out17_2_16_bfp.out25_37", - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_75_bfp.out25_39", + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39", "/transformer_blocks.8/norm2_context/LayerNormalization_output_0.out14_35_bfp.out15_35", "/transformer_blocks.8/Add_6_output_0.out0_0_35_bfp.out1_88", "/transformer_blocks.8/ff_context/net.0/Mul_5_output_0.out17_2_17_bfp.out25_38", - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_76_bfp.out25_40", + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40", "/transformer_blocks.8/Add_7_output_0.out10_87", "/transformer_blocks.9/norm1_context/norm/LayerNormalization_output_0.out14_36_bfp.out15_36", "/transformer_blocks.8/Add_3_output_0.out10_85", @@ -45948,7 +45948,7 @@ "/transformer_blocks.9/norm2/LayerNormalization_output_0.out14_38_bfp.out15_38", "/transformer_blocks.9/Add_2_output_0.out0_0_38_bfp.out1_90", "/transformer_blocks.9/ff/net.0/Mul_5_output_0.out17_2_18_bfp.out25_41", - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_83_bfp.out25_43", + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43", "/transformer_blocks.9/Add_3_output_0.out10_89", "/Add_9_output_0.out_35_1_13", "/transformer_blocks.10/norm1/norm/LayerNormalization_output_0.out14_40_bfp.out15_40", @@ -45956,7 +45956,7 @@ "/transformer_blocks.9/norm2_context/LayerNormalization_output_0.out14_39_bfp.out15_39", "/transformer_blocks.9/Add_6_output_0.out0_0_39_bfp.out1_92", "/transformer_blocks.9/ff_context/net.0/Mul_5_output_0.out17_2_19_bfp.out25_42", - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_44", + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44", "/transformer_blocks.9/Add_7_output_0.out10_91", "/transformer_blocks.10/norm1_context/norm/LayerNormalization_output_0.out14_41_bfp.out15_41", "/transformer_blocks.10/norm1_context/Add_2_output_0.out0_0_41_bfp.out1_8", @@ -45969,14 +45969,14 @@ "/transformer_blocks.10/norm2/LayerNormalization_output_0.out14_42_bfp.out15_42", "/transformer_blocks.10/Add_2_output_0.out0_0_42_bfp.out1_7", "/transformer_blocks.10/ff/net.0/Mul_5_output_0.out17_2_20_bfp.out25_45", - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_91_bfp.out25_47", + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47", "/transformer_blocks.10/Add_3_output_0.out10_7", "/transformer_blocks.10/attn/to_add_out/Add_output_0.out6_1_21_bfp.out7_21", "/transformer_blocks.10/Add_4_output_0.out10_8", "/transformer_blocks.10/norm2_context/LayerNormalization_output_0.out14_43_bfp.out15_43", "/transformer_blocks.10/Add_6_output_0.out0_0_43_bfp.out1_9", "/transformer_blocks.10/ff_context/net.0/Mul_5_output_0.out17_2_21_bfp.out25_46", - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_92_bfp.out25_48", + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48", "/transformer_blocks.10/Add_7_output_0.out10_9", "/transformer_blocks.11/norm1_context/norm/LayerNormalization_output_0.out14_44_bfp.out15_44", "/Add_10_output_0.out_35_1_14", @@ -45992,7 +45992,7 @@ "/transformer_blocks.11/norm2/LayerNormalization_output_0.out14_46_bfp.out15_46", "/transformer_blocks.11/Add_2_output_0.out0_0_46_bfp.out1_11", "/transformer_blocks.11/ff/net.0/Mul_5_output_0.out17_2_22_bfp.out25_49", - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_99_bfp.out25_51", + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51", "/transformer_blocks.11/Add_3_output_0.out10_11", "/Add_11_output_0.out_35_1_15", "/transformer_blocks.11/attn/to_add_out/Add_output_0.out6_1_23_bfp.out7_23", @@ -46000,7 +46000,7 @@ "/transformer_blocks.11/norm2_context/LayerNormalization_output_0.out14_47_bfp.out15_47", "/transformer_blocks.11/Add_6_output_0.out0_0_47_bfp.out1_13", "/transformer_blocks.11/ff_context/net.0/Mul_5_output_0.out17_2_23_bfp.out25_50", - "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_3_100_bfp.out25_52", + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52", "/transformer_blocks.11/Add_7_output_0.out10_13", "/transformer_blocks.12/norm1/norm/LayerNormalization_output_0.out14_49_bfp.out15_49", "/transformer_blocks.12/norm1/Add_2_output_0.out0_0_49_bfp.out1_14", @@ -46015,14 +46015,14 @@ "/transformer_blocks.12/norm2/LayerNormalization_output_0.out14_50_bfp.out15_50", "/transformer_blocks.12/Add_2_output_0.out0_0_50_bfp.out1_15", "/transformer_blocks.12/ff/net.0/Mul_5_output_0.out17_2_24_bfp.out25_53", - "/transformer_blocks.12/ff/net.2/Add_output_0.out17_3_107_bfp.out25_55", + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55", "/transformer_blocks.12/Add_3_output_0.out10_15", "/transformer_blocks.12/attn/to_add_out/Add_output_0.out6_1_25_bfp.out7_25", "/transformer_blocks.12/Add_4_output_0.out10_16", "/transformer_blocks.12/norm2_context/LayerNormalization_output_0.out14_51_bfp.out15_51", "/transformer_blocks.12/Add_6_output_0.out0_0_51_bfp.out1_17", "/transformer_blocks.12/ff_context/net.0/Mul_5_output_0.out17_2_25_bfp.out25_54", - "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_3_108_bfp.out25_56", + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56", "/transformer_blocks.12/Add_7_output_0.out10_17", "/transformer_blocks.13/norm1_context/norm/LayerNormalization_output_0.out14_52_bfp.out15_52", "/Add_12_output_0.out_35_1_16", @@ -46038,7 +46038,7 @@ "/transformer_blocks.13/norm2/LayerNormalization_output_0.out14_54_bfp.out15_54", "/transformer_blocks.13/Add_2_output_0.out0_0_54_bfp.out1_19", "/transformer_blocks.13/ff/net.0/Mul_5_output_0.out17_2_26_bfp.out25_57", - "/transformer_blocks.13/ff/net.2/Add_output_0.out17_3_115_bfp.out25_59", + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59", "/transformer_blocks.13/Add_3_output_0.out10_19", "/Add_13_output_0.out_35_1_17", "/transformer_blocks.13/attn/to_add_out/Add_output_0.out6_1_27_bfp.out7_27", @@ -46046,7 +46046,7 @@ "/transformer_blocks.13/norm2_context/LayerNormalization_output_0.out14_55_bfp.out15_55", "/transformer_blocks.13/Add_6_output_0.out0_0_55_bfp.out1_21", "/transformer_blocks.13/ff_context/net.0/Mul_5_output_0.out17_2_27_bfp.out25_58", - "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_3_116_bfp.out25_60", + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60", "/transformer_blocks.13/Add_7_output_0.out10_21", "/transformer_blocks.14/norm1/norm/LayerNormalization_output_0.out14_57_bfp.out15_57", "/transformer_blocks.14/norm1/Add_2_output_0.out0_0_57_bfp.out1_22", @@ -46061,14 +46061,14 @@ "/transformer_blocks.14/norm2/LayerNormalization_output_0.out14_58_bfp.out15_58", "/transformer_blocks.14/Add_2_output_0.out0_0_58_bfp.out1_23", "/transformer_blocks.14/ff/net.0/Mul_5_output_0.out17_2_28_bfp.out25_61", - "/transformer_blocks.14/ff/net.2/Add_output_0.out17_3_123_bfp.out25_63", + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63", "/transformer_blocks.14/Add_3_output_0.out10_23", "/transformer_blocks.14/attn/to_add_out/Add_output_0.out6_1_29_bfp.out7_29", "/transformer_blocks.14/Add_4_output_0.out10_24", "/transformer_blocks.14/norm2_context/LayerNormalization_output_0.out14_59_bfp.out15_59", "/transformer_blocks.14/Add_6_output_0.out0_0_59_bfp.out1_25", "/transformer_blocks.14/ff_context/net.0/Mul_5_output_0.out17_2_29_bfp.out25_62", - "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_3_124_bfp.out25_64", + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64", "/transformer_blocks.14/Add_7_output_0.out10_25", "/transformer_blocks.15/norm1_context/norm/LayerNormalization_output_0.out14_60_bfp.out15_60", "/Add_14_output_0.out_35_1_18", @@ -46084,7 +46084,7 @@ "/transformer_blocks.15/norm2/LayerNormalization_output_0.out14_62_bfp.out15_62", "/transformer_blocks.15/Add_2_output_0.out0_0_62_bfp.out1_27", "/transformer_blocks.15/ff/net.0/Mul_5_output_0.out17_2_30_bfp.out25_65", - "/transformer_blocks.15/ff/net.2/Add_output_0.out17_3_131_bfp.out25_67", + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67", "/transformer_blocks.15/Add_3_output_0.out10_27", "/Add_15_output_0.out_35_1_19", "/transformer_blocks.15/attn/to_add_out/Add_output_0.out6_1_31_bfp.out7_31", @@ -46092,7 +46092,7 @@ "/transformer_blocks.15/norm2_context/LayerNormalization_output_0.out14_63_bfp.out15_63", "/transformer_blocks.15/Add_6_output_0.out0_0_63_bfp.out1_29", "/transformer_blocks.15/ff_context/net.0/Mul_5_output_0.out17_2_31_bfp.out25_66", - "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_3_132_bfp.out25_68", + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68", "/transformer_blocks.15/Add_7_output_0.out10_29", "/transformer_blocks.16/norm1/norm/LayerNormalization_output_0.out14_65_bfp.out15_65", "/transformer_blocks.16/norm1/Add_2_output_0.out0_0_65_bfp.out1_30", @@ -46107,14 +46107,14 @@ "/transformer_blocks.16/norm2/LayerNormalization_output_0.out14_66_bfp.out15_66", "/transformer_blocks.16/Add_2_output_0.out0_0_66_bfp.out1_31", "/transformer_blocks.16/ff/net.0/Mul_5_output_0.out17_2_32_bfp.out25_69", - "/transformer_blocks.16/ff/net.2/Add_output_0.out17_3_139_bfp.out25_71", + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71", "/transformer_blocks.16/Add_3_output_0.out10_31", "/transformer_blocks.16/attn/to_add_out/Add_output_0.out6_1_33_bfp.out7_33", "/transformer_blocks.16/Add_4_output_0.out10_32", "/transformer_blocks.16/norm2_context/LayerNormalization_output_0.out14_67_bfp.out15_67", "/transformer_blocks.16/Add_6_output_0.out0_0_67_bfp.out1_33", "/transformer_blocks.16/ff_context/net.0/Mul_5_output_0.out17_2_33_bfp.out25_70", - "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_3_140_bfp.out25_72", + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72", "/transformer_blocks.16/Add_7_output_0.out10_33", "/transformer_blocks.17/norm1_context/norm/LayerNormalization_output_0.out14_68_bfp.out15_68", "/Add_16_output_0.out_35_1_20", @@ -46130,7 +46130,7 @@ "/transformer_blocks.17/norm2/LayerNormalization_output_0.out14_70_bfp.out15_70", "/transformer_blocks.17/Add_2_output_0.out0_0_70_bfp.out1_35", "/transformer_blocks.17/ff/net.0/Mul_5_output_0.out17_2_34_bfp.out25_73", - "/transformer_blocks.17/ff/net.2/Add_output_0.out17_3_147_bfp.out25_75", + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75", "/transformer_blocks.17/Add_3_output_0.out10_35", "/Add_17_output_0.out_35_1_21", "/transformer_blocks.17/attn/to_add_out/Add_output_0.out6_1_35_bfp.out7_35", @@ -46138,7 +46138,7 @@ "/transformer_blocks.17/norm2_context/LayerNormalization_output_0.out14_71_bfp.out15_71", "/transformer_blocks.17/Add_6_output_0.out0_0_71_bfp.out1_37", "/transformer_blocks.17/ff_context/net.0/Mul_5_output_0.out17_2_35_bfp.out25_74", - "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_3_148_bfp.out25_76", + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76", "/transformer_blocks.17/Add_7_output_0.out10_37", "/transformer_blocks.18/norm1/norm/LayerNormalization_output_0.out14_73_bfp.out15_73", "/transformer_blocks.18/norm1/Add_2_output_0.out0_0_73_bfp.out1_38", @@ -46153,14 +46153,14 @@ "/transformer_blocks.18/norm2/LayerNormalization_output_0.out14_74_bfp.out15_74", "/transformer_blocks.18/Add_2_output_0.out0_0_74_bfp.out1_39", "/transformer_blocks.18/ff/net.0/Mul_5_output_0.out17_2_36_bfp.out25_77", - "/transformer_blocks.18/ff/net.2/Add_output_0.out17_3_155_bfp.out25_79", + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79", "/transformer_blocks.18/Add_3_output_0.out10_39", "/transformer_blocks.18/attn/to_add_out/Add_output_0.out6_1_37_bfp.out7_37", "/transformer_blocks.18/Add_4_output_0.out10_40", "/transformer_blocks.18/norm2_context/LayerNormalization_output_0.out14_75_bfp.out15_75", "/transformer_blocks.18/Add_6_output_0.out0_0_75_bfp.out1_41", "/transformer_blocks.18/ff_context/net.0/Mul_5_output_0.out17_2_37_bfp.out25_78", - "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_3_156_bfp.out25_80", + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80", "/transformer_blocks.18/Add_7_output_0.out10_41", "/transformer_blocks.19/norm1_context/norm/LayerNormalization_output_0.out14_76_bfp.out15_76", "/Add_18_output_0.out_35_1_22", @@ -46176,7 +46176,7 @@ "/transformer_blocks.19/norm2/LayerNormalization_output_0.out14_78_bfp.out15_78", "/transformer_blocks.19/Add_2_output_0.out0_0_78_bfp.out1_43", "/transformer_blocks.19/ff/net.0/Mul_5_output_0.out17_2_38_bfp.out25_81", - "/transformer_blocks.19/ff/net.2/Add_output_0.out17_3_163_bfp.out25_83", + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83", "/transformer_blocks.19/Add_3_output_0.out10_43", "/Add_19_output_0.out_35_1_23", "/transformer_blocks.19/attn/to_add_out/Add_output_0.out6_1_39_bfp.out7_39", @@ -46184,7 +46184,7 @@ "/transformer_blocks.19/norm2_context/LayerNormalization_output_0.out14_79_bfp.out15_79", "/transformer_blocks.19/Add_6_output_0.out0_0_79_bfp.out1_45", "/transformer_blocks.19/ff_context/net.0/Mul_5_output_0.out17_2_39_bfp.out25_82", - "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_3_164_bfp.out25_84", + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84", "/transformer_blocks.19/Add_7_output_0.out10_45", "/transformer_blocks.20/norm1/norm/LayerNormalization_output_0.out14_81_bfp.out15_81", "/transformer_blocks.20/norm1_context/norm/LayerNormalization_output_0.out14_80_bfp.out15_80", @@ -46200,13 +46200,13 @@ "/transformer_blocks.20/norm2/LayerNormalization_output_0.out14_82_bfp.out15_82", "/transformer_blocks.20/Add_2_output_0.out0_0_82_bfp.out1_51", "/transformer_blocks.20/ff/net.0/Mul_5_output_0.out17_2_40_bfp.out25_85", - "/transformer_blocks.20/ff/net.2/Add_output_0.out17_3_171_bfp.out25_87", + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87", "/transformer_blocks.20/Add_3_output_0.out10_51", "/transformer_blocks.20/Add_4_output_0.out10_52", "/transformer_blocks.20/norm2_context/LayerNormalization_output_0.out14_83_bfp.out15_83", "/transformer_blocks.20/Add_6_output_0.out0_0_83_bfp.out1_53", "/transformer_blocks.20/ff_context/net.0/Mul_5_output_0.out17_2_41_bfp.out25_86", - "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_3_172_bfp.out25_88", + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88", "/transformer_blocks.20/Add_7_output_0.out10_53", "/transformer_blocks.21/norm1_context/norm/LayerNormalization_output_0.out14_84_bfp.out15_84", "/Add_20_output_0.out_35_1_24", @@ -46222,7 +46222,7 @@ "/transformer_blocks.21/norm2/LayerNormalization_output_0.out14_86_bfp.out15_86", "/transformer_blocks.21/Add_2_output_0.out0_0_86_bfp.out1_55", "/transformer_blocks.21/ff/net.0/Mul_5_output_0.out17_2_42_bfp.out25_89", - "/transformer_blocks.21/ff/net.2/Add_output_0.out17_3_179_bfp.out25_91", + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91", "/transformer_blocks.21/Add_3_output_0.out10_55", "/Add_21_output_0.out_35_1_25", "/transformer_blocks.21/attn/to_add_out/Add_output_0.out6_1_43_bfp.out7_43", @@ -46230,7 +46230,7 @@ "/transformer_blocks.21/norm2_context/LayerNormalization_output_0.out14_87_bfp.out15_87", "/transformer_blocks.21/Add_6_output_0.out0_0_87_bfp.out1_57", "/transformer_blocks.21/ff_context/net.0/Mul_5_output_0.out17_2_43_bfp.out25_90", - "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_3_180_bfp.out25_92", + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92", "/transformer_blocks.21/Add_7_output_0.out10_57", "/transformer_blocks.22/norm1/norm/LayerNormalization_output_0.out14_89_bfp.out15_89", "/transformer_blocks.22/norm1/Add_2_output_0.out0_0_89_bfp.out1_58", @@ -46245,14 +46245,14 @@ "/transformer_blocks.22/norm2/LayerNormalization_output_0.out14_90_bfp.out15_90", "/transformer_blocks.22/Add_2_output_0.out0_0_90_bfp.out1_59", "/transformer_blocks.22/ff/net.0/Mul_5_output_0.out17_2_44_bfp.out25_93", - "/transformer_blocks.22/ff/net.2/Add_output_0.out17_3_187_bfp.out25_95", + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95", "/transformer_blocks.22/Add_3_output_0.out10_59", "/transformer_blocks.22/attn/to_add_out/Add_output_0.out6_1_45_bfp.out7_45", "/transformer_blocks.22/Add_4_output_0.out10_60", "/transformer_blocks.22/norm2_context/LayerNormalization_output_0.out14_91_bfp.out15_91", "/transformer_blocks.22/Add_6_output_0.out0_0_91_bfp.out1_61", "/transformer_blocks.22/ff_context/net.0/Mul_5_output_0.out17_2_45_bfp.out25_94", - "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_3_188_bfp.out25_96", + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96", "/transformer_blocks.22/Add_7_output_0.out10_61", "/transformer_blocks.23/norm1_context/norm/LayerNormalization_output_0.out14_92_bfp.out15_92", "/Add_22_output_0.out_35_1_26", @@ -46268,7 +46268,7 @@ "/transformer_blocks.23/norm2/LayerNormalization_output_0.out14_94_bfp.out15_94", "/transformer_blocks.23/Add_2_output_0.out0_0_94_bfp.out1_63", "/transformer_blocks.23/ff/net.0/Mul_5_output_0.out17_2_46_bfp.out25_97", - "/transformer_blocks.23/ff/net.2/Add_output_0.out17_3_195_bfp.out25_98", + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98", "/transformer_blocks.23/Add_3_output_0.out10_63", "/norm_out/norm/LayerNormalization_output_0.out14_95_bfp.out15_95", "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93", @@ -46280,613 +46280,613 @@ "xrt_arg_id": 3, "packed_tensors": [ "existing_model.pos_embed.proj.weight", - "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0", + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_45_0_2", "/time_text_embed/timestep_embedder/act/Sigmoid.weights2_1", - "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1", - "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2", + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_45_0_3", + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_45_0_0", "/time_text_embed/text_embedder/act_1/Sigmoid.weights2_0", - "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3", + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_45_0_1", "/transformer_blocks.0/norm1/silu/Sigmoid.weights2_2", - "encoder_hidden_states.out17_3_0_bfp.wts", - "onnx::MatMul_7905", - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts", - "/transformer_blocks.0/norm1_context/norm/Constant_output_0", - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0", + "encoder_hidden_states.out17_4_0_bfp.wts", + "onnx::MatMul_7905_45_1_0", + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts", + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0", + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0", "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_0_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_1", - "/transformer_blocks.0/norm1/norm/Constant_output_0", - "/transformer_blocks.0/norm1/norm/Constant_1_output_0", + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1", + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1", "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_0_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_1", "onnx::MatMul_7909_onnx::MatMul_7906", "onnx::MatMul_7910_onnx::MatMul_7907", "onnx::MatMul_7911_onnx::MatMul_7908", - "onnx::MatMul_7926", + "onnx::MatMul_7926_45_5_1", "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_2_gma", - "/transformer_blocks.0/norm2_context/Constant_output_0", - "/transformer_blocks.0/norm2_context/Constant_1_output_0", + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3", + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3", "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_3_existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_4", - "onnx::MatMul_7929", - "onnx::MatMul_7930", + "onnx::MatMul_7929_45_1_2", + "onnx::MatMul_7930_45_1_4", "existing_model.transformer_blocks.0.norm1_context.linear.weight_5_1_5_27_1_5_gma", - "onnx::MatMul_7925", + "onnx::MatMul_7925_45_5_0", "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_2_gma", - "/transformer_blocks.0/norm2/Constant_output_0", - "/transformer_blocks.0/norm2/Constant_1_output_0", + "/transformer_blocks.0/norm2/Constant_output_0_14_2", + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2", "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_3_existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_4", - "onnx::MatMul_7927", - "onnx::MatMul_7928", + "onnx::MatMul_7927_45_1_1", + "onnx::MatMul_7928_45_1_3", "existing_model.transformer_blocks.0.norm1.linear.weight_5_1_4_27_0_5_gma", - "/transformer_blocks.1/norm1_context/norm/Constant_output_0", - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.1/norm1/norm/Constant_output_0", - "/transformer_blocks.1/norm1/norm/Constant_1_output_0", + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_4", + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_4", + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_5", + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_5", "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_0_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_1", "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_0_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_1", "onnx::MatMul_7934_onnx::MatMul_7931", "onnx::MatMul_7935_onnx::MatMul_7932", "onnx::MatMul_7936_onnx::MatMul_7933", - "onnx::MatMul_7950", + "onnx::MatMul_7950_45_5_2", "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_2_gma", - "/transformer_blocks.1/norm2/Constant_output_0", - "/transformer_blocks.1/norm2/Constant_1_output_0", + "/transformer_blocks.1/norm2/Constant_output_0_14_6", + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6", "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_3_existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_4", - "onnx::MatMul_7952", - "onnx::MatMul_7953", + "onnx::MatMul_7952_45_1_5", + "onnx::MatMul_7953_45_1_7", "existing_model.transformer_blocks.1.norm1.linear.weight_5_1_6_27_2_5_gma", - "onnx::MatMul_7951", + "onnx::MatMul_7951_45_5_3", "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_2_gma", - "/transformer_blocks.1/norm2_context/Constant_output_0", - "/transformer_blocks.1/norm2_context/Constant_1_output_0", + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7", + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7", "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_3_existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_4", - "onnx::MatMul_7954", - "onnx::MatMul_7955", + "onnx::MatMul_7954_45_1_6", + "onnx::MatMul_7955_45_1_8", "existing_model.transformer_blocks.1.norm1_context.linear.weight_5_1_7_27_3_5_gma", - "/transformer_blocks.2/norm1/norm/Constant_output_0", - "/transformer_blocks.2/norm1/norm/Constant_1_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_output_0", - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_9", + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_9", + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_8", + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_8", "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_0_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_1", "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_0_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_1", "onnx::MatMul_7959_onnx::MatMul_7956", "onnx::MatMul_7960_onnx::MatMul_7957", "onnx::MatMul_7961_onnx::MatMul_7958", - "onnx::MatMul_7976", + "onnx::MatMul_7976_45_5_5", "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_2_gma", - "onnx::MatMul_7975", + "onnx::MatMul_7975_45_5_4", "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_2_gma", - "/transformer_blocks.2/norm2/Constant_output_0", - "/transformer_blocks.2/norm2/Constant_1_output_0", + "/transformer_blocks.2/norm2/Constant_output_0_14_10", + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10", "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_3_existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_4", - "onnx::MatMul_7977", - "onnx::MatMul_7978", - "/transformer_blocks.2/norm2_context/Constant_output_0", - "/transformer_blocks.2/norm2_context/Constant_1_output_0", + "onnx::MatMul_7977_45_1_9", + "onnx::MatMul_7978_45_1_11", + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11", + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11", "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_3_existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_4", - "onnx::MatMul_7979", - "onnx::MatMul_7980", + "onnx::MatMul_7979_45_1_10", + "onnx::MatMul_7980_45_1_12", "existing_model.transformer_blocks.2.norm1_context.linear.weight_5_1_9_27_5_5_gma", - "/transformer_blocks.3/norm1_context/norm/Constant_output_0", - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_12", + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_12", "existing_model.transformer_blocks.2.norm1.linear.weight_5_1_8_27_4_5_gma", - "/transformer_blocks.3/norm1/norm/Constant_output_0", - "/transformer_blocks.3/norm1/norm/Constant_1_output_0", + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_13", + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_13", "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_0_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_1", "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_0_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_1", "onnx::MatMul_7984_onnx::MatMul_7981", "onnx::MatMul_7985_onnx::MatMul_7982", "onnx::MatMul_7986_onnx::MatMul_7983", - "onnx::MatMul_8001", + "onnx::MatMul_8001_45_5_7", "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_2_gma", - "onnx::MatMul_8000", + "onnx::MatMul_8000_45_5_6", "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_2_gma", - "/transformer_blocks.3/norm2/Constant_output_0", - "/transformer_blocks.3/norm2/Constant_1_output_0", + "/transformer_blocks.3/norm2/Constant_output_0_14_14", + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14", "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_3_existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_4", - "onnx::MatMul_8002", - "onnx::MatMul_8003", + "onnx::MatMul_8002_45_1_13", + "onnx::MatMul_8003_45_1_15", "existing_model.transformer_blocks.3.norm1.linear.weight_5_1_10_27_6_5_gma", - "/transformer_blocks.3/norm2_context/Constant_output_0", - "/transformer_blocks.3/norm2_context/Constant_1_output_0", + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15", + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15", "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_3_existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_4", - "onnx::MatMul_8004", - "onnx::MatMul_8005", + "onnx::MatMul_8004_45_1_14", + "onnx::MatMul_8005_45_1_16", "existing_model.transformer_blocks.3.norm1_context.linear.weight_5_1_11_27_7_5_gma", - "/transformer_blocks.4/norm1/norm/Constant_output_0", - "/transformer_blocks.4/norm1/norm/Constant_1_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_output_0", - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_17", + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_17", + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_16", + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_16", "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_0_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_1", "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_0_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_1", "onnx::MatMul_8009_onnx::MatMul_8006", "onnx::MatMul_8010_onnx::MatMul_8007", "onnx::MatMul_8011_onnx::MatMul_8008", - "onnx::MatMul_8026", + "onnx::MatMul_8026_45_5_9", "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_2_gma", - "onnx::MatMul_8025", + "onnx::MatMul_8025_45_5_8", "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_2_gma", - "/transformer_blocks.4/norm2/Constant_output_0", - "/transformer_blocks.4/norm2/Constant_1_output_0", + "/transformer_blocks.4/norm2/Constant_output_0_14_18", + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18", "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_3_existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_4", - "onnx::MatMul_8027", - "onnx::MatMul_8028", - "/transformer_blocks.4/norm2_context/Constant_output_0", - "/transformer_blocks.4/norm2_context/Constant_1_output_0", + "onnx::MatMul_8027_45_1_17", + "onnx::MatMul_8028_45_1_19", + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19", + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19", "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_3_existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_4", - "onnx::MatMul_8029", - "onnx::MatMul_8030", + "onnx::MatMul_8029_45_1_18", + "onnx::MatMul_8030_45_1_20", "existing_model.transformer_blocks.4.norm1_context.linear.weight_5_1_13_27_9_5_gma", - "/transformer_blocks.5/norm1_context/norm/Constant_output_0", - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_20", + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_20", "existing_model.transformer_blocks.4.norm1.linear.weight_5_1_12_27_8_5_gma", - "/transformer_blocks.5/norm1/norm/Constant_output_0", - "/transformer_blocks.5/norm1/norm/Constant_1_output_0", + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_21", + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_21", "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_0_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_1", "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_0_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_1", "onnx::MatMul_8034_onnx::MatMul_8031", "onnx::MatMul_8035_onnx::MatMul_8032", "onnx::MatMul_8036_onnx::MatMul_8033", - "onnx::MatMul_8051", + "onnx::MatMul_8051_45_5_11", "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_2_gma", - "onnx::MatMul_8050", + "onnx::MatMul_8050_45_5_10", "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_2_gma", - "/transformer_blocks.5/norm2/Constant_output_0", - "/transformer_blocks.5/norm2/Constant_1_output_0", + "/transformer_blocks.5/norm2/Constant_output_0_14_22", + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22", "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_3_existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_4", - "onnx::MatMul_8052", - "onnx::MatMul_8053", + "onnx::MatMul_8052_45_1_21", + "onnx::MatMul_8053_45_1_23", "existing_model.transformer_blocks.5.norm1.linear.weight_5_1_14_27_10_5_gma", - "/transformer_blocks.5/norm2_context/Constant_output_0", - "/transformer_blocks.5/norm2_context/Constant_1_output_0", + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23", + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23", "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_3_existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_4", - "onnx::MatMul_8054", - "onnx::MatMul_8055", + "onnx::MatMul_8054_45_1_22", + "onnx::MatMul_8055_45_1_24", "existing_model.transformer_blocks.5.norm1_context.linear.weight_5_1_15_27_11_5_gma", - "/transformer_blocks.6/norm1/norm/Constant_output_0", - "/transformer_blocks.6/norm1/norm/Constant_1_output_0", - "/transformer_blocks.6/norm1_context/norm/Constant_output_0", - "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_25", + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_25", + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_24", + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_24", "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_0_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_1", "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_0_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_1", "onnx::MatMul_8059_onnx::MatMul_8056", "onnx::MatMul_8060_onnx::MatMul_8057", "onnx::MatMul_8061_onnx::MatMul_8058", - "onnx::MatMul_8076", + "onnx::MatMul_8076_45_5_13", "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_2_gma", - "onnx::MatMul_8075", + "onnx::MatMul_8075_45_5_12", "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_2_gma", - "/transformer_blocks.6/norm2/Constant_output_0", - "/transformer_blocks.6/norm2/Constant_1_output_0", + "/transformer_blocks.6/norm2/Constant_output_0_14_26", + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26", "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_3_existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_4", - "onnx::MatMul_8077", - "onnx::MatMul_8078", - "/transformer_blocks.6/norm2_context/Constant_output_0", - "/transformer_blocks.6/norm2_context/Constant_1_output_0", + "onnx::MatMul_8077_45_1_25", + "onnx::MatMul_8078_45_1_27", + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27", + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27", "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_3_existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_4", - "onnx::MatMul_8079", - "onnx::MatMul_8080", + "onnx::MatMul_8079_45_1_26", + "onnx::MatMul_8080_45_1_28", "existing_model.transformer_blocks.6.norm1_context.linear.weight_5_1_17_27_13_5_gma", - "/transformer_blocks.7/norm1_context/norm/Constant_output_0", - "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_28", + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_28", "existing_model.transformer_blocks.6.norm1.linear.weight_5_1_16_27_12_5_gma", - "/transformer_blocks.7/norm1/norm/Constant_output_0", - "/transformer_blocks.7/norm1/norm/Constant_1_output_0", + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_29", + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_29", "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_0_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_1", "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_0_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_1", "onnx::MatMul_8084_onnx::MatMul_8081", "onnx::MatMul_8085_onnx::MatMul_8082", "onnx::MatMul_8086_onnx::MatMul_8083", - "onnx::MatMul_8101", + "onnx::MatMul_8101_45_5_15", "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_2_gma", - "onnx::MatMul_8100", + "onnx::MatMul_8100_45_5_14", "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_2_gma", - "/transformer_blocks.7/norm2/Constant_output_0", - "/transformer_blocks.7/norm2/Constant_1_output_0", + "/transformer_blocks.7/norm2/Constant_output_0_14_30", + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30", "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_3_existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_4", - "onnx::MatMul_8102", - "onnx::MatMul_8103", + "onnx::MatMul_8102_45_1_29", + "onnx::MatMul_8103_45_1_31", "existing_model.transformer_blocks.7.norm1.linear.weight_5_1_18_27_14_5_gma", - "/transformer_blocks.7/norm2_context/Constant_output_0", - "/transformer_blocks.7/norm2_context/Constant_1_output_0", + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31", + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31", "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_3_existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_4", - "onnx::MatMul_8104", - "onnx::MatMul_8105", + "onnx::MatMul_8104_45_1_30", + "onnx::MatMul_8105_45_1_32", "existing_model.transformer_blocks.7.norm1_context.linear.weight_5_1_19_27_15_5_gma", - "/transformer_blocks.8/norm1/norm/Constant_output_0", - "/transformer_blocks.8/norm1/norm/Constant_1_output_0", - "/transformer_blocks.8/norm1_context/norm/Constant_output_0", - "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_33", + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_33", + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_32", + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_32", "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_0_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_1", "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_0_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_1", "onnx::MatMul_8109_onnx::MatMul_8106", "onnx::MatMul_8110_onnx::MatMul_8107", "onnx::MatMul_8111_onnx::MatMul_8108", - "onnx::MatMul_8126", + "onnx::MatMul_8126_45_5_17", "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_2_gma", - "onnx::MatMul_8125", + "onnx::MatMul_8125_45_5_16", "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_2_gma", - "/transformer_blocks.8/norm2/Constant_output_0", - "/transformer_blocks.8/norm2/Constant_1_output_0", + "/transformer_blocks.8/norm2/Constant_output_0_14_34", + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34", "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_3_existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_4", - "onnx::MatMul_8127", - "onnx::MatMul_8128", - "/transformer_blocks.8/norm2_context/Constant_output_0", - "/transformer_blocks.8/norm2_context/Constant_1_output_0", + "onnx::MatMul_8127_45_1_33", + "onnx::MatMul_8128_45_1_35", + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35", + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35", "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_3_existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_4", - "onnx::MatMul_8129", - "onnx::MatMul_8130", + "onnx::MatMul_8129_45_1_34", + "onnx::MatMul_8130_45_1_36", "existing_model.transformer_blocks.8.norm1_context.linear.weight_5_1_21_27_17_5_gma", - "/transformer_blocks.9/norm1_context/norm/Constant_output_0", - "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_36", + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_36", "existing_model.transformer_blocks.8.norm1.linear.weight_5_1_20_27_16_5_gma", - "/transformer_blocks.9/norm1/norm/Constant_output_0", - "/transformer_blocks.9/norm1/norm/Constant_1_output_0", + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_37", + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_37", "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_0_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_1", "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_0_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_1", "onnx::MatMul_8134_onnx::MatMul_8131", "onnx::MatMul_8135_onnx::MatMul_8132", "onnx::MatMul_8136_onnx::MatMul_8133", - "onnx::MatMul_8151", + "onnx::MatMul_8151_45_5_19", "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_2_gma", - "onnx::MatMul_8150", + "onnx::MatMul_8150_45_5_18", "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_2_gma", - "/transformer_blocks.9/norm2/Constant_output_0", - "/transformer_blocks.9/norm2/Constant_1_output_0", + "/transformer_blocks.9/norm2/Constant_output_0_14_38", + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38", "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_3_existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_4", - "onnx::MatMul_8152", - "onnx::MatMul_8153", + "onnx::MatMul_8152_45_1_37", + "onnx::MatMul_8153_45_1_39", "existing_model.transformer_blocks.9.norm1.linear.weight_5_1_22_27_18_5_gma", - "/transformer_blocks.10/norm1/norm/Constant_output_0", - "/transformer_blocks.10/norm1/norm/Constant_1_output_0", + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40", + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40", "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_0_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_1", - "/transformer_blocks.9/norm2_context/Constant_output_0", - "/transformer_blocks.9/norm2_context/Constant_1_output_0", + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39", + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39", "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_3_existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_4", - "onnx::MatMul_8154", - "onnx::MatMul_8155", + "onnx::MatMul_8154_45_1_38", + "onnx::MatMul_8155_45_1_40", "existing_model.transformer_blocks.9.norm1_context.linear.weight_5_1_23_27_19_5_gma", - "/transformer_blocks.10/norm1_context/norm/Constant_output_0", - "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41", + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41", "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_0_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_1", "onnx::MatMul_8159_onnx::MatMul_8156", "onnx::MatMul_8160_onnx::MatMul_8157", "onnx::MatMul_8161_onnx::MatMul_8158", - "onnx::MatMul_8175", + "onnx::MatMul_8175_45_5_20", "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_2_gma", - "/transformer_blocks.10/norm2/Constant_output_0", - "/transformer_blocks.10/norm2/Constant_1_output_0", + "/transformer_blocks.10/norm2/Constant_output_0_14_42", + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42", "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_3_existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_4", - "onnx::MatMul_8177", - "onnx::MatMul_8178", + "onnx::MatMul_8177_45_1_41", + "onnx::MatMul_8178_45_1_43", "existing_model.transformer_blocks.10.norm1.linear.weight_5_1_24_27_20_5_gma", - "onnx::MatMul_8176", + "onnx::MatMul_8176_45_5_21", "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_2_gma", - "/transformer_blocks.10/norm2_context/Constant_output_0", - "/transformer_blocks.10/norm2_context/Constant_1_output_0", + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43", + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43", "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_3_existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_4", - "onnx::MatMul_8179", - "onnx::MatMul_8180", + "onnx::MatMul_8179_45_1_42", + "onnx::MatMul_8180_45_1_44", "existing_model.transformer_blocks.10.norm1_context.linear.weight_5_1_25_27_21_5_gma", - "/transformer_blocks.11/norm1_context/norm/Constant_output_0", - "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.11/norm1/norm/Constant_output_0", - "/transformer_blocks.11/norm1/norm/Constant_1_output_0", + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_44", + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_44", + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_45", + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_45", "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_0_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_1", "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_0_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_1", "onnx::MatMul_8184_onnx::MatMul_8181", "onnx::MatMul_8185_onnx::MatMul_8182", "onnx::MatMul_8186_onnx::MatMul_8183", - "onnx::MatMul_8200", + "onnx::MatMul_8200_45_5_22", "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_2_gma", - "/transformer_blocks.11/norm2/Constant_output_0", - "/transformer_blocks.11/norm2/Constant_1_output_0", + "/transformer_blocks.11/norm2/Constant_output_0_14_46", + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46", "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_3_existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_4", - "onnx::MatMul_8202", - "onnx::MatMul_8203", + "onnx::MatMul_8202_45_1_45", + "onnx::MatMul_8203_45_1_47", "existing_model.transformer_blocks.11.norm1.linear.weight_5_1_26_27_22_5_gma", - "onnx::MatMul_8201", + "onnx::MatMul_8201_45_5_23", "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_2_gma", - "/transformer_blocks.11/norm2_context/Constant_output_0", - "/transformer_blocks.11/norm2_context/Constant_1_output_0", + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47", + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47", "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_3_existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_4", - "onnx::MatMul_8204", - "onnx::MatMul_8205", + "onnx::MatMul_8204_45_1_46", + "onnx::MatMul_8205_45_1_48", "existing_model.transformer_blocks.11.norm1_context.linear.weight_5_1_27_27_23_5_gma", - "/transformer_blocks.12/norm1/norm/Constant_output_0", - "/transformer_blocks.12/norm1/norm/Constant_1_output_0", + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_49", + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_49", "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_0_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_1", - "/transformer_blocks.12/norm1_context/norm/Constant_output_0", - "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_48", + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_48", "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_0_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_1", "onnx::MatMul_8209_onnx::MatMul_8206", "onnx::MatMul_8210_onnx::MatMul_8207", "onnx::MatMul_8211_onnx::MatMul_8208", - "onnx::MatMul_8225", + "onnx::MatMul_8225_45_5_24", "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_2_gma", - "/transformer_blocks.12/norm2/Constant_output_0", - "/transformer_blocks.12/norm2/Constant_1_output_0", + "/transformer_blocks.12/norm2/Constant_output_0_14_50", + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50", "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_3_existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_4", - "onnx::MatMul_8227", - "onnx::MatMul_8228", + "onnx::MatMul_8227_45_1_49", + "onnx::MatMul_8228_45_1_51", "existing_model.transformer_blocks.12.norm1.linear.weight_5_1_28_27_24_5_gma", - "onnx::MatMul_8226", + "onnx::MatMul_8226_45_5_25", "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_2_gma", - "/transformer_blocks.12/norm2_context/Constant_output_0", - "/transformer_blocks.12/norm2_context/Constant_1_output_0", + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51", + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51", "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_3_existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_4", - "onnx::MatMul_8229", - "onnx::MatMul_8230", + "onnx::MatMul_8229_45_1_50", + "onnx::MatMul_8230_45_1_52", "existing_model.transformer_blocks.12.norm1_context.linear.weight_5_1_29_27_25_5_gma", - "/transformer_blocks.13/norm1_context/norm/Constant_output_0", - "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.13/norm1/norm/Constant_output_0", - "/transformer_blocks.13/norm1/norm/Constant_1_output_0", + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_52", + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_52", + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_53", + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_53", "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_0_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_1", "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_0_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_1", "onnx::MatMul_8234_onnx::MatMul_8231", "onnx::MatMul_8235_onnx::MatMul_8232", "onnx::MatMul_8236_onnx::MatMul_8233", - "onnx::MatMul_8250", + "onnx::MatMul_8250_45_5_26", "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_2_gma", - "/transformer_blocks.13/norm2/Constant_output_0", - "/transformer_blocks.13/norm2/Constant_1_output_0", + "/transformer_blocks.13/norm2/Constant_output_0_14_54", + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54", "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_3_existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_4", - "onnx::MatMul_8252", - "onnx::MatMul_8253", + "onnx::MatMul_8252_45_1_53", + "onnx::MatMul_8253_45_1_55", "existing_model.transformer_blocks.13.norm1.linear.weight_5_1_30_27_26_5_gma", - "onnx::MatMul_8251", + "onnx::MatMul_8251_45_5_27", "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_2_gma", - "/transformer_blocks.13/norm2_context/Constant_output_0", - "/transformer_blocks.13/norm2_context/Constant_1_output_0", + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55", + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55", "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_3_existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_4", - "onnx::MatMul_8254", - "onnx::MatMul_8255", + "onnx::MatMul_8254_45_1_54", + "onnx::MatMul_8255_45_1_56", "existing_model.transformer_blocks.13.norm1_context.linear.weight_5_1_31_27_27_5_gma", - "/transformer_blocks.14/norm1/norm/Constant_output_0", - "/transformer_blocks.14/norm1/norm/Constant_1_output_0", + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_57", + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_57", "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_0_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_1", - "/transformer_blocks.14/norm1_context/norm/Constant_output_0", - "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_56", + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_56", "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_0_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_1", "onnx::MatMul_8259_onnx::MatMul_8256", "onnx::MatMul_8260_onnx::MatMul_8257", "onnx::MatMul_8261_onnx::MatMul_8258", - "onnx::MatMul_8275", + "onnx::MatMul_8275_45_5_28", "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_2_gma", - "/transformer_blocks.14/norm2/Constant_output_0", - "/transformer_blocks.14/norm2/Constant_1_output_0", + "/transformer_blocks.14/norm2/Constant_output_0_14_58", + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58", "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_3_existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_4", - "onnx::MatMul_8277", - "onnx::MatMul_8278", + "onnx::MatMul_8277_45_1_57", + "onnx::MatMul_8278_45_1_59", "existing_model.transformer_blocks.14.norm1.linear.weight_5_1_32_27_28_5_gma", - "onnx::MatMul_8276", + "onnx::MatMul_8276_45_5_29", "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_2_gma", - "/transformer_blocks.14/norm2_context/Constant_output_0", - "/transformer_blocks.14/norm2_context/Constant_1_output_0", + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59", + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59", "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_3_existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_4", - "onnx::MatMul_8279", - "onnx::MatMul_8280", + "onnx::MatMul_8279_45_1_58", + "onnx::MatMul_8280_45_1_60", "existing_model.transformer_blocks.14.norm1_context.linear.weight_5_1_33_27_29_5_gma", - "/transformer_blocks.15/norm1_context/norm/Constant_output_0", - "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.15/norm1/norm/Constant_output_0", - "/transformer_blocks.15/norm1/norm/Constant_1_output_0", + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_60", + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_60", + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_61", + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_61", "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_0_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_1", "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_0_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_1", "onnx::MatMul_8284_onnx::MatMul_8281", "onnx::MatMul_8285_onnx::MatMul_8282", "onnx::MatMul_8286_onnx::MatMul_8283", - "onnx::MatMul_8300", + "onnx::MatMul_8300_45_5_30", "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_2_gma", - "/transformer_blocks.15/norm2/Constant_output_0", - "/transformer_blocks.15/norm2/Constant_1_output_0", + "/transformer_blocks.15/norm2/Constant_output_0_14_62", + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62", "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_3_existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_4", - "onnx::MatMul_8302", - "onnx::MatMul_8303", + "onnx::MatMul_8302_45_1_61", + "onnx::MatMul_8303_45_1_63", "existing_model.transformer_blocks.15.norm1.linear.weight_5_1_34_27_30_5_gma", - "onnx::MatMul_8301", + "onnx::MatMul_8301_45_5_31", "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_2_gma", - "/transformer_blocks.15/norm2_context/Constant_output_0", - "/transformer_blocks.15/norm2_context/Constant_1_output_0", + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63", + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63", "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_3_existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_4", - "onnx::MatMul_8304", - "onnx::MatMul_8305", + "onnx::MatMul_8304_45_1_62", + "onnx::MatMul_8305_45_1_64", "existing_model.transformer_blocks.15.norm1_context.linear.weight_5_1_35_27_31_5_gma", - "/transformer_blocks.16/norm1/norm/Constant_output_0", - "/transformer_blocks.16/norm1/norm/Constant_1_output_0", + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_65", + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_65", "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_0_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_1", - "/transformer_blocks.16/norm1_context/norm/Constant_output_0", - "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_64", + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_64", "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_0_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_1", "onnx::MatMul_8309_onnx::MatMul_8306", "onnx::MatMul_8310_onnx::MatMul_8307", "onnx::MatMul_8311_onnx::MatMul_8308", - "onnx::MatMul_8325", + "onnx::MatMul_8325_45_5_32", "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_2_gma", - "/transformer_blocks.16/norm2/Constant_output_0", - "/transformer_blocks.16/norm2/Constant_1_output_0", + "/transformer_blocks.16/norm2/Constant_output_0_14_66", + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66", "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_3_existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_4", - "onnx::MatMul_8327", - "onnx::MatMul_8328", + "onnx::MatMul_8327_45_1_65", + "onnx::MatMul_8328_45_1_67", "existing_model.transformer_blocks.16.norm1.linear.weight_5_1_36_27_32_5_gma", - "onnx::MatMul_8326", + "onnx::MatMul_8326_45_5_33", "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_2_gma", - "/transformer_blocks.16/norm2_context/Constant_output_0", - "/transformer_blocks.16/norm2_context/Constant_1_output_0", + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67", + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67", "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_3_existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_4", - "onnx::MatMul_8329", - "onnx::MatMul_8330", + "onnx::MatMul_8329_45_1_66", + "onnx::MatMul_8330_45_1_68", "existing_model.transformer_blocks.16.norm1_context.linear.weight_5_1_37_27_33_5_gma", - "/transformer_blocks.17/norm1_context/norm/Constant_output_0", - "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.17/norm1/norm/Constant_output_0", - "/transformer_blocks.17/norm1/norm/Constant_1_output_0", + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_68", + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_68", + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_69", + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_69", "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_0_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_1", "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_0_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_1", "onnx::MatMul_8334_onnx::MatMul_8331", "onnx::MatMul_8335_onnx::MatMul_8332", "onnx::MatMul_8336_onnx::MatMul_8333", - "onnx::MatMul_8350", + "onnx::MatMul_8350_45_5_34", "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_2_gma", - "/transformer_blocks.17/norm2/Constant_output_0", - "/transformer_blocks.17/norm2/Constant_1_output_0", + "/transformer_blocks.17/norm2/Constant_output_0_14_70", + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70", "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_3_existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_4", - "onnx::MatMul_8352", - "onnx::MatMul_8353", + "onnx::MatMul_8352_45_1_69", + "onnx::MatMul_8353_45_1_71", "existing_model.transformer_blocks.17.norm1.linear.weight_5_1_38_27_34_5_gma", - "onnx::MatMul_8351", + "onnx::MatMul_8351_45_5_35", "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_2_gma", - "/transformer_blocks.17/norm2_context/Constant_output_0", - "/transformer_blocks.17/norm2_context/Constant_1_output_0", + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71", + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71", "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_3_existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_4", - "onnx::MatMul_8354", - "onnx::MatMul_8355", + "onnx::MatMul_8354_45_1_70", + "onnx::MatMul_8355_45_1_72", "existing_model.transformer_blocks.17.norm1_context.linear.weight_5_1_39_27_35_5_gma", - "/transformer_blocks.18/norm1/norm/Constant_output_0", - "/transformer_blocks.18/norm1/norm/Constant_1_output_0", + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_73", + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_73", "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_0_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_1", - "/transformer_blocks.18/norm1_context/norm/Constant_output_0", - "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_72", + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_72", "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_0_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_1", "onnx::MatMul_8359_onnx::MatMul_8356", "onnx::MatMul_8360_onnx::MatMul_8357", "onnx::MatMul_8361_onnx::MatMul_8358", - "onnx::MatMul_8375", + "onnx::MatMul_8375_45_5_36", "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_2_gma", - "/transformer_blocks.18/norm2/Constant_output_0", - "/transformer_blocks.18/norm2/Constant_1_output_0", + "/transformer_blocks.18/norm2/Constant_output_0_14_74", + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74", "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_3_existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_4", - "onnx::MatMul_8377", - "onnx::MatMul_8378", + "onnx::MatMul_8377_45_1_73", + "onnx::MatMul_8378_45_1_75", "existing_model.transformer_blocks.18.norm1.linear.weight_5_1_40_27_36_5_gma", - "onnx::MatMul_8376", + "onnx::MatMul_8376_45_5_37", "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_2_gma", - "/transformer_blocks.18/norm2_context/Constant_output_0", - "/transformer_blocks.18/norm2_context/Constant_1_output_0", + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75", + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75", "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_3_existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_4", - "onnx::MatMul_8379", - "onnx::MatMul_8380", + "onnx::MatMul_8379_45_1_74", + "onnx::MatMul_8380_45_1_76", "existing_model.transformer_blocks.18.norm1_context.linear.weight_5_1_41_27_37_5_gma", - "/transformer_blocks.19/norm1_context/norm/Constant_output_0", - "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.19/norm1/norm/Constant_output_0", - "/transformer_blocks.19/norm1/norm/Constant_1_output_0", + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_76", + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_76", + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_77", + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_77", "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_0_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_1", "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_0_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_1", "onnx::MatMul_8384_onnx::MatMul_8381", "onnx::MatMul_8385_onnx::MatMul_8382", "onnx::MatMul_8386_onnx::MatMul_8383", - "onnx::MatMul_8400", + "onnx::MatMul_8400_45_5_38", "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_2_gma", - "/transformer_blocks.19/norm2/Constant_output_0", - "/transformer_blocks.19/norm2/Constant_1_output_0", + "/transformer_blocks.19/norm2/Constant_output_0_14_78", + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78", "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_3_existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_4", - "onnx::MatMul_8402", - "onnx::MatMul_8403", + "onnx::MatMul_8402_45_1_77", + "onnx::MatMul_8403_45_1_79", "existing_model.transformer_blocks.19.norm1.linear.weight_5_1_42_27_38_5_gma", - "onnx::MatMul_8401", + "onnx::MatMul_8401_45_5_39", "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_2_gma", - "/transformer_blocks.19/norm2_context/Constant_output_0", - "/transformer_blocks.19/norm2_context/Constant_1_output_0", + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79", + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79", "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_3_existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_4", - "onnx::MatMul_8404", - "onnx::MatMul_8405", + "onnx::MatMul_8404_45_1_78", + "onnx::MatMul_8405_45_1_80", "existing_model.transformer_blocks.19.norm1_context.linear.weight_5_1_43_27_39_5_gma", - "/transformer_blocks.20/norm1/norm/Constant_output_0", - "/transformer_blocks.20/norm1/norm/Constant_1_output_0", - "/transformer_blocks.20/norm1_context/norm/Constant_output_0", - "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_81", + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_81", + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_80", + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_80", "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_0_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_1", "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_0_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_1", "onnx::MatMul_8409_onnx::MatMul_8406", "onnx::MatMul_8410_onnx::MatMul_8407", "onnx::MatMul_8411_onnx::MatMul_8408", - "onnx::MatMul_8426", - "onnx::MatMul_8425", + "onnx::MatMul_8426_45_5_41", + "onnx::MatMul_8425_45_5_40", "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_2_gma", - "/transformer_blocks.20/norm2/Constant_output_0", - "/transformer_blocks.20/norm2/Constant_1_output_0", + "/transformer_blocks.20/norm2/Constant_output_0_14_82", + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82", "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_3_existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_4", - "onnx::MatMul_8427", - "onnx::MatMul_8428", + "onnx::MatMul_8427_45_1_81", + "onnx::MatMul_8428_45_1_83", "existing_model.transformer_blocks.20.norm1.linear.weight_5_1_44_27_40_5_gma", "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_2_gma", - "/transformer_blocks.20/norm2_context/Constant_output_0", - "/transformer_blocks.20/norm2_context/Constant_1_output_0", + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83", + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83", "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_3_existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_4", - "onnx::MatMul_8429", - "onnx::MatMul_8430", + "onnx::MatMul_8429_45_1_82", + "onnx::MatMul_8430_45_1_84", "existing_model.transformer_blocks.20.norm1_context.linear.weight_5_1_45_27_41_5_gma", - "/transformer_blocks.21/norm1_context/norm/Constant_output_0", - "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.21/norm1/norm/Constant_output_0", - "/transformer_blocks.21/norm1/norm/Constant_1_output_0", + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_84", + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_84", + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_85", + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_85", "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_0_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_1", "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_0_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_1", "onnx::MatMul_8434_onnx::MatMul_8431", "onnx::MatMul_8435_onnx::MatMul_8432", "onnx::MatMul_8436_onnx::MatMul_8433", - "onnx::MatMul_8450", + "onnx::MatMul_8450_45_5_42", "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_2_gma", - "/transformer_blocks.21/norm2/Constant_output_0", - "/transformer_blocks.21/norm2/Constant_1_output_0", + "/transformer_blocks.21/norm2/Constant_output_0_14_86", + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86", "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_3_existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_4", - "onnx::MatMul_8452", - "onnx::MatMul_8453", + "onnx::MatMul_8452_45_1_85", + "onnx::MatMul_8453_45_1_87", "existing_model.transformer_blocks.21.norm1.linear.weight_5_1_46_27_42_5_gma", - "onnx::MatMul_8451", + "onnx::MatMul_8451_45_5_43", "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_2_gma", - "/transformer_blocks.21/norm2_context/Constant_output_0", - "/transformer_blocks.21/norm2_context/Constant_1_output_0", + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87", + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87", "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_3_existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_4", - "onnx::MatMul_8454", - "onnx::MatMul_8455", + "onnx::MatMul_8454_45_1_86", + "onnx::MatMul_8455_45_1_88", "existing_model.transformer_blocks.21.norm1_context.linear.weight_5_1_47_27_43_5_gma", - "/transformer_blocks.22/norm1/norm/Constant_output_0", - "/transformer_blocks.22/norm1/norm/Constant_1_output_0", + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_89", + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_89", "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_0_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_1", - "/transformer_blocks.22/norm1_context/norm/Constant_output_0", - "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0", + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_88", + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_88", "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_0_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_1", "onnx::MatMul_8459_onnx::MatMul_8456", "onnx::MatMul_8460_onnx::MatMul_8457", "onnx::MatMul_8461_onnx::MatMul_8458", - "onnx::MatMul_8475", + "onnx::MatMul_8475_45_5_44", "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_2_gma", - "/transformer_blocks.22/norm2/Constant_output_0", - "/transformer_blocks.22/norm2/Constant_1_output_0", + "/transformer_blocks.22/norm2/Constant_output_0_14_90", + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90", "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_3_existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_4", - "onnx::MatMul_8477", - "onnx::MatMul_8478", + "onnx::MatMul_8477_45_1_89", + "onnx::MatMul_8478_45_1_91", "existing_model.transformer_blocks.22.norm1.linear.weight_5_1_48_27_44_5_gma", - "onnx::MatMul_8476", + "onnx::MatMul_8476_45_5_45", "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_2_gma", - "/transformer_blocks.22/norm2_context/Constant_output_0", - "/transformer_blocks.22/norm2_context/Constant_1_output_0", + "/transformer_blocks.22/norm2_context/Constant_output_0_14_91", + "/transformer_blocks.22/norm2_context/Constant_1_output_0_14_91", "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_3_existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_4", - "onnx::MatMul_8479", - "onnx::MatMul_8480", + "onnx::MatMul_8479_45_1_90", + "onnx::MatMul_8480_45_1_92", "existing_model.transformer_blocks.22.norm1_context.linear.weight_5_1_49_27_45_5_gma", - "/transformer_blocks.23/norm1_context/norm/Constant_output_0", - "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0", - "/transformer_blocks.23/norm1/norm/Constant_output_0", - "/transformer_blocks.23/norm1/norm/Constant_1_output_0", + "/transformer_blocks.23/norm1_context/norm/Constant_output_0_14_92", + "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0_14_92", + "/transformer_blocks.23/norm1/norm/Constant_output_0_14_93", + "/transformer_blocks.23/norm1/norm/Constant_1_output_0_14_93", "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_0_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_1", "existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_0_existing_model.transformer_blocks.23.norm1_context.linear.weight_5_1_51_27_47_1", "onnx::MatMul_8484_onnx::MatMul_8481", "onnx::MatMul_8485_onnx::MatMul_8482", "onnx::MatMul_8486_onnx::MatMul_8483", - "onnx::MatMul_8497", + "onnx::MatMul_8497_45_5_46", "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_2_gma", - "/transformer_blocks.23/norm2/Constant_output_0", - "/transformer_blocks.23/norm2/Constant_1_output_0", + "/transformer_blocks.23/norm2/Constant_output_0_14_94", + "/transformer_blocks.23/norm2/Constant_1_output_0_14_94", "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_3_existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_4", - "onnx::MatMul_8498", - "onnx::MatMul_8499", + "onnx::MatMul_8498_45_1_93", + "onnx::MatMul_8499_45_1_94", "existing_model.transformer_blocks.23.norm1.linear.weight_5_1_50_27_46_5_gma", - "/norm_out/norm/Constant_output_0", - "/norm_out/norm/Constant_1_output_0", + "/norm_out/norm/Constant_output_0_14_95", + "/norm_out/norm/Constant_1_output_0_14_95", "existing_model.norm_out.linear.weight_5_1_52_27_48_0_existing_model.norm_out.linear.weight_5_1_52_27_48_1", "/norm_out/Add_2_output_0.out0_0_95_bfp.out1_93_bfp.wts", - "onnx::MatMul_8500" + "onnx::MatMul_8500_45_0_4" ] }, "super_instr": { @@ -46909,7 +46909,7 @@ "size_in_bytes": 32, "op_tensor_size": 32, "dynamic_shapes": [ - "batch_size", + "batch_4", "w", "h", "False" @@ -46934,7 +46934,7 @@ ], "offset": 32 }, - "/time_text_embed/Cast_output_0.out17_3_3": { + "/time_text_embed/Cast_output_0.out17_4_3": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -46945,12 +46945,12 @@ "size_in_bytes": 512, "op_tensor_size": 512, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 3104 }, - "pooled_projections.out17_3_1": { + "pooled_projections.out17_4_1": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -46961,12 +46961,12 @@ "size_in_bytes": 4096, "op_tensor_size": 4096, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 3616 }, - "encoder_hidden_states.out17_3_0": { + "encoder_hidden_states.out17_4_0": { "packed_buffer_label": "in", "xrt_arg_id": 0, "dtype": "bfloat16", @@ -46978,7 +46978,7 @@ "size_in_bytes": 8192, "op_tensor_size": 8192, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -46996,11 +46996,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 18976 + "offset": 15904 }, "block_controlnet_hidden_states_1.out_35_1_6": { "packed_buffer_label": "in", @@ -47014,11 +47014,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 25120 + "offset": 18976 }, "block_controlnet_hidden_states_2.out_35_1_8": { "packed_buffer_label": "in", @@ -47032,11 +47032,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 31264 + "offset": 22048 }, "block_controlnet_hidden_states_3.out_35_1_10": { "packed_buffer_label": "in", @@ -47050,11 +47050,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 37408 + "offset": 25120 }, "block_controlnet_hidden_states_4.out_35_1_12": { "packed_buffer_label": "in", @@ -47068,11 +47068,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 43552 + "offset": 28192 }, "block_controlnet_hidden_states_5.out_35_1_14": { "packed_buffer_label": "in", @@ -47086,11 +47086,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 49696 + "offset": 31264 }, "block_controlnet_hidden_states_6.out_35_1_16": { "packed_buffer_label": "in", @@ -47104,11 +47104,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 55840 + "offset": 34336 }, "block_controlnet_hidden_states_7.out_35_1_18": { "packed_buffer_label": "in", @@ -47122,11 +47122,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 61984 + "offset": 37408 }, "block_controlnet_hidden_states_8.out_35_1_20": { "packed_buffer_label": "in", @@ -47140,11 +47140,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 68128 + "offset": 40480 }, "block_controlnet_hidden_states_9.out_35_1_22": { "packed_buffer_label": "in", @@ -47158,11 +47158,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 74272 + "offset": 43552 }, "block_controlnet_hidden_states_10.out_35_1_24": { "packed_buffer_label": "in", @@ -47176,11 +47176,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 80416 + "offset": 46624 }, "block_controlnet_hidden_states_11.out_35_1_26": { "packed_buffer_label": "in", @@ -47194,11 +47194,11 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], - "offset": 83488 + "offset": 49696 }, "/Reshape_output_0.out17_0_0": { "packed_buffer_label": "out", @@ -47215,7 +47215,7 @@ "size_in_bytes": 128, "op_tensor_size": 128, "dynamic_shapes": [ - "batch_size", + "batch_4", "floor(w/2)", "floor(h/2)", "False", @@ -47236,7 +47236,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47254,13 +47254,13 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 3072 }, - "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_3_3": { + "/time_text_embed/timestep_embedder/linear_1/Gemm_output_0.out17_4_3": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -47271,7 +47271,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 6144 @@ -47287,12 +47287,12 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 9216 }, - "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_3_4": { + "/time_text_embed/timestep_embedder/linear_2/Gemm_output_0.out17_4_4": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -47303,12 +47303,12 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 12288 }, - "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_3_1": { + "/time_text_embed/text_embedder/linear_1/Gemm_output_0.out17_4_1": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -47319,7 +47319,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 15360 @@ -47335,12 +47335,12 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 18432 }, - "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_3_2": { + "/time_text_embed/text_embedder/linear_2/Gemm_output_0.out17_4_2": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -47351,7 +47351,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 21504 @@ -47367,7 +47367,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 24576 @@ -47383,12 +47383,12 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "False" ], "offset": 27648 }, - "encoder_hidden_states.out17_3_0_bfp.out25_0": { + "encoder_hidden_states.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -47400,13 +47400,13 @@ "size_in_bytes": 4096, "op_tensor_size": 4096, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 30720 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -47418,13 +47418,13 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 34816 }, - "/context_embedder/Add_output_0.out17_3_0": { + "/context_embedder/Add_output_0.out17_4_0": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "bfloat16", @@ -47436,7 +47436,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47454,7 +47454,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47472,7 +47472,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47490,7 +47490,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47508,7 +47508,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47527,7 +47527,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -47547,7 +47547,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -47567,7 +47567,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -47586,7 +47586,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -47604,7 +47604,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47622,7 +47622,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47640,7 +47640,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47658,7 +47658,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47676,13 +47676,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 59392 }, - "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_3_12_bfp.out25_8": { + "/transformer_blocks.0/ff_context/net.2/Add_output_0.out17_4_12_bfp.out25_8": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -47694,7 +47694,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47712,7 +47712,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47730,7 +47730,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47748,7 +47748,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47766,7 +47766,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47784,7 +47784,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47802,13 +47802,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 77824 }, - "/transformer_blocks.0/ff/net.2/Add_output_0.out17_3_11_bfp.out25_7": { + "/transformer_blocks.0/ff/net.2/Add_output_0.out17_4_11_bfp.out25_7": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -47820,7 +47820,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47838,7 +47838,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47856,7 +47856,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47874,7 +47874,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47892,7 +47892,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47910,7 +47910,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -47928,7 +47928,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -47947,7 +47947,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -47967,7 +47967,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -47987,7 +47987,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -48006,7 +48006,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -48024,7 +48024,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48042,7 +48042,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48060,7 +48060,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48078,7 +48078,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48096,13 +48096,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 111616 }, - "/transformer_blocks.1/ff/net.2/Add_output_0.out17_3_19_bfp.out25_11": { + "/transformer_blocks.1/ff/net.2/Add_output_0.out17_4_19_bfp.out25_11": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -48114,7 +48114,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48132,7 +48132,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48150,7 +48150,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48168,7 +48168,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48186,7 +48186,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48204,7 +48204,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48222,7 +48222,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48240,13 +48240,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 133120 }, - "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_3_20_bfp.out25_12": { + "/transformer_blocks.1/ff_context/net.2/Add_output_0.out17_4_20_bfp.out25_12": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -48258,7 +48258,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48276,7 +48276,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48294,7 +48294,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48312,7 +48312,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48330,7 +48330,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48348,7 +48348,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48367,7 +48367,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -48387,7 +48387,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -48407,7 +48407,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -48426,7 +48426,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -48444,7 +48444,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48462,7 +48462,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48480,7 +48480,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48498,7 +48498,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48516,7 +48516,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48534,7 +48534,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48552,13 +48552,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 168448 }, - "/transformer_blocks.2/ff/net.2/Add_output_0.out17_3_27_bfp.out25_15": { + "/transformer_blocks.2/ff/net.2/Add_output_0.out17_4_27_bfp.out25_15": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -48570,7 +48570,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48588,7 +48588,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48606,7 +48606,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48624,13 +48624,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 179200 }, - "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_3_28_bfp.out25_16": { + "/transformer_blocks.2/ff_context/net.2/Add_output_0.out17_4_28_bfp.out25_16": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -48642,7 +48642,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48660,7 +48660,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48678,7 +48678,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48696,7 +48696,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48714,7 +48714,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48732,7 +48732,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48750,7 +48750,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48768,7 +48768,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48787,7 +48787,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -48807,7 +48807,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -48827,7 +48827,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -48846,7 +48846,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -48864,7 +48864,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48882,7 +48882,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -48900,7 +48900,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48918,7 +48918,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48936,7 +48936,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48954,7 +48954,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -48972,13 +48972,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 220672 }, - "/transformer_blocks.3/ff/net.2/Add_output_0.out17_3_35_bfp.out25_19": { + "/transformer_blocks.3/ff/net.2/Add_output_0.out17_4_35_bfp.out25_19": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -48990,7 +48990,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49008,7 +49008,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49026,7 +49026,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49044,7 +49044,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49062,7 +49062,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49080,13 +49080,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 237568 }, - "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_3_36_bfp.out25_20": { + "/transformer_blocks.3/ff_context/net.2/Add_output_0.out17_4_36_bfp.out25_20": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -49098,7 +49098,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49116,7 +49116,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49134,7 +49134,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49152,7 +49152,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49170,7 +49170,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49188,7 +49188,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49207,7 +49207,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -49227,7 +49227,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -49247,7 +49247,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -49266,7 +49266,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -49284,7 +49284,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49302,7 +49302,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49320,7 +49320,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49338,7 +49338,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49356,7 +49356,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49374,7 +49374,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49392,13 +49392,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 272896 }, - "/transformer_blocks.4/ff/net.2/Add_output_0.out17_3_43_bfp.out25_23": { + "/transformer_blocks.4/ff/net.2/Add_output_0.out17_4_43_bfp.out25_23": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -49410,7 +49410,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49428,7 +49428,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49446,7 +49446,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49464,13 +49464,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 283648 }, - "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_3_44_bfp.out25_24": { + "/transformer_blocks.4/ff_context/net.2/Add_output_0.out17_4_44_bfp.out25_24": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -49482,7 +49482,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49500,7 +49500,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49518,7 +49518,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49536,7 +49536,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49554,7 +49554,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49572,7 +49572,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49590,7 +49590,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49608,7 +49608,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49627,7 +49627,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -49647,7 +49647,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -49667,7 +49667,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -49686,7 +49686,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -49704,7 +49704,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49722,7 +49722,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49740,7 +49740,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49758,7 +49758,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49776,7 +49776,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49794,7 +49794,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49812,13 +49812,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 325120 }, - "/transformer_blocks.5/ff/net.2/Add_output_0.out17_3_51_bfp.out25_27": { + "/transformer_blocks.5/ff/net.2/Add_output_0.out17_4_51_bfp.out25_27": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -49830,7 +49830,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49848,7 +49848,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49866,7 +49866,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49884,7 +49884,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49902,7 +49902,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49920,13 +49920,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 342016 }, - "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_3_52_bfp.out25_28": { + "/transformer_blocks.5/ff_context/net.2/Add_output_0.out17_4_52_bfp.out25_28": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -49938,7 +49938,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49956,7 +49956,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -49974,7 +49974,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -49992,7 +49992,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50010,7 +50010,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50028,7 +50028,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50047,7 +50047,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50067,7 +50067,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50087,7 +50087,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50106,7 +50106,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -50124,7 +50124,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50142,7 +50142,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50160,7 +50160,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50178,7 +50178,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50196,7 +50196,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50214,7 +50214,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50232,13 +50232,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 377344 }, - "/transformer_blocks.6/ff/net.2/Add_output_0.out17_3_59_bfp.out25_31": { + "/transformer_blocks.6/ff/net.2/Add_output_0.out17_4_59_bfp.out25_31": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -50250,7 +50250,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50268,7 +50268,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50286,7 +50286,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50304,13 +50304,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 388096 }, - "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_3_60_bfp.out25_32": { + "/transformer_blocks.6/ff_context/net.2/Add_output_0.out17_4_60_bfp.out25_32": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -50322,7 +50322,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50340,7 +50340,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50358,7 +50358,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50376,7 +50376,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50394,7 +50394,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50412,7 +50412,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50430,7 +50430,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50448,7 +50448,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50467,7 +50467,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50487,7 +50487,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50507,7 +50507,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50526,7 +50526,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -50544,7 +50544,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50562,7 +50562,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50580,7 +50580,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50598,7 +50598,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50616,7 +50616,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50634,7 +50634,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50652,13 +50652,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 429568 }, - "/transformer_blocks.7/ff/net.2/Add_output_0.out17_3_67_bfp.out25_35": { + "/transformer_blocks.7/ff/net.2/Add_output_0.out17_4_67_bfp.out25_35": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -50670,7 +50670,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50688,7 +50688,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50706,7 +50706,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50724,7 +50724,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50742,7 +50742,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50760,13 +50760,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 446464 }, - "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_3_68_bfp.out25_36": { + "/transformer_blocks.7/ff_context/net.2/Add_output_0.out17_4_68_bfp.out25_36": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -50778,7 +50778,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50796,7 +50796,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50814,7 +50814,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50832,7 +50832,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50850,7 +50850,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -50868,7 +50868,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50887,7 +50887,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50907,7 +50907,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50927,7 +50927,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -50946,7 +50946,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -50964,7 +50964,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -50982,7 +50982,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51000,7 +51000,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51018,7 +51018,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51036,7 +51036,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51054,7 +51054,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51072,13 +51072,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 481792 }, - "/transformer_blocks.8/ff/net.2/Add_output_0.out17_3_75_bfp.out25_39": { + "/transformer_blocks.8/ff/net.2/Add_output_0.out17_4_75_bfp.out25_39": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -51090,7 +51090,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51108,7 +51108,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51126,7 +51126,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51144,13 +51144,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 492544 }, - "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_3_76_bfp.out25_40": { + "/transformer_blocks.8/ff_context/net.2/Add_output_0.out17_4_76_bfp.out25_40": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -51162,7 +51162,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51180,7 +51180,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51198,7 +51198,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51216,7 +51216,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51234,7 +51234,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51252,7 +51252,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51270,7 +51270,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51288,7 +51288,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51307,7 +51307,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -51327,7 +51327,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -51347,7 +51347,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -51366,7 +51366,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -51384,7 +51384,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51402,7 +51402,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51420,7 +51420,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51438,7 +51438,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51456,7 +51456,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51474,7 +51474,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51492,13 +51492,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 534016 }, - "/transformer_blocks.9/ff/net.2/Add_output_0.out17_3_83_bfp.out25_43": { + "/transformer_blocks.9/ff/net.2/Add_output_0.out17_4_83_bfp.out25_43": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -51510,7 +51510,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51528,7 +51528,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51546,7 +51546,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51564,7 +51564,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51582,7 +51582,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51600,7 +51600,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51618,7 +51618,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51636,13 +51636,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 553984 }, - "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_3_84_bfp.out25_44": { + "/transformer_blocks.9/ff_context/net.2/Add_output_0.out17_4_84_bfp.out25_44": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -51654,7 +51654,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51672,7 +51672,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51690,7 +51690,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51708,7 +51708,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51727,7 +51727,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -51747,7 +51747,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -51767,7 +51767,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -51786,7 +51786,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -51804,7 +51804,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51822,7 +51822,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51840,7 +51840,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51858,7 +51858,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51876,13 +51876,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 581632 }, - "/transformer_blocks.10/ff/net.2/Add_output_0.out17_3_91_bfp.out25_47": { + "/transformer_blocks.10/ff/net.2/Add_output_0.out17_4_91_bfp.out25_47": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -51894,7 +51894,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51912,7 +51912,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -51930,7 +51930,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51948,7 +51948,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51966,7 +51966,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -51984,7 +51984,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52002,13 +52002,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 600064 }, - "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_3_92_bfp.out25_48": { + "/transformer_blocks.10/ff_context/net.2/Add_output_0.out17_4_92_bfp.out25_48": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -52020,7 +52020,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52038,7 +52038,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52056,7 +52056,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52074,7 +52074,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52092,7 +52092,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52110,7 +52110,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52128,7 +52128,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52147,7 +52147,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -52167,7 +52167,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -52187,7 +52187,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -52206,7 +52206,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -52224,7 +52224,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52242,7 +52242,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52260,7 +52260,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52278,7 +52278,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52296,13 +52296,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 633856 }, - "/transformer_blocks.11/ff/net.2/Add_output_0.out17_3_99_bfp.out25_51": { + "/transformer_blocks.11/ff/net.2/Add_output_0.out17_4_99_bfp.out25_51": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -52314,7 +52314,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52332,7 +52332,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52350,7 +52350,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52368,7 +52368,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52386,7 +52386,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52404,7 +52404,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52422,7 +52422,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52440,13 +52440,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 655360 }, - "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_3_100_bfp.out25_52": { + "/transformer_blocks.11/ff_context/net.2/Add_output_0.out17_4_100_bfp.out25_52": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -52458,7 +52458,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52476,7 +52476,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52494,7 +52494,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52512,7 +52512,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52530,7 +52530,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52548,7 +52548,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52567,7 +52567,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -52587,7 +52587,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -52607,7 +52607,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -52626,7 +52626,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -52644,7 +52644,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52662,7 +52662,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52680,7 +52680,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52698,7 +52698,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52716,13 +52716,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 686080 }, - "/transformer_blocks.12/ff/net.2/Add_output_0.out17_3_107_bfp.out25_55": { + "/transformer_blocks.12/ff/net.2/Add_output_0.out17_4_107_bfp.out25_55": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -52734,7 +52734,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52752,7 +52752,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52770,7 +52770,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52788,7 +52788,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52806,7 +52806,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52824,7 +52824,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52842,13 +52842,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 704512 }, - "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_3_108_bfp.out25_56": { + "/transformer_blocks.12/ff_context/net.2/Add_output_0.out17_4_108_bfp.out25_56": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -52860,7 +52860,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52878,7 +52878,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52896,7 +52896,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52914,7 +52914,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52932,7 +52932,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52950,7 +52950,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -52968,7 +52968,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -52987,7 +52987,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53007,7 +53007,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53027,7 +53027,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53046,7 +53046,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -53064,7 +53064,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53082,7 +53082,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53100,7 +53100,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53118,7 +53118,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53136,13 +53136,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 738304 }, - "/transformer_blocks.13/ff/net.2/Add_output_0.out17_3_115_bfp.out25_59": { + "/transformer_blocks.13/ff/net.2/Add_output_0.out17_4_115_bfp.out25_59": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -53154,7 +53154,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53172,7 +53172,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53190,7 +53190,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53208,7 +53208,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53226,7 +53226,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53244,7 +53244,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53262,7 +53262,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53280,13 +53280,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 759808 }, - "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_3_116_bfp.out25_60": { + "/transformer_blocks.13/ff_context/net.2/Add_output_0.out17_4_116_bfp.out25_60": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -53298,7 +53298,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53316,7 +53316,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53334,7 +53334,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53352,7 +53352,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53370,7 +53370,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53388,7 +53388,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53407,7 +53407,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53427,7 +53427,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53447,7 +53447,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53466,7 +53466,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -53484,7 +53484,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53502,7 +53502,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53520,7 +53520,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53538,7 +53538,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53556,13 +53556,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 790528 }, - "/transformer_blocks.14/ff/net.2/Add_output_0.out17_3_123_bfp.out25_63": { + "/transformer_blocks.14/ff/net.2/Add_output_0.out17_4_123_bfp.out25_63": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -53574,7 +53574,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53592,7 +53592,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53610,7 +53610,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53628,7 +53628,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53646,7 +53646,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53664,7 +53664,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53682,13 +53682,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 808960 }, - "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_3_124_bfp.out25_64": { + "/transformer_blocks.14/ff_context/net.2/Add_output_0.out17_4_124_bfp.out25_64": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -53700,7 +53700,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53718,7 +53718,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53736,7 +53736,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53754,7 +53754,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53772,7 +53772,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53790,7 +53790,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53808,7 +53808,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -53827,7 +53827,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53847,7 +53847,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53867,7 +53867,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -53886,7 +53886,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -53904,7 +53904,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53922,7 +53922,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53940,7 +53940,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53958,7 +53958,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -53976,13 +53976,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 842752 }, - "/transformer_blocks.15/ff/net.2/Add_output_0.out17_3_131_bfp.out25_67": { + "/transformer_blocks.15/ff/net.2/Add_output_0.out17_4_131_bfp.out25_67": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -53994,7 +53994,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54012,7 +54012,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54030,7 +54030,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54048,7 +54048,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54066,7 +54066,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54084,7 +54084,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54102,7 +54102,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54120,13 +54120,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 864256 }, - "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_3_132_bfp.out25_68": { + "/transformer_blocks.15/ff_context/net.2/Add_output_0.out17_4_132_bfp.out25_68": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -54138,7 +54138,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54156,7 +54156,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54174,7 +54174,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54192,7 +54192,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54210,7 +54210,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54228,7 +54228,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54247,7 +54247,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -54267,7 +54267,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -54287,7 +54287,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -54306,7 +54306,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -54324,7 +54324,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54342,7 +54342,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54360,7 +54360,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54378,7 +54378,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54396,13 +54396,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 894976 }, - "/transformer_blocks.16/ff/net.2/Add_output_0.out17_3_139_bfp.out25_71": { + "/transformer_blocks.16/ff/net.2/Add_output_0.out17_4_139_bfp.out25_71": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -54414,7 +54414,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54432,7 +54432,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54450,7 +54450,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54468,7 +54468,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54486,7 +54486,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54504,7 +54504,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54522,13 +54522,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 913408 }, - "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_3_140_bfp.out25_72": { + "/transformer_blocks.16/ff_context/net.2/Add_output_0.out17_4_140_bfp.out25_72": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -54540,7 +54540,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54558,7 +54558,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54576,7 +54576,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54594,7 +54594,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54612,7 +54612,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54630,7 +54630,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54648,7 +54648,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54667,7 +54667,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -54687,7 +54687,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -54707,7 +54707,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -54726,7 +54726,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -54744,7 +54744,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54762,7 +54762,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54780,7 +54780,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54798,7 +54798,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54816,13 +54816,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 947200 }, - "/transformer_blocks.17/ff/net.2/Add_output_0.out17_3_147_bfp.out25_75": { + "/transformer_blocks.17/ff/net.2/Add_output_0.out17_4_147_bfp.out25_75": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -54834,7 +54834,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54852,7 +54852,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54870,7 +54870,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -54888,7 +54888,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54906,7 +54906,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54924,7 +54924,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54942,7 +54942,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54960,13 +54960,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 968704 }, - "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_3_148_bfp.out25_76": { + "/transformer_blocks.17/ff_context/net.2/Add_output_0.out17_4_148_bfp.out25_76": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -54978,7 +54978,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -54996,7 +54996,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55014,7 +55014,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55032,7 +55032,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55050,7 +55050,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55068,7 +55068,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55087,7 +55087,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55107,7 +55107,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55127,7 +55127,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55146,7 +55146,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -55164,7 +55164,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55182,7 +55182,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55200,7 +55200,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55218,7 +55218,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55236,13 +55236,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 999424 }, - "/transformer_blocks.18/ff/net.2/Add_output_0.out17_3_155_bfp.out25_79": { + "/transformer_blocks.18/ff/net.2/Add_output_0.out17_4_155_bfp.out25_79": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -55254,7 +55254,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55272,7 +55272,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55290,7 +55290,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55308,7 +55308,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55326,7 +55326,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55344,7 +55344,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55362,13 +55362,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 1017856 }, - "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_3_156_bfp.out25_80": { + "/transformer_blocks.18/ff_context/net.2/Add_output_0.out17_4_156_bfp.out25_80": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -55380,7 +55380,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55398,7 +55398,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55416,7 +55416,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55434,7 +55434,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55452,7 +55452,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55470,7 +55470,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55488,7 +55488,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55507,7 +55507,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55527,7 +55527,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55547,7 +55547,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55566,7 +55566,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -55584,7 +55584,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55602,7 +55602,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55620,7 +55620,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55638,7 +55638,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55656,13 +55656,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 1051648 }, - "/transformer_blocks.19/ff/net.2/Add_output_0.out17_3_163_bfp.out25_83": { + "/transformer_blocks.19/ff/net.2/Add_output_0.out17_4_163_bfp.out25_83": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -55674,7 +55674,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55692,7 +55692,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55710,7 +55710,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55728,7 +55728,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55746,7 +55746,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55764,7 +55764,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55782,7 +55782,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55800,13 +55800,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 1073152 }, - "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_3_164_bfp.out25_84": { + "/transformer_blocks.19/ff_context/net.2/Add_output_0.out17_4_164_bfp.out25_84": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -55818,7 +55818,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55836,7 +55836,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55854,7 +55854,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55872,7 +55872,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55890,7 +55890,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -55908,7 +55908,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -55927,7 +55927,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55947,7 +55947,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55967,7 +55967,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -55986,7 +55986,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -56004,7 +56004,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56022,7 +56022,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56040,7 +56040,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56058,7 +56058,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56076,7 +56076,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56094,13 +56094,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 1105408 }, - "/transformer_blocks.20/ff/net.2/Add_output_0.out17_3_171_bfp.out25_87": { + "/transformer_blocks.20/ff/net.2/Add_output_0.out17_4_171_bfp.out25_87": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -56112,7 +56112,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56130,7 +56130,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56148,7 +56148,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56166,7 +56166,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56184,7 +56184,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56202,13 +56202,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 1122304 }, - "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_3_172_bfp.out25_88": { + "/transformer_blocks.20/ff_context/net.2/Add_output_0.out17_4_172_bfp.out25_88": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -56220,7 +56220,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56238,7 +56238,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56256,7 +56256,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56274,7 +56274,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56292,7 +56292,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56310,7 +56310,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56328,7 +56328,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56347,7 +56347,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -56367,7 +56367,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -56387,7 +56387,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -56406,7 +56406,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -56424,7 +56424,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56442,7 +56442,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56460,7 +56460,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56478,7 +56478,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56496,13 +56496,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 1156096 }, - "/transformer_blocks.21/ff/net.2/Add_output_0.out17_3_179_bfp.out25_91": { + "/transformer_blocks.21/ff/net.2/Add_output_0.out17_4_179_bfp.out25_91": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -56514,7 +56514,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56532,7 +56532,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56550,7 +56550,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56568,7 +56568,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56586,7 +56586,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56604,7 +56604,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56622,7 +56622,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56640,13 +56640,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 1177600 }, - "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_3_180_bfp.out25_92": { + "/transformer_blocks.21/ff_context/net.2/Add_output_0.out17_4_180_bfp.out25_92": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -56658,7 +56658,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56676,7 +56676,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56694,7 +56694,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56712,7 +56712,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56730,7 +56730,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56748,7 +56748,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56767,7 +56767,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -56787,7 +56787,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -56807,7 +56807,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -56826,7 +56826,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -56844,7 +56844,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56862,7 +56862,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56880,7 +56880,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56898,7 +56898,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56916,13 +56916,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 1208320 }, - "/transformer_blocks.22/ff/net.2/Add_output_0.out17_3_187_bfp.out25_95": { + "/transformer_blocks.22/ff/net.2/Add_output_0.out17_4_187_bfp.out25_95": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -56934,7 +56934,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56952,7 +56952,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -56970,7 +56970,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -56988,7 +56988,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -57006,7 +57006,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -57024,7 +57024,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -57042,13 +57042,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], "offset": 1226752 }, - "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_3_188_bfp.out25_96": { + "/transformer_blocks.22/ff_context/net.2/Add_output_0.out17_4_188_bfp.out25_96": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -57060,7 +57060,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -57078,7 +57078,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -57096,7 +57096,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -57114,7 +57114,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57132,7 +57132,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57150,7 +57150,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57168,7 +57168,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length", "False" ], @@ -57187,7 +57187,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -57207,7 +57207,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -57227,7 +57227,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "False", "max_length + state_dim1", "False" @@ -57246,7 +57246,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "max_length + state_dim1", "False" ], @@ -57264,7 +57264,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57282,7 +57282,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57300,7 +57300,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57318,7 +57318,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57336,13 +57336,13 @@ "size_in_bytes": 6144, "op_tensor_size": 6144, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], "offset": 1260544 }, - "/transformer_blocks.23/ff/net.2/Add_output_0.out17_3_195_bfp.out25_98": { + "/transformer_blocks.23/ff/net.2/Add_output_0.out17_4_195_bfp.out25_98": { "packed_buffer_label": "scratch", "xrt_arg_id": 2, "dtype": "uint8", @@ -57354,7 +57354,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57372,7 +57372,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57390,7 +57390,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57408,7 +57408,7 @@ "size_in_bytes": 1536, "op_tensor_size": 1536, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57426,7 +57426,7 @@ "size_in_bytes": 3072, "op_tensor_size": 3072, "dynamic_shapes": [ - "batch_size", + "batch_4", "state_dim1", "False" ], @@ -57445,7 +57445,7 @@ "file_name": "cache/pos_embedprojConv_0.const", "file_size": 135168 }, - "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0": { + "existing_model.time_text_embed.timestep_embedder.linear_1.weight_5_1_0_45_0_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57471,7 +57471,7 @@ "file_name": "cache/pos_embedprojConv_2.const", "file_size": 256 }, - "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1": { + "existing_model.time_text_embed.timestep_embedder.linear_2.weight_5_1_1_45_0_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57484,7 +57484,7 @@ "file_name": "cache/pos_embedprojConv_3.const", "file_size": 2691072 }, - "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2": { + "existing_model.time_text_embed.text_embedder.linear_1.weight_5_1_2_45_0_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57510,7 +57510,7 @@ "file_name": "cache/pos_embedprojConv_5.const", "file_size": 256 }, - "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3": { + "existing_model.time_text_embed.text_embedder.linear_2.weight_5_1_3_45_0_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57536,7 +57536,7 @@ "file_name": "cache/pos_embedprojConv_7.const", "file_size": 256 }, - "encoder_hidden_states.out17_3_0_bfp.wts": { + "encoder_hidden_states.out17_4_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57549,7 +57549,7 @@ "file_name": "cache/pos_embedprojConv_8.const", "file_size": 128 }, - "onnx::MatMul_7905": { + "onnx::MatMul_7905_45_1_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57562,7 +57562,7 @@ "file_name": "cache/pos_embedprojConv_9.const", "file_size": 7274496 }, - "/context_embedder/Add_output_0.out17_3_0_bfp.out25_0_bfp.wts": { + "/context_embedder/Add_output_0.out17_4_0_bfp.out25_0_bfp.wts": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57575,7 +57575,7 @@ "file_name": "cache/pos_embedprojConv_10.const", "file_size": 128 }, - "/transformer_blocks.0/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57588,7 +57588,7 @@ "file_name": "cache/pos_embedprojConv_11.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1_context/norm/Constant_1_output_0_14_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57614,7 +57614,7 @@ "file_name": "cache/pos_embedprojConv_13.const", "file_size": 5382144 }, - "/transformer_blocks.0/norm1/norm/Constant_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57627,7 +57627,7 @@ "file_name": "cache/pos_embedprojConv_14.const", "file_size": 3072 }, - "/transformer_blocks.0/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.0/norm1/norm/Constant_1_output_0_14_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57692,7 +57692,7 @@ "file_name": "cache/pos_embedprojConv_19.const", "file_size": 5382144 }, - "onnx::MatMul_7926": { + "onnx::MatMul_7926_45_5_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57718,7 +57718,7 @@ "file_name": "cache/pos_embedprojConv_21.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2_context/Constant_output_0": { + "/transformer_blocks.0/norm2_context/Constant_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57731,7 +57731,7 @@ "file_name": "cache/pos_embedprojConv_22.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2_context/Constant_1_output_0": { + "/transformer_blocks.0/norm2_context/Constant_1_output_0_14_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57757,7 +57757,7 @@ "file_name": "cache/pos_embedprojConv_24.const", "file_size": 5382144 }, - "onnx::MatMul_7929": { + "onnx::MatMul_7929_45_1_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57770,7 +57770,7 @@ "file_name": "cache/pos_embedprojConv_25.const", "file_size": 10764288 }, - "onnx::MatMul_7930": { + "onnx::MatMul_7930_45_1_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57796,7 +57796,7 @@ "file_name": "cache/pos_embedprojConv_27.const", "file_size": 2691072 }, - "onnx::MatMul_7925": { + "onnx::MatMul_7925_45_5_0": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57822,7 +57822,7 @@ "file_name": "cache/pos_embedprojConv_29.const", "file_size": 2691072 }, - "/transformer_blocks.0/norm2/Constant_output_0": { + "/transformer_blocks.0/norm2/Constant_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57835,7 +57835,7 @@ "file_name": "cache/pos_embedprojConv_30.const", "file_size": 3072 }, - "/transformer_blocks.0/norm2/Constant_1_output_0": { + "/transformer_blocks.0/norm2/Constant_1_output_0_14_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57861,7 +57861,7 @@ "file_name": "cache/pos_embedprojConv_32.const", "file_size": 5382144 }, - "onnx::MatMul_7927": { + "onnx::MatMul_7927_45_1_1": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57874,7 +57874,7 @@ "file_name": "cache/pos_embedprojConv_33.const", "file_size": 10764288 }, - "onnx::MatMul_7928": { + "onnx::MatMul_7928_45_1_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -57900,7 +57900,7 @@ "file_name": "cache/pos_embedprojConv_35.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57913,7 +57913,7 @@ "file_name": "cache/pos_embedprojConv_36.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1_context/norm/Constant_1_output_0_14_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57926,7 +57926,7 @@ "file_name": "cache/pos_embedprojConv_37.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1/norm/Constant_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -57939,7 +57939,7 @@ "file_name": "cache/pos_embedprojConv_38.const", "file_size": 3072 }, - "/transformer_blocks.1/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.1/norm1/norm/Constant_1_output_0_14_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58017,7 +58017,7 @@ "file_name": "cache/pos_embedprojConv_44.const", "file_size": 5382144 }, - "onnx::MatMul_7950": { + "onnx::MatMul_7950_45_5_2": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58043,7 +58043,7 @@ "file_name": "cache/pos_embedprojConv_46.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2/Constant_output_0": { + "/transformer_blocks.1/norm2/Constant_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58056,7 +58056,7 @@ "file_name": "cache/pos_embedprojConv_47.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2/Constant_1_output_0": { + "/transformer_blocks.1/norm2/Constant_1_output_0_14_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58082,7 +58082,7 @@ "file_name": "cache/pos_embedprojConv_49.const", "file_size": 5382144 }, - "onnx::MatMul_7952": { + "onnx::MatMul_7952_45_1_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58095,7 +58095,7 @@ "file_name": "cache/pos_embedprojConv_50.const", "file_size": 10764288 }, - "onnx::MatMul_7953": { + "onnx::MatMul_7953_45_1_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58121,7 +58121,7 @@ "file_name": "cache/pos_embedprojConv_52.const", "file_size": 2691072 }, - "onnx::MatMul_7951": { + "onnx::MatMul_7951_45_5_3": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58147,7 +58147,7 @@ "file_name": "cache/pos_embedprojConv_54.const", "file_size": 2691072 }, - "/transformer_blocks.1/norm2_context/Constant_output_0": { + "/transformer_blocks.1/norm2_context/Constant_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58160,7 +58160,7 @@ "file_name": "cache/pos_embedprojConv_55.const", "file_size": 3072 }, - "/transformer_blocks.1/norm2_context/Constant_1_output_0": { + "/transformer_blocks.1/norm2_context/Constant_1_output_0_14_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58186,7 +58186,7 @@ "file_name": "cache/pos_embedprojConv_57.const", "file_size": 5382144 }, - "onnx::MatMul_7954": { + "onnx::MatMul_7954_45_1_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58199,7 +58199,7 @@ "file_name": "cache/pos_embedprojConv_58.const", "file_size": 10764288 }, - "onnx::MatMul_7955": { + "onnx::MatMul_7955_45_1_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58225,7 +58225,7 @@ "file_name": "cache/pos_embedprojConv_60.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm1/norm/Constant_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58238,7 +58238,7 @@ "file_name": "cache/pos_embedprojConv_61.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1/norm/Constant_1_output_0_14_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58251,7 +58251,7 @@ "file_name": "cache/pos_embedprojConv_62.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58264,7 +58264,7 @@ "file_name": "cache/pos_embedprojConv_63.const", "file_size": 3072 }, - "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.2/norm1_context/norm/Constant_1_output_0_14_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58342,7 +58342,7 @@ "file_name": "cache/pos_embedprojConv_69.const", "file_size": 5382144 }, - "onnx::MatMul_7976": { + "onnx::MatMul_7976_45_5_5": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58368,7 +58368,7 @@ "file_name": "cache/pos_embedprojConv_71.const", "file_size": 2691072 }, - "onnx::MatMul_7975": { + "onnx::MatMul_7975_45_5_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58394,7 +58394,7 @@ "file_name": "cache/pos_embedprojConv_73.const", "file_size": 2691072 }, - "/transformer_blocks.2/norm2/Constant_output_0": { + "/transformer_blocks.2/norm2/Constant_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58407,7 +58407,7 @@ "file_name": "cache/pos_embedprojConv_74.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2/Constant_1_output_0": { + "/transformer_blocks.2/norm2/Constant_1_output_0_14_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58433,7 +58433,7 @@ "file_name": "cache/pos_embedprojConv_76.const", "file_size": 5382144 }, - "onnx::MatMul_7977": { + "onnx::MatMul_7977_45_1_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58446,7 +58446,7 @@ "file_name": "cache/pos_embedprojConv_77.const", "file_size": 10764288 }, - "onnx::MatMul_7978": { + "onnx::MatMul_7978_45_1_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58459,7 +58459,7 @@ "file_name": "cache/pos_embedprojConv_78.const", "file_size": 10764288 }, - "/transformer_blocks.2/norm2_context/Constant_output_0": { + "/transformer_blocks.2/norm2_context/Constant_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58472,7 +58472,7 @@ "file_name": "cache/pos_embedprojConv_79.const", "file_size": 3072 }, - "/transformer_blocks.2/norm2_context/Constant_1_output_0": { + "/transformer_blocks.2/norm2_context/Constant_1_output_0_14_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58498,7 +58498,7 @@ "file_name": "cache/pos_embedprojConv_81.const", "file_size": 5382144 }, - "onnx::MatMul_7979": { + "onnx::MatMul_7979_45_1_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58511,7 +58511,7 @@ "file_name": "cache/pos_embedprojConv_82.const", "file_size": 10764288 }, - "onnx::MatMul_7980": { + "onnx::MatMul_7980_45_1_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58537,7 +58537,7 @@ "file_name": "cache/pos_embedprojConv_84.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58550,7 +58550,7 @@ "file_name": "cache/pos_embedprojConv_85.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1_context/norm/Constant_1_output_0_14_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58576,7 +58576,7 @@ "file_name": "cache/pos_embedprojConv_87.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm1/norm/Constant_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58589,7 +58589,7 @@ "file_name": "cache/pos_embedprojConv_88.const", "file_size": 3072 }, - "/transformer_blocks.3/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.3/norm1/norm/Constant_1_output_0_14_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58667,7 +58667,7 @@ "file_name": "cache/pos_embedprojConv_94.const", "file_size": 5382144 }, - "onnx::MatMul_8001": { + "onnx::MatMul_8001_45_5_7": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58693,7 +58693,7 @@ "file_name": "cache/pos_embedprojConv_96.const", "file_size": 2691072 }, - "onnx::MatMul_8000": { + "onnx::MatMul_8000_45_5_6": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58719,7 +58719,7 @@ "file_name": "cache/pos_embedprojConv_98.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2/Constant_output_0": { + "/transformer_blocks.3/norm2/Constant_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58732,7 +58732,7 @@ "file_name": "cache/pos_embedprojConv_99.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2/Constant_1_output_0": { + "/transformer_blocks.3/norm2/Constant_1_output_0_14_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58758,7 +58758,7 @@ "file_name": "cache/pos_embedprojConv_101.const", "file_size": 5382144 }, - "onnx::MatMul_8002": { + "onnx::MatMul_8002_45_1_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58771,7 +58771,7 @@ "file_name": "cache/pos_embedprojConv_102.const", "file_size": 10764288 }, - "onnx::MatMul_8003": { + "onnx::MatMul_8003_45_1_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58797,7 +58797,7 @@ "file_name": "cache/pos_embedprojConv_104.const", "file_size": 2691072 }, - "/transformer_blocks.3/norm2_context/Constant_output_0": { + "/transformer_blocks.3/norm2_context/Constant_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58810,7 +58810,7 @@ "file_name": "cache/pos_embedprojConv_105.const", "file_size": 3072 }, - "/transformer_blocks.3/norm2_context/Constant_1_output_0": { + "/transformer_blocks.3/norm2_context/Constant_1_output_0_14_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58836,7 +58836,7 @@ "file_name": "cache/pos_embedprojConv_107.const", "file_size": 5382144 }, - "onnx::MatMul_8004": { + "onnx::MatMul_8004_45_1_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58849,7 +58849,7 @@ "file_name": "cache/pos_embedprojConv_108.const", "file_size": 10764288 }, - "onnx::MatMul_8005": { + "onnx::MatMul_8005_45_1_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -58875,7 +58875,7 @@ "file_name": "cache/pos_embedprojConv_110.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm1/norm/Constant_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58888,7 +58888,7 @@ "file_name": "cache/pos_embedprojConv_111.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1/norm/Constant_1_output_0_14_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58901,7 +58901,7 @@ "file_name": "cache/pos_embedprojConv_112.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58914,7 +58914,7 @@ "file_name": "cache/pos_embedprojConv_113.const", "file_size": 3072 }, - "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.4/norm1_context/norm/Constant_1_output_0_14_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -58992,7 +58992,7 @@ "file_name": "cache/pos_embedprojConv_119.const", "file_size": 5382144 }, - "onnx::MatMul_8026": { + "onnx::MatMul_8026_45_5_9": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59018,7 +59018,7 @@ "file_name": "cache/pos_embedprojConv_121.const", "file_size": 2691072 }, - "onnx::MatMul_8025": { + "onnx::MatMul_8025_45_5_8": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59044,7 +59044,7 @@ "file_name": "cache/pos_embedprojConv_123.const", "file_size": 2691072 }, - "/transformer_blocks.4/norm2/Constant_output_0": { + "/transformer_blocks.4/norm2/Constant_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59057,7 +59057,7 @@ "file_name": "cache/pos_embedprojConv_124.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2/Constant_1_output_0": { + "/transformer_blocks.4/norm2/Constant_1_output_0_14_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59083,7 +59083,7 @@ "file_name": "cache/pos_embedprojConv_126.const", "file_size": 5382144 }, - "onnx::MatMul_8027": { + "onnx::MatMul_8027_45_1_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59096,7 +59096,7 @@ "file_name": "cache/pos_embedprojConv_127.const", "file_size": 10764288 }, - "onnx::MatMul_8028": { + "onnx::MatMul_8028_45_1_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59109,7 +59109,7 @@ "file_name": "cache/pos_embedprojConv_128.const", "file_size": 10764288 }, - "/transformer_blocks.4/norm2_context/Constant_output_0": { + "/transformer_blocks.4/norm2_context/Constant_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59122,7 +59122,7 @@ "file_name": "cache/pos_embedprojConv_129.const", "file_size": 3072 }, - "/transformer_blocks.4/norm2_context/Constant_1_output_0": { + "/transformer_blocks.4/norm2_context/Constant_1_output_0_14_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59148,7 +59148,7 @@ "file_name": "cache/pos_embedprojConv_131.const", "file_size": 5382144 }, - "onnx::MatMul_8029": { + "onnx::MatMul_8029_45_1_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59161,7 +59161,7 @@ "file_name": "cache/pos_embedprojConv_132.const", "file_size": 10764288 }, - "onnx::MatMul_8030": { + "onnx::MatMul_8030_45_1_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59187,7 +59187,7 @@ "file_name": "cache/pos_embedprojConv_134.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59200,7 +59200,7 @@ "file_name": "cache/pos_embedprojConv_135.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1_context/norm/Constant_1_output_0_14_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59226,7 +59226,7 @@ "file_name": "cache/pos_embedprojConv_137.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm1/norm/Constant_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59239,7 +59239,7 @@ "file_name": "cache/pos_embedprojConv_138.const", "file_size": 3072 }, - "/transformer_blocks.5/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.5/norm1/norm/Constant_1_output_0_14_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59317,7 +59317,7 @@ "file_name": "cache/pos_embedprojConv_144.const", "file_size": 5382144 }, - "onnx::MatMul_8051": { + "onnx::MatMul_8051_45_5_11": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59343,7 +59343,7 @@ "file_name": "cache/pos_embedprojConv_146.const", "file_size": 2691072 }, - "onnx::MatMul_8050": { + "onnx::MatMul_8050_45_5_10": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59369,7 +59369,7 @@ "file_name": "cache/pos_embedprojConv_148.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm2/Constant_output_0": { + "/transformer_blocks.5/norm2/Constant_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59382,7 +59382,7 @@ "file_name": "cache/pos_embedprojConv_149.const", "file_size": 3072 }, - "/transformer_blocks.5/norm2/Constant_1_output_0": { + "/transformer_blocks.5/norm2/Constant_1_output_0_14_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59408,7 +59408,7 @@ "file_name": "cache/pos_embedprojConv_151.const", "file_size": 5382144 }, - "onnx::MatMul_8052": { + "onnx::MatMul_8052_45_1_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59421,7 +59421,7 @@ "file_name": "cache/pos_embedprojConv_152.const", "file_size": 10764288 }, - "onnx::MatMul_8053": { + "onnx::MatMul_8053_45_1_23": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59447,7 +59447,7 @@ "file_name": "cache/pos_embedprojConv_154.const", "file_size": 2691072 }, - "/transformer_blocks.5/norm2_context/Constant_output_0": { + "/transformer_blocks.5/norm2_context/Constant_output_0_14_23": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59460,7 +59460,7 @@ "file_name": "cache/pos_embedprojConv_155.const", "file_size": 3072 }, - "/transformer_blocks.5/norm2_context/Constant_1_output_0": { + "/transformer_blocks.5/norm2_context/Constant_1_output_0_14_23": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59486,7 +59486,7 @@ "file_name": "cache/pos_embedprojConv_157.const", "file_size": 5382144 }, - "onnx::MatMul_8054": { + "onnx::MatMul_8054_45_1_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59499,7 +59499,7 @@ "file_name": "cache/pos_embedprojConv_158.const", "file_size": 10764288 }, - "onnx::MatMul_8055": { + "onnx::MatMul_8055_45_1_24": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59525,7 +59525,7 @@ "file_name": "cache/pos_embedprojConv_160.const", "file_size": 2691072 }, - "/transformer_blocks.6/norm1/norm/Constant_output_0": { + "/transformer_blocks.6/norm1/norm/Constant_output_0_14_25": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59538,7 +59538,7 @@ "file_name": "cache/pos_embedprojConv_161.const", "file_size": 3072 }, - "/transformer_blocks.6/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.6/norm1/norm/Constant_1_output_0_14_25": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59551,7 +59551,7 @@ "file_name": "cache/pos_embedprojConv_162.const", "file_size": 3072 }, - "/transformer_blocks.6/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.6/norm1_context/norm/Constant_output_0_14_24": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59564,7 +59564,7 @@ "file_name": "cache/pos_embedprojConv_163.const", "file_size": 3072 }, - "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.6/norm1_context/norm/Constant_1_output_0_14_24": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59642,7 +59642,7 @@ "file_name": "cache/pos_embedprojConv_169.const", "file_size": 5382144 }, - "onnx::MatMul_8076": { + "onnx::MatMul_8076_45_5_13": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59668,7 +59668,7 @@ "file_name": "cache/pos_embedprojConv_171.const", "file_size": 2691072 }, - "onnx::MatMul_8075": { + "onnx::MatMul_8075_45_5_12": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59694,7 +59694,7 @@ "file_name": "cache/pos_embedprojConv_173.const", "file_size": 2691072 }, - "/transformer_blocks.6/norm2/Constant_output_0": { + "/transformer_blocks.6/norm2/Constant_output_0_14_26": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59707,7 +59707,7 @@ "file_name": "cache/pos_embedprojConv_174.const", "file_size": 3072 }, - "/transformer_blocks.6/norm2/Constant_1_output_0": { + "/transformer_blocks.6/norm2/Constant_1_output_0_14_26": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59733,7 +59733,7 @@ "file_name": "cache/pos_embedprojConv_176.const", "file_size": 5382144 }, - "onnx::MatMul_8077": { + "onnx::MatMul_8077_45_1_25": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59746,7 +59746,7 @@ "file_name": "cache/pos_embedprojConv_177.const", "file_size": 10764288 }, - "onnx::MatMul_8078": { + "onnx::MatMul_8078_45_1_27": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59759,7 +59759,7 @@ "file_name": "cache/pos_embedprojConv_178.const", "file_size": 10764288 }, - "/transformer_blocks.6/norm2_context/Constant_output_0": { + "/transformer_blocks.6/norm2_context/Constant_output_0_14_27": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59772,7 +59772,7 @@ "file_name": "cache/pos_embedprojConv_179.const", "file_size": 3072 }, - "/transformer_blocks.6/norm2_context/Constant_1_output_0": { + "/transformer_blocks.6/norm2_context/Constant_1_output_0_14_27": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59798,7 +59798,7 @@ "file_name": "cache/pos_embedprojConv_181.const", "file_size": 5382144 }, - "onnx::MatMul_8079": { + "onnx::MatMul_8079_45_1_26": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59811,7 +59811,7 @@ "file_name": "cache/pos_embedprojConv_182.const", "file_size": 10764288 }, - "onnx::MatMul_8080": { + "onnx::MatMul_8080_45_1_28": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59837,7 +59837,7 @@ "file_name": "cache/pos_embedprojConv_184.const", "file_size": 2691072 }, - "/transformer_blocks.7/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.7/norm1_context/norm/Constant_output_0_14_28": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59850,7 +59850,7 @@ "file_name": "cache/pos_embedprojConv_185.const", "file_size": 3072 }, - "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.7/norm1_context/norm/Constant_1_output_0_14_28": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59876,7 +59876,7 @@ "file_name": "cache/pos_embedprojConv_187.const", "file_size": 2691072 }, - "/transformer_blocks.7/norm1/norm/Constant_output_0": { + "/transformer_blocks.7/norm1/norm/Constant_output_0_14_29": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59889,7 +59889,7 @@ "file_name": "cache/pos_embedprojConv_188.const", "file_size": 3072 }, - "/transformer_blocks.7/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.7/norm1/norm/Constant_1_output_0_14_29": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -59967,7 +59967,7 @@ "file_name": "cache/pos_embedprojConv_194.const", "file_size": 5382144 }, - "onnx::MatMul_8101": { + "onnx::MatMul_8101_45_5_15": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -59993,7 +59993,7 @@ "file_name": "cache/pos_embedprojConv_196.const", "file_size": 2691072 }, - "onnx::MatMul_8100": { + "onnx::MatMul_8100_45_5_14": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60019,7 +60019,7 @@ "file_name": "cache/pos_embedprojConv_198.const", "file_size": 2691072 }, - "/transformer_blocks.7/norm2/Constant_output_0": { + "/transformer_blocks.7/norm2/Constant_output_0_14_30": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60032,7 +60032,7 @@ "file_name": "cache/pos_embedprojConv_199.const", "file_size": 3072 }, - "/transformer_blocks.7/norm2/Constant_1_output_0": { + "/transformer_blocks.7/norm2/Constant_1_output_0_14_30": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60058,7 +60058,7 @@ "file_name": "cache/pos_embedprojConv_201.const", "file_size": 5382144 }, - "onnx::MatMul_8102": { + "onnx::MatMul_8102_45_1_29": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60071,7 +60071,7 @@ "file_name": "cache/pos_embedprojConv_202.const", "file_size": 10764288 }, - "onnx::MatMul_8103": { + "onnx::MatMul_8103_45_1_31": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60097,7 +60097,7 @@ "file_name": "cache/pos_embedprojConv_204.const", "file_size": 2691072 }, - "/transformer_blocks.7/norm2_context/Constant_output_0": { + "/transformer_blocks.7/norm2_context/Constant_output_0_14_31": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60110,7 +60110,7 @@ "file_name": "cache/pos_embedprojConv_205.const", "file_size": 3072 }, - "/transformer_blocks.7/norm2_context/Constant_1_output_0": { + "/transformer_blocks.7/norm2_context/Constant_1_output_0_14_31": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60136,7 +60136,7 @@ "file_name": "cache/pos_embedprojConv_207.const", "file_size": 5382144 }, - "onnx::MatMul_8104": { + "onnx::MatMul_8104_45_1_30": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60149,7 +60149,7 @@ "file_name": "cache/pos_embedprojConv_208.const", "file_size": 10764288 }, - "onnx::MatMul_8105": { + "onnx::MatMul_8105_45_1_32": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60175,7 +60175,7 @@ "file_name": "cache/pos_embedprojConv_210.const", "file_size": 2691072 }, - "/transformer_blocks.8/norm1/norm/Constant_output_0": { + "/transformer_blocks.8/norm1/norm/Constant_output_0_14_33": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60188,7 +60188,7 @@ "file_name": "cache/pos_embedprojConv_211.const", "file_size": 3072 }, - "/transformer_blocks.8/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.8/norm1/norm/Constant_1_output_0_14_33": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60201,7 +60201,7 @@ "file_name": "cache/pos_embedprojConv_212.const", "file_size": 3072 }, - "/transformer_blocks.8/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.8/norm1_context/norm/Constant_output_0_14_32": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60214,7 +60214,7 @@ "file_name": "cache/pos_embedprojConv_213.const", "file_size": 3072 }, - "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.8/norm1_context/norm/Constant_1_output_0_14_32": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60292,7 +60292,7 @@ "file_name": "cache/pos_embedprojConv_219.const", "file_size": 5382144 }, - "onnx::MatMul_8126": { + "onnx::MatMul_8126_45_5_17": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60318,7 +60318,7 @@ "file_name": "cache/pos_embedprojConv_221.const", "file_size": 2691072 }, - "onnx::MatMul_8125": { + "onnx::MatMul_8125_45_5_16": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60344,7 +60344,7 @@ "file_name": "cache/pos_embedprojConv_223.const", "file_size": 2691072 }, - "/transformer_blocks.8/norm2/Constant_output_0": { + "/transformer_blocks.8/norm2/Constant_output_0_14_34": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60357,7 +60357,7 @@ "file_name": "cache/pos_embedprojConv_224.const", "file_size": 3072 }, - "/transformer_blocks.8/norm2/Constant_1_output_0": { + "/transformer_blocks.8/norm2/Constant_1_output_0_14_34": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60383,7 +60383,7 @@ "file_name": "cache/pos_embedprojConv_226.const", "file_size": 5382144 }, - "onnx::MatMul_8127": { + "onnx::MatMul_8127_45_1_33": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60396,7 +60396,7 @@ "file_name": "cache/pos_embedprojConv_227.const", "file_size": 10764288 }, - "onnx::MatMul_8128": { + "onnx::MatMul_8128_45_1_35": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60409,7 +60409,7 @@ "file_name": "cache/pos_embedprojConv_228.const", "file_size": 10764288 }, - "/transformer_blocks.8/norm2_context/Constant_output_0": { + "/transformer_blocks.8/norm2_context/Constant_output_0_14_35": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60422,7 +60422,7 @@ "file_name": "cache/pos_embedprojConv_229.const", "file_size": 3072 }, - "/transformer_blocks.8/norm2_context/Constant_1_output_0": { + "/transformer_blocks.8/norm2_context/Constant_1_output_0_14_35": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60448,7 +60448,7 @@ "file_name": "cache/pos_embedprojConv_231.const", "file_size": 5382144 }, - "onnx::MatMul_8129": { + "onnx::MatMul_8129_45_1_34": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60461,7 +60461,7 @@ "file_name": "cache/pos_embedprojConv_232.const", "file_size": 10764288 }, - "onnx::MatMul_8130": { + "onnx::MatMul_8130_45_1_36": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60487,7 +60487,7 @@ "file_name": "cache/pos_embedprojConv_234.const", "file_size": 2691072 }, - "/transformer_blocks.9/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.9/norm1_context/norm/Constant_output_0_14_36": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60500,7 +60500,7 @@ "file_name": "cache/pos_embedprojConv_235.const", "file_size": 3072 }, - "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.9/norm1_context/norm/Constant_1_output_0_14_36": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60526,7 +60526,7 @@ "file_name": "cache/pos_embedprojConv_237.const", "file_size": 2691072 }, - "/transformer_blocks.9/norm1/norm/Constant_output_0": { + "/transformer_blocks.9/norm1/norm/Constant_output_0_14_37": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60539,7 +60539,7 @@ "file_name": "cache/pos_embedprojConv_238.const", "file_size": 3072 }, - "/transformer_blocks.9/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.9/norm1/norm/Constant_1_output_0_14_37": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60617,7 +60617,7 @@ "file_name": "cache/pos_embedprojConv_244.const", "file_size": 5382144 }, - "onnx::MatMul_8151": { + "onnx::MatMul_8151_45_5_19": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60643,7 +60643,7 @@ "file_name": "cache/pos_embedprojConv_246.const", "file_size": 2691072 }, - "onnx::MatMul_8150": { + "onnx::MatMul_8150_45_5_18": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60669,7 +60669,7 @@ "file_name": "cache/pos_embedprojConv_248.const", "file_size": 2691072 }, - "/transformer_blocks.9/norm2/Constant_output_0": { + "/transformer_blocks.9/norm2/Constant_output_0_14_38": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60682,7 +60682,7 @@ "file_name": "cache/pos_embedprojConv_249.const", "file_size": 3072 }, - "/transformer_blocks.9/norm2/Constant_1_output_0": { + "/transformer_blocks.9/norm2/Constant_1_output_0_14_38": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60708,7 +60708,7 @@ "file_name": "cache/pos_embedprojConv_251.const", "file_size": 5382144 }, - "onnx::MatMul_8152": { + "onnx::MatMul_8152_45_1_37": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60721,7 +60721,7 @@ "file_name": "cache/pos_embedprojConv_252.const", "file_size": 10764288 }, - "onnx::MatMul_8153": { + "onnx::MatMul_8153_45_1_39": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60747,7 +60747,7 @@ "file_name": "cache/pos_embedprojConv_254.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm1/norm/Constant_output_0": { + "/transformer_blocks.10/norm1/norm/Constant_output_0_14_40": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60760,7 +60760,7 @@ "file_name": "cache/pos_embedprojConv_255.const", "file_size": 3072 }, - "/transformer_blocks.10/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.10/norm1/norm/Constant_1_output_0_14_40": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60786,7 +60786,7 @@ "file_name": "cache/pos_embedprojConv_257.const", "file_size": 5382144 }, - "/transformer_blocks.9/norm2_context/Constant_output_0": { + "/transformer_blocks.9/norm2_context/Constant_output_0_14_39": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60799,7 +60799,7 @@ "file_name": "cache/pos_embedprojConv_258.const", "file_size": 3072 }, - "/transformer_blocks.9/norm2_context/Constant_1_output_0": { + "/transformer_blocks.9/norm2_context/Constant_1_output_0_14_39": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60825,7 +60825,7 @@ "file_name": "cache/pos_embedprojConv_260.const", "file_size": 5382144 }, - "onnx::MatMul_8154": { + "onnx::MatMul_8154_45_1_38": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60838,7 +60838,7 @@ "file_name": "cache/pos_embedprojConv_261.const", "file_size": 10764288 }, - "onnx::MatMul_8155": { + "onnx::MatMul_8155_45_1_40": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60864,7 +60864,7 @@ "file_name": "cache/pos_embedprojConv_263.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.10/norm1_context/norm/Constant_output_0_14_41": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60877,7 +60877,7 @@ "file_name": "cache/pos_embedprojConv_264.const", "file_size": 3072 }, - "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.10/norm1_context/norm/Constant_1_output_0_14_41": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60942,7 +60942,7 @@ "file_name": "cache/pos_embedprojConv_269.const", "file_size": 5382144 }, - "onnx::MatMul_8175": { + "onnx::MatMul_8175_45_5_20": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -60968,7 +60968,7 @@ "file_name": "cache/pos_embedprojConv_271.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm2/Constant_output_0": { + "/transformer_blocks.10/norm2/Constant_output_0_14_42": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -60981,7 +60981,7 @@ "file_name": "cache/pos_embedprojConv_272.const", "file_size": 3072 }, - "/transformer_blocks.10/norm2/Constant_1_output_0": { + "/transformer_blocks.10/norm2/Constant_1_output_0_14_42": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61007,7 +61007,7 @@ "file_name": "cache/pos_embedprojConv_274.const", "file_size": 5382144 }, - "onnx::MatMul_8177": { + "onnx::MatMul_8177_45_1_41": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61020,7 +61020,7 @@ "file_name": "cache/pos_embedprojConv_275.const", "file_size": 10764288 }, - "onnx::MatMul_8178": { + "onnx::MatMul_8178_45_1_43": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61046,7 +61046,7 @@ "file_name": "cache/pos_embedprojConv_277.const", "file_size": 2691072 }, - "onnx::MatMul_8176": { + "onnx::MatMul_8176_45_5_21": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61072,7 +61072,7 @@ "file_name": "cache/pos_embedprojConv_279.const", "file_size": 2691072 }, - "/transformer_blocks.10/norm2_context/Constant_output_0": { + "/transformer_blocks.10/norm2_context/Constant_output_0_14_43": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61085,7 +61085,7 @@ "file_name": "cache/pos_embedprojConv_280.const", "file_size": 3072 }, - "/transformer_blocks.10/norm2_context/Constant_1_output_0": { + "/transformer_blocks.10/norm2_context/Constant_1_output_0_14_43": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61111,7 +61111,7 @@ "file_name": "cache/pos_embedprojConv_282.const", "file_size": 5382144 }, - "onnx::MatMul_8179": { + "onnx::MatMul_8179_45_1_42": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61124,7 +61124,7 @@ "file_name": "cache/pos_embedprojConv_283.const", "file_size": 10764288 }, - "onnx::MatMul_8180": { + "onnx::MatMul_8180_45_1_44": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61150,7 +61150,7 @@ "file_name": "cache/pos_embedprojConv_285.const", "file_size": 2691072 }, - "/transformer_blocks.11/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.11/norm1_context/norm/Constant_output_0_14_44": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61163,7 +61163,7 @@ "file_name": "cache/pos_embedprojConv_286.const", "file_size": 3072 }, - "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.11/norm1_context/norm/Constant_1_output_0_14_44": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61176,7 +61176,7 @@ "file_name": "cache/pos_embedprojConv_287.const", "file_size": 3072 }, - "/transformer_blocks.11/norm1/norm/Constant_output_0": { + "/transformer_blocks.11/norm1/norm/Constant_output_0_14_45": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61189,7 +61189,7 @@ "file_name": "cache/pos_embedprojConv_288.const", "file_size": 3072 }, - "/transformer_blocks.11/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.11/norm1/norm/Constant_1_output_0_14_45": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61267,7 +61267,7 @@ "file_name": "cache/pos_embedprojConv_294.const", "file_size": 5382144 }, - "onnx::MatMul_8200": { + "onnx::MatMul_8200_45_5_22": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61293,7 +61293,7 @@ "file_name": "cache/pos_embedprojConv_296.const", "file_size": 2691072 }, - "/transformer_blocks.11/norm2/Constant_output_0": { + "/transformer_blocks.11/norm2/Constant_output_0_14_46": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61306,7 +61306,7 @@ "file_name": "cache/pos_embedprojConv_297.const", "file_size": 3072 }, - "/transformer_blocks.11/norm2/Constant_1_output_0": { + "/transformer_blocks.11/norm2/Constant_1_output_0_14_46": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61332,7 +61332,7 @@ "file_name": "cache/pos_embedprojConv_299.const", "file_size": 5382144 }, - "onnx::MatMul_8202": { + "onnx::MatMul_8202_45_1_45": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61345,7 +61345,7 @@ "file_name": "cache/pos_embedprojConv_300.const", "file_size": 10764288 }, - "onnx::MatMul_8203": { + "onnx::MatMul_8203_45_1_47": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61371,7 +61371,7 @@ "file_name": "cache/pos_embedprojConv_302.const", "file_size": 2691072 }, - "onnx::MatMul_8201": { + "onnx::MatMul_8201_45_5_23": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61397,7 +61397,7 @@ "file_name": "cache/pos_embedprojConv_304.const", "file_size": 2691072 }, - "/transformer_blocks.11/norm2_context/Constant_output_0": { + "/transformer_blocks.11/norm2_context/Constant_output_0_14_47": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61410,7 +61410,7 @@ "file_name": "cache/pos_embedprojConv_305.const", "file_size": 3072 }, - "/transformer_blocks.11/norm2_context/Constant_1_output_0": { + "/transformer_blocks.11/norm2_context/Constant_1_output_0_14_47": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61436,7 +61436,7 @@ "file_name": "cache/pos_embedprojConv_307.const", "file_size": 5382144 }, - "onnx::MatMul_8204": { + "onnx::MatMul_8204_45_1_46": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61449,7 +61449,7 @@ "file_name": "cache/pos_embedprojConv_308.const", "file_size": 10764288 }, - "onnx::MatMul_8205": { + "onnx::MatMul_8205_45_1_48": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61475,7 +61475,7 @@ "file_name": "cache/pos_embedprojConv_310.const", "file_size": 2691072 }, - "/transformer_blocks.12/norm1/norm/Constant_output_0": { + "/transformer_blocks.12/norm1/norm/Constant_output_0_14_49": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61488,7 +61488,7 @@ "file_name": "cache/pos_embedprojConv_311.const", "file_size": 3072 }, - "/transformer_blocks.12/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.12/norm1/norm/Constant_1_output_0_14_49": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61514,7 +61514,7 @@ "file_name": "cache/pos_embedprojConv_313.const", "file_size": 5382144 }, - "/transformer_blocks.12/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.12/norm1_context/norm/Constant_output_0_14_48": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61527,7 +61527,7 @@ "file_name": "cache/pos_embedprojConv_314.const", "file_size": 3072 }, - "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.12/norm1_context/norm/Constant_1_output_0_14_48": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61592,7 +61592,7 @@ "file_name": "cache/pos_embedprojConv_319.const", "file_size": 5382144 }, - "onnx::MatMul_8225": { + "onnx::MatMul_8225_45_5_24": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61618,7 +61618,7 @@ "file_name": "cache/pos_embedprojConv_321.const", "file_size": 2691072 }, - "/transformer_blocks.12/norm2/Constant_output_0": { + "/transformer_blocks.12/norm2/Constant_output_0_14_50": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61631,7 +61631,7 @@ "file_name": "cache/pos_embedprojConv_322.const", "file_size": 3072 }, - "/transformer_blocks.12/norm2/Constant_1_output_0": { + "/transformer_blocks.12/norm2/Constant_1_output_0_14_50": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61657,7 +61657,7 @@ "file_name": "cache/pos_embedprojConv_324.const", "file_size": 5382144 }, - "onnx::MatMul_8227": { + "onnx::MatMul_8227_45_1_49": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61670,7 +61670,7 @@ "file_name": "cache/pos_embedprojConv_325.const", "file_size": 10764288 }, - "onnx::MatMul_8228": { + "onnx::MatMul_8228_45_1_51": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61696,7 +61696,7 @@ "file_name": "cache/pos_embedprojConv_327.const", "file_size": 2691072 }, - "onnx::MatMul_8226": { + "onnx::MatMul_8226_45_5_25": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61722,7 +61722,7 @@ "file_name": "cache/pos_embedprojConv_329.const", "file_size": 2691072 }, - "/transformer_blocks.12/norm2_context/Constant_output_0": { + "/transformer_blocks.12/norm2_context/Constant_output_0_14_51": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61735,7 +61735,7 @@ "file_name": "cache/pos_embedprojConv_330.const", "file_size": 3072 }, - "/transformer_blocks.12/norm2_context/Constant_1_output_0": { + "/transformer_blocks.12/norm2_context/Constant_1_output_0_14_51": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61761,7 +61761,7 @@ "file_name": "cache/pos_embedprojConv_332.const", "file_size": 5382144 }, - "onnx::MatMul_8229": { + "onnx::MatMul_8229_45_1_50": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61774,7 +61774,7 @@ "file_name": "cache/pos_embedprojConv_333.const", "file_size": 10764288 }, - "onnx::MatMul_8230": { + "onnx::MatMul_8230_45_1_52": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61800,7 +61800,7 @@ "file_name": "cache/pos_embedprojConv_335.const", "file_size": 2691072 }, - "/transformer_blocks.13/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.13/norm1_context/norm/Constant_output_0_14_52": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61813,7 +61813,7 @@ "file_name": "cache/pos_embedprojConv_336.const", "file_size": 3072 }, - "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.13/norm1_context/norm/Constant_1_output_0_14_52": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61826,7 +61826,7 @@ "file_name": "cache/pos_embedprojConv_337.const", "file_size": 3072 }, - "/transformer_blocks.13/norm1/norm/Constant_output_0": { + "/transformer_blocks.13/norm1/norm/Constant_output_0_14_53": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61839,7 +61839,7 @@ "file_name": "cache/pos_embedprojConv_338.const", "file_size": 3072 }, - "/transformer_blocks.13/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.13/norm1/norm/Constant_1_output_0_14_53": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61917,7 +61917,7 @@ "file_name": "cache/pos_embedprojConv_344.const", "file_size": 5382144 }, - "onnx::MatMul_8250": { + "onnx::MatMul_8250_45_5_26": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61943,7 +61943,7 @@ "file_name": "cache/pos_embedprojConv_346.const", "file_size": 2691072 }, - "/transformer_blocks.13/norm2/Constant_output_0": { + "/transformer_blocks.13/norm2/Constant_output_0_14_54": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61956,7 +61956,7 @@ "file_name": "cache/pos_embedprojConv_347.const", "file_size": 3072 }, - "/transformer_blocks.13/norm2/Constant_1_output_0": { + "/transformer_blocks.13/norm2/Constant_1_output_0_14_54": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -61982,7 +61982,7 @@ "file_name": "cache/pos_embedprojConv_349.const", "file_size": 5382144 }, - "onnx::MatMul_8252": { + "onnx::MatMul_8252_45_1_53": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -61995,7 +61995,7 @@ "file_name": "cache/pos_embedprojConv_350.const", "file_size": 10764288 }, - "onnx::MatMul_8253": { + "onnx::MatMul_8253_45_1_55": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62021,7 +62021,7 @@ "file_name": "cache/pos_embedprojConv_352.const", "file_size": 2691072 }, - "onnx::MatMul_8251": { + "onnx::MatMul_8251_45_5_27": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62047,7 +62047,7 @@ "file_name": "cache/pos_embedprojConv_354.const", "file_size": 2691072 }, - "/transformer_blocks.13/norm2_context/Constant_output_0": { + "/transformer_blocks.13/norm2_context/Constant_output_0_14_55": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62060,7 +62060,7 @@ "file_name": "cache/pos_embedprojConv_355.const", "file_size": 3072 }, - "/transformer_blocks.13/norm2_context/Constant_1_output_0": { + "/transformer_blocks.13/norm2_context/Constant_1_output_0_14_55": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62086,7 +62086,7 @@ "file_name": "cache/pos_embedprojConv_357.const", "file_size": 5382144 }, - "onnx::MatMul_8254": { + "onnx::MatMul_8254_45_1_54": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62099,7 +62099,7 @@ "file_name": "cache/pos_embedprojConv_358.const", "file_size": 10764288 }, - "onnx::MatMul_8255": { + "onnx::MatMul_8255_45_1_56": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62125,7 +62125,7 @@ "file_name": "cache/pos_embedprojConv_360.const", "file_size": 2691072 }, - "/transformer_blocks.14/norm1/norm/Constant_output_0": { + "/transformer_blocks.14/norm1/norm/Constant_output_0_14_57": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62138,7 +62138,7 @@ "file_name": "cache/pos_embedprojConv_361.const", "file_size": 3072 }, - "/transformer_blocks.14/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.14/norm1/norm/Constant_1_output_0_14_57": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62164,7 +62164,7 @@ "file_name": "cache/pos_embedprojConv_363.const", "file_size": 5382144 }, - "/transformer_blocks.14/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.14/norm1_context/norm/Constant_output_0_14_56": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62177,7 +62177,7 @@ "file_name": "cache/pos_embedprojConv_364.const", "file_size": 3072 }, - "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.14/norm1_context/norm/Constant_1_output_0_14_56": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62242,7 +62242,7 @@ "file_name": "cache/pos_embedprojConv_369.const", "file_size": 5382144 }, - "onnx::MatMul_8275": { + "onnx::MatMul_8275_45_5_28": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62268,7 +62268,7 @@ "file_name": "cache/pos_embedprojConv_371.const", "file_size": 2691072 }, - "/transformer_blocks.14/norm2/Constant_output_0": { + "/transformer_blocks.14/norm2/Constant_output_0_14_58": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62281,7 +62281,7 @@ "file_name": "cache/pos_embedprojConv_372.const", "file_size": 3072 }, - "/transformer_blocks.14/norm2/Constant_1_output_0": { + "/transformer_blocks.14/norm2/Constant_1_output_0_14_58": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62307,7 +62307,7 @@ "file_name": "cache/pos_embedprojConv_374.const", "file_size": 5382144 }, - "onnx::MatMul_8277": { + "onnx::MatMul_8277_45_1_57": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62320,7 +62320,7 @@ "file_name": "cache/pos_embedprojConv_375.const", "file_size": 10764288 }, - "onnx::MatMul_8278": { + "onnx::MatMul_8278_45_1_59": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62346,7 +62346,7 @@ "file_name": "cache/pos_embedprojConv_377.const", "file_size": 2691072 }, - "onnx::MatMul_8276": { + "onnx::MatMul_8276_45_5_29": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62372,7 +62372,7 @@ "file_name": "cache/pos_embedprojConv_379.const", "file_size": 2691072 }, - "/transformer_blocks.14/norm2_context/Constant_output_0": { + "/transformer_blocks.14/norm2_context/Constant_output_0_14_59": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62385,7 +62385,7 @@ "file_name": "cache/pos_embedprojConv_380.const", "file_size": 3072 }, - "/transformer_blocks.14/norm2_context/Constant_1_output_0": { + "/transformer_blocks.14/norm2_context/Constant_1_output_0_14_59": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62411,7 +62411,7 @@ "file_name": "cache/pos_embedprojConv_382.const", "file_size": 5382144 }, - "onnx::MatMul_8279": { + "onnx::MatMul_8279_45_1_58": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62424,7 +62424,7 @@ "file_name": "cache/pos_embedprojConv_383.const", "file_size": 10764288 }, - "onnx::MatMul_8280": { + "onnx::MatMul_8280_45_1_60": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62450,7 +62450,7 @@ "file_name": "cache/pos_embedprojConv_385.const", "file_size": 2691072 }, - "/transformer_blocks.15/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.15/norm1_context/norm/Constant_output_0_14_60": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62463,7 +62463,7 @@ "file_name": "cache/pos_embedprojConv_386.const", "file_size": 3072 }, - "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.15/norm1_context/norm/Constant_1_output_0_14_60": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62476,7 +62476,7 @@ "file_name": "cache/pos_embedprojConv_387.const", "file_size": 3072 }, - "/transformer_blocks.15/norm1/norm/Constant_output_0": { + "/transformer_blocks.15/norm1/norm/Constant_output_0_14_61": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62489,7 +62489,7 @@ "file_name": "cache/pos_embedprojConv_388.const", "file_size": 3072 }, - "/transformer_blocks.15/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.15/norm1/norm/Constant_1_output_0_14_61": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62567,7 +62567,7 @@ "file_name": "cache/pos_embedprojConv_394.const", "file_size": 5382144 }, - "onnx::MatMul_8300": { + "onnx::MatMul_8300_45_5_30": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62593,7 +62593,7 @@ "file_name": "cache/pos_embedprojConv_396.const", "file_size": 2691072 }, - "/transformer_blocks.15/norm2/Constant_output_0": { + "/transformer_blocks.15/norm2/Constant_output_0_14_62": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62606,7 +62606,7 @@ "file_name": "cache/pos_embedprojConv_397.const", "file_size": 3072 }, - "/transformer_blocks.15/norm2/Constant_1_output_0": { + "/transformer_blocks.15/norm2/Constant_1_output_0_14_62": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62632,7 +62632,7 @@ "file_name": "cache/pos_embedprojConv_399.const", "file_size": 5382144 }, - "onnx::MatMul_8302": { + "onnx::MatMul_8302_45_1_61": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62645,7 +62645,7 @@ "file_name": "cache/pos_embedprojConv_400.const", "file_size": 10764288 }, - "onnx::MatMul_8303": { + "onnx::MatMul_8303_45_1_63": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62671,7 +62671,7 @@ "file_name": "cache/pos_embedprojConv_402.const", "file_size": 2691072 }, - "onnx::MatMul_8301": { + "onnx::MatMul_8301_45_5_31": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62697,7 +62697,7 @@ "file_name": "cache/pos_embedprojConv_404.const", "file_size": 2691072 }, - "/transformer_blocks.15/norm2_context/Constant_output_0": { + "/transformer_blocks.15/norm2_context/Constant_output_0_14_63": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62710,7 +62710,7 @@ "file_name": "cache/pos_embedprojConv_405.const", "file_size": 3072 }, - "/transformer_blocks.15/norm2_context/Constant_1_output_0": { + "/transformer_blocks.15/norm2_context/Constant_1_output_0_14_63": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62736,7 +62736,7 @@ "file_name": "cache/pos_embedprojConv_407.const", "file_size": 5382144 }, - "onnx::MatMul_8304": { + "onnx::MatMul_8304_45_1_62": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62749,7 +62749,7 @@ "file_name": "cache/pos_embedprojConv_408.const", "file_size": 10764288 }, - "onnx::MatMul_8305": { + "onnx::MatMul_8305_45_1_64": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62775,7 +62775,7 @@ "file_name": "cache/pos_embedprojConv_410.const", "file_size": 2691072 }, - "/transformer_blocks.16/norm1/norm/Constant_output_0": { + "/transformer_blocks.16/norm1/norm/Constant_output_0_14_65": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62788,7 +62788,7 @@ "file_name": "cache/pos_embedprojConv_411.const", "file_size": 3072 }, - "/transformer_blocks.16/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.16/norm1/norm/Constant_1_output_0_14_65": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62814,7 +62814,7 @@ "file_name": "cache/pos_embedprojConv_413.const", "file_size": 5382144 }, - "/transformer_blocks.16/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.16/norm1_context/norm/Constant_output_0_14_64": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62827,7 +62827,7 @@ "file_name": "cache/pos_embedprojConv_414.const", "file_size": 3072 }, - "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.16/norm1_context/norm/Constant_1_output_0_14_64": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62892,7 +62892,7 @@ "file_name": "cache/pos_embedprojConv_419.const", "file_size": 5382144 }, - "onnx::MatMul_8325": { + "onnx::MatMul_8325_45_5_32": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62918,7 +62918,7 @@ "file_name": "cache/pos_embedprojConv_421.const", "file_size": 2691072 }, - "/transformer_blocks.16/norm2/Constant_output_0": { + "/transformer_blocks.16/norm2/Constant_output_0_14_66": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62931,7 +62931,7 @@ "file_name": "cache/pos_embedprojConv_422.const", "file_size": 3072 }, - "/transformer_blocks.16/norm2/Constant_1_output_0": { + "/transformer_blocks.16/norm2/Constant_1_output_0_14_66": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -62957,7 +62957,7 @@ "file_name": "cache/pos_embedprojConv_424.const", "file_size": 5382144 }, - "onnx::MatMul_8327": { + "onnx::MatMul_8327_45_1_65": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62970,7 +62970,7 @@ "file_name": "cache/pos_embedprojConv_425.const", "file_size": 10764288 }, - "onnx::MatMul_8328": { + "onnx::MatMul_8328_45_1_67": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -62996,7 +62996,7 @@ "file_name": "cache/pos_embedprojConv_427.const", "file_size": 2691072 }, - "onnx::MatMul_8326": { + "onnx::MatMul_8326_45_5_33": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63022,7 +63022,7 @@ "file_name": "cache/pos_embedprojConv_429.const", "file_size": 2691072 }, - "/transformer_blocks.16/norm2_context/Constant_output_0": { + "/transformer_blocks.16/norm2_context/Constant_output_0_14_67": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63035,7 +63035,7 @@ "file_name": "cache/pos_embedprojConv_430.const", "file_size": 3072 }, - "/transformer_blocks.16/norm2_context/Constant_1_output_0": { + "/transformer_blocks.16/norm2_context/Constant_1_output_0_14_67": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63061,7 +63061,7 @@ "file_name": "cache/pos_embedprojConv_432.const", "file_size": 5382144 }, - "onnx::MatMul_8329": { + "onnx::MatMul_8329_45_1_66": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63074,7 +63074,7 @@ "file_name": "cache/pos_embedprojConv_433.const", "file_size": 10764288 }, - "onnx::MatMul_8330": { + "onnx::MatMul_8330_45_1_68": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63100,7 +63100,7 @@ "file_name": "cache/pos_embedprojConv_435.const", "file_size": 2691072 }, - "/transformer_blocks.17/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.17/norm1_context/norm/Constant_output_0_14_68": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63113,7 +63113,7 @@ "file_name": "cache/pos_embedprojConv_436.const", "file_size": 3072 }, - "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.17/norm1_context/norm/Constant_1_output_0_14_68": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63126,7 +63126,7 @@ "file_name": "cache/pos_embedprojConv_437.const", "file_size": 3072 }, - "/transformer_blocks.17/norm1/norm/Constant_output_0": { + "/transformer_blocks.17/norm1/norm/Constant_output_0_14_69": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63139,7 +63139,7 @@ "file_name": "cache/pos_embedprojConv_438.const", "file_size": 3072 }, - "/transformer_blocks.17/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.17/norm1/norm/Constant_1_output_0_14_69": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63217,7 +63217,7 @@ "file_name": "cache/pos_embedprojConv_444.const", "file_size": 5382144 }, - "onnx::MatMul_8350": { + "onnx::MatMul_8350_45_5_34": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63243,7 +63243,7 @@ "file_name": "cache/pos_embedprojConv_446.const", "file_size": 2691072 }, - "/transformer_blocks.17/norm2/Constant_output_0": { + "/transformer_blocks.17/norm2/Constant_output_0_14_70": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63256,7 +63256,7 @@ "file_name": "cache/pos_embedprojConv_447.const", "file_size": 3072 }, - "/transformer_blocks.17/norm2/Constant_1_output_0": { + "/transformer_blocks.17/norm2/Constant_1_output_0_14_70": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63282,7 +63282,7 @@ "file_name": "cache/pos_embedprojConv_449.const", "file_size": 5382144 }, - "onnx::MatMul_8352": { + "onnx::MatMul_8352_45_1_69": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63295,7 +63295,7 @@ "file_name": "cache/pos_embedprojConv_450.const", "file_size": 10764288 }, - "onnx::MatMul_8353": { + "onnx::MatMul_8353_45_1_71": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63321,7 +63321,7 @@ "file_name": "cache/pos_embedprojConv_452.const", "file_size": 2691072 }, - "onnx::MatMul_8351": { + "onnx::MatMul_8351_45_5_35": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63347,7 +63347,7 @@ "file_name": "cache/pos_embedprojConv_454.const", "file_size": 2691072 }, - "/transformer_blocks.17/norm2_context/Constant_output_0": { + "/transformer_blocks.17/norm2_context/Constant_output_0_14_71": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63360,7 +63360,7 @@ "file_name": "cache/pos_embedprojConv_455.const", "file_size": 3072 }, - "/transformer_blocks.17/norm2_context/Constant_1_output_0": { + "/transformer_blocks.17/norm2_context/Constant_1_output_0_14_71": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63386,7 +63386,7 @@ "file_name": "cache/pos_embedprojConv_457.const", "file_size": 5382144 }, - "onnx::MatMul_8354": { + "onnx::MatMul_8354_45_1_70": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63399,7 +63399,7 @@ "file_name": "cache/pos_embedprojConv_458.const", "file_size": 10764288 }, - "onnx::MatMul_8355": { + "onnx::MatMul_8355_45_1_72": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63425,7 +63425,7 @@ "file_name": "cache/pos_embedprojConv_460.const", "file_size": 2691072 }, - "/transformer_blocks.18/norm1/norm/Constant_output_0": { + "/transformer_blocks.18/norm1/norm/Constant_output_0_14_73": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63438,7 +63438,7 @@ "file_name": "cache/pos_embedprojConv_461.const", "file_size": 3072 }, - "/transformer_blocks.18/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.18/norm1/norm/Constant_1_output_0_14_73": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63464,7 +63464,7 @@ "file_name": "cache/pos_embedprojConv_463.const", "file_size": 5382144 }, - "/transformer_blocks.18/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.18/norm1_context/norm/Constant_output_0_14_72": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63477,7 +63477,7 @@ "file_name": "cache/pos_embedprojConv_464.const", "file_size": 3072 }, - "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.18/norm1_context/norm/Constant_1_output_0_14_72": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63542,7 +63542,7 @@ "file_name": "cache/pos_embedprojConv_469.const", "file_size": 5382144 }, - "onnx::MatMul_8375": { + "onnx::MatMul_8375_45_5_36": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63568,7 +63568,7 @@ "file_name": "cache/pos_embedprojConv_471.const", "file_size": 2691072 }, - "/transformer_blocks.18/norm2/Constant_output_0": { + "/transformer_blocks.18/norm2/Constant_output_0_14_74": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63581,7 +63581,7 @@ "file_name": "cache/pos_embedprojConv_472.const", "file_size": 3072 }, - "/transformer_blocks.18/norm2/Constant_1_output_0": { + "/transformer_blocks.18/norm2/Constant_1_output_0_14_74": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63607,7 +63607,7 @@ "file_name": "cache/pos_embedprojConv_474.const", "file_size": 5382144 }, - "onnx::MatMul_8377": { + "onnx::MatMul_8377_45_1_73": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63620,7 +63620,7 @@ "file_name": "cache/pos_embedprojConv_475.const", "file_size": 10764288 }, - "onnx::MatMul_8378": { + "onnx::MatMul_8378_45_1_75": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63646,7 +63646,7 @@ "file_name": "cache/pos_embedprojConv_477.const", "file_size": 2691072 }, - "onnx::MatMul_8376": { + "onnx::MatMul_8376_45_5_37": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63672,7 +63672,7 @@ "file_name": "cache/pos_embedprojConv_479.const", "file_size": 2691072 }, - "/transformer_blocks.18/norm2_context/Constant_output_0": { + "/transformer_blocks.18/norm2_context/Constant_output_0_14_75": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63685,7 +63685,7 @@ "file_name": "cache/pos_embedprojConv_480.const", "file_size": 3072 }, - "/transformer_blocks.18/norm2_context/Constant_1_output_0": { + "/transformer_blocks.18/norm2_context/Constant_1_output_0_14_75": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63711,7 +63711,7 @@ "file_name": "cache/pos_embedprojConv_482.const", "file_size": 5382144 }, - "onnx::MatMul_8379": { + "onnx::MatMul_8379_45_1_74": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63724,7 +63724,7 @@ "file_name": "cache/pos_embedprojConv_483.const", "file_size": 10764288 }, - "onnx::MatMul_8380": { + "onnx::MatMul_8380_45_1_76": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63750,7 +63750,7 @@ "file_name": "cache/pos_embedprojConv_485.const", "file_size": 2691072 }, - "/transformer_blocks.19/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.19/norm1_context/norm/Constant_output_0_14_76": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63763,7 +63763,7 @@ "file_name": "cache/pos_embedprojConv_486.const", "file_size": 3072 }, - "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.19/norm1_context/norm/Constant_1_output_0_14_76": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63776,7 +63776,7 @@ "file_name": "cache/pos_embedprojConv_487.const", "file_size": 3072 }, - "/transformer_blocks.19/norm1/norm/Constant_output_0": { + "/transformer_blocks.19/norm1/norm/Constant_output_0_14_77": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63789,7 +63789,7 @@ "file_name": "cache/pos_embedprojConv_488.const", "file_size": 3072 }, - "/transformer_blocks.19/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.19/norm1/norm/Constant_1_output_0_14_77": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63867,7 +63867,7 @@ "file_name": "cache/pos_embedprojConv_494.const", "file_size": 5382144 }, - "onnx::MatMul_8400": { + "onnx::MatMul_8400_45_5_38": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63893,7 +63893,7 @@ "file_name": "cache/pos_embedprojConv_496.const", "file_size": 2691072 }, - "/transformer_blocks.19/norm2/Constant_output_0": { + "/transformer_blocks.19/norm2/Constant_output_0_14_78": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63906,7 +63906,7 @@ "file_name": "cache/pos_embedprojConv_497.const", "file_size": 3072 }, - "/transformer_blocks.19/norm2/Constant_1_output_0": { + "/transformer_blocks.19/norm2/Constant_1_output_0_14_78": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -63932,7 +63932,7 @@ "file_name": "cache/pos_embedprojConv_499.const", "file_size": 5382144 }, - "onnx::MatMul_8402": { + "onnx::MatMul_8402_45_1_77": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63945,7 +63945,7 @@ "file_name": "cache/pos_embedprojConv_500.const", "file_size": 10764288 }, - "onnx::MatMul_8403": { + "onnx::MatMul_8403_45_1_79": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63971,7 +63971,7 @@ "file_name": "cache/pos_embedprojConv_502.const", "file_size": 2691072 }, - "onnx::MatMul_8401": { + "onnx::MatMul_8401_45_5_39": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -63997,7 +63997,7 @@ "file_name": "cache/pos_embedprojConv_504.const", "file_size": 2691072 }, - "/transformer_blocks.19/norm2_context/Constant_output_0": { + "/transformer_blocks.19/norm2_context/Constant_output_0_14_79": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64010,7 +64010,7 @@ "file_name": "cache/pos_embedprojConv_505.const", "file_size": 3072 }, - "/transformer_blocks.19/norm2_context/Constant_1_output_0": { + "/transformer_blocks.19/norm2_context/Constant_1_output_0_14_79": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64036,7 +64036,7 @@ "file_name": "cache/pos_embedprojConv_507.const", "file_size": 5382144 }, - "onnx::MatMul_8404": { + "onnx::MatMul_8404_45_1_78": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64049,7 +64049,7 @@ "file_name": "cache/pos_embedprojConv_508.const", "file_size": 10764288 }, - "onnx::MatMul_8405": { + "onnx::MatMul_8405_45_1_80": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64075,7 +64075,7 @@ "file_name": "cache/pos_embedprojConv_510.const", "file_size": 2691072 }, - "/transformer_blocks.20/norm1/norm/Constant_output_0": { + "/transformer_blocks.20/norm1/norm/Constant_output_0_14_81": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64088,7 +64088,7 @@ "file_name": "cache/pos_embedprojConv_511.const", "file_size": 3072 }, - "/transformer_blocks.20/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.20/norm1/norm/Constant_1_output_0_14_81": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64101,7 +64101,7 @@ "file_name": "cache/pos_embedprojConv_512.const", "file_size": 3072 }, - "/transformer_blocks.20/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.20/norm1_context/norm/Constant_output_0_14_80": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64114,7 +64114,7 @@ "file_name": "cache/pos_embedprojConv_513.const", "file_size": 3072 }, - "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.20/norm1_context/norm/Constant_1_output_0_14_80": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64192,7 +64192,7 @@ "file_name": "cache/pos_embedprojConv_519.const", "file_size": 5382144 }, - "onnx::MatMul_8426": { + "onnx::MatMul_8426_45_5_41": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64205,7 +64205,7 @@ "file_name": "cache/pos_embedprojConv_520.const", "file_size": 2691072 }, - "onnx::MatMul_8425": { + "onnx::MatMul_8425_45_5_40": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64231,7 +64231,7 @@ "file_name": "cache/pos_embedprojConv_522.const", "file_size": 2691072 }, - "/transformer_blocks.20/norm2/Constant_output_0": { + "/transformer_blocks.20/norm2/Constant_output_0_14_82": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64244,7 +64244,7 @@ "file_name": "cache/pos_embedprojConv_523.const", "file_size": 3072 }, - "/transformer_blocks.20/norm2/Constant_1_output_0": { + "/transformer_blocks.20/norm2/Constant_1_output_0_14_82": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64270,7 +64270,7 @@ "file_name": "cache/pos_embedprojConv_525.const", "file_size": 5382144 }, - "onnx::MatMul_8427": { + "onnx::MatMul_8427_45_1_81": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64283,7 +64283,7 @@ "file_name": "cache/pos_embedprojConv_526.const", "file_size": 10764288 }, - "onnx::MatMul_8428": { + "onnx::MatMul_8428_45_1_83": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64322,7 +64322,7 @@ "file_name": "cache/pos_embedprojConv_529.const", "file_size": 2691072 }, - "/transformer_blocks.20/norm2_context/Constant_output_0": { + "/transformer_blocks.20/norm2_context/Constant_output_0_14_83": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64335,7 +64335,7 @@ "file_name": "cache/pos_embedprojConv_530.const", "file_size": 3072 }, - "/transformer_blocks.20/norm2_context/Constant_1_output_0": { + "/transformer_blocks.20/norm2_context/Constant_1_output_0_14_83": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64361,7 +64361,7 @@ "file_name": "cache/pos_embedprojConv_532.const", "file_size": 5382144 }, - "onnx::MatMul_8429": { + "onnx::MatMul_8429_45_1_82": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64374,7 +64374,7 @@ "file_name": "cache/pos_embedprojConv_533.const", "file_size": 10764288 }, - "onnx::MatMul_8430": { + "onnx::MatMul_8430_45_1_84": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64400,7 +64400,7 @@ "file_name": "cache/pos_embedprojConv_535.const", "file_size": 2691072 }, - "/transformer_blocks.21/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.21/norm1_context/norm/Constant_output_0_14_84": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64413,7 +64413,7 @@ "file_name": "cache/pos_embedprojConv_536.const", "file_size": 3072 }, - "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.21/norm1_context/norm/Constant_1_output_0_14_84": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64426,7 +64426,7 @@ "file_name": "cache/pos_embedprojConv_537.const", "file_size": 3072 }, - "/transformer_blocks.21/norm1/norm/Constant_output_0": { + "/transformer_blocks.21/norm1/norm/Constant_output_0_14_85": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64439,7 +64439,7 @@ "file_name": "cache/pos_embedprojConv_538.const", "file_size": 3072 }, - "/transformer_blocks.21/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.21/norm1/norm/Constant_1_output_0_14_85": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64517,7 +64517,7 @@ "file_name": "cache/pos_embedprojConv_544.const", "file_size": 5382144 }, - "onnx::MatMul_8450": { + "onnx::MatMul_8450_45_5_42": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64543,7 +64543,7 @@ "file_name": "cache/pos_embedprojConv_546.const", "file_size": 2691072 }, - "/transformer_blocks.21/norm2/Constant_output_0": { + "/transformer_blocks.21/norm2/Constant_output_0_14_86": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64556,7 +64556,7 @@ "file_name": "cache/pos_embedprojConv_547.const", "file_size": 3072 }, - "/transformer_blocks.21/norm2/Constant_1_output_0": { + "/transformer_blocks.21/norm2/Constant_1_output_0_14_86": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64582,7 +64582,7 @@ "file_name": "cache/pos_embedprojConv_549.const", "file_size": 5382144 }, - "onnx::MatMul_8452": { + "onnx::MatMul_8452_45_1_85": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64595,7 +64595,7 @@ "file_name": "cache/pos_embedprojConv_550.const", "file_size": 10764288 }, - "onnx::MatMul_8453": { + "onnx::MatMul_8453_45_1_87": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64621,7 +64621,7 @@ "file_name": "cache/pos_embedprojConv_552.const", "file_size": 2691072 }, - "onnx::MatMul_8451": { + "onnx::MatMul_8451_45_5_43": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64647,7 +64647,7 @@ "file_name": "cache/pos_embedprojConv_554.const", "file_size": 2691072 }, - "/transformer_blocks.21/norm2_context/Constant_output_0": { + "/transformer_blocks.21/norm2_context/Constant_output_0_14_87": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64660,7 +64660,7 @@ "file_name": "cache/pos_embedprojConv_555.const", "file_size": 3072 }, - "/transformer_blocks.21/norm2_context/Constant_1_output_0": { + "/transformer_blocks.21/norm2_context/Constant_1_output_0_14_87": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64686,7 +64686,7 @@ "file_name": "cache/pos_embedprojConv_557.const", "file_size": 5382144 }, - "onnx::MatMul_8454": { + "onnx::MatMul_8454_45_1_86": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64699,7 +64699,7 @@ "file_name": "cache/pos_embedprojConv_558.const", "file_size": 10764288 }, - "onnx::MatMul_8455": { + "onnx::MatMul_8455_45_1_88": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64725,7 +64725,7 @@ "file_name": "cache/pos_embedprojConv_560.const", "file_size": 2691072 }, - "/transformer_blocks.22/norm1/norm/Constant_output_0": { + "/transformer_blocks.22/norm1/norm/Constant_output_0_14_89": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64738,7 +64738,7 @@ "file_name": "cache/pos_embedprojConv_561.const", "file_size": 3072 }, - "/transformer_blocks.22/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.22/norm1/norm/Constant_1_output_0_14_89": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64764,7 +64764,7 @@ "file_name": "cache/pos_embedprojConv_563.const", "file_size": 5382144 }, - "/transformer_blocks.22/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.22/norm1_context/norm/Constant_output_0_14_88": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64777,7 +64777,7 @@ "file_name": "cache/pos_embedprojConv_564.const", "file_size": 3072 }, - "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.22/norm1_context/norm/Constant_1_output_0_14_88": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64842,7 +64842,7 @@ "file_name": "cache/pos_embedprojConv_569.const", "file_size": 5382144 }, - "onnx::MatMul_8475": { + "onnx::MatMul_8475_45_5_44": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64868,7 +64868,7 @@ "file_name": "cache/pos_embedprojConv_571.const", "file_size": 2691072 }, - "/transformer_blocks.22/norm2/Constant_output_0": { + "/transformer_blocks.22/norm2/Constant_output_0_14_90": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64881,7 +64881,7 @@ "file_name": "cache/pos_embedprojConv_572.const", "file_size": 3072 }, - "/transformer_blocks.22/norm2/Constant_1_output_0": { + "/transformer_blocks.22/norm2/Constant_1_output_0_14_90": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64907,7 +64907,7 @@ "file_name": "cache/pos_embedprojConv_574.const", "file_size": 5382144 }, - "onnx::MatMul_8477": { + "onnx::MatMul_8477_45_1_89": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64920,7 +64920,7 @@ "file_name": "cache/pos_embedprojConv_575.const", "file_size": 10764288 }, - "onnx::MatMul_8478": { + "onnx::MatMul_8478_45_1_91": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64946,7 +64946,7 @@ "file_name": "cache/pos_embedprojConv_577.const", "file_size": 2691072 }, - "onnx::MatMul_8476": { + "onnx::MatMul_8476_45_5_45": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -64972,7 +64972,7 @@ "file_name": "cache/pos_embedprojConv_579.const", "file_size": 2691072 }, - "/transformer_blocks.22/norm2_context/Constant_output_0": { + "/transformer_blocks.22/norm2_context/Constant_output_0_14_91": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -64985,7 +64985,7 @@ "file_name": "cache/pos_embedprojConv_580.const", "file_size": 3072 }, - "/transformer_blocks.22/norm2_context/Constant_1_output_0": { + "/transformer_blocks.22/norm2_context/Constant_1_output_0_14_91": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65011,7 +65011,7 @@ "file_name": "cache/pos_embedprojConv_582.const", "file_size": 5382144 }, - "onnx::MatMul_8479": { + "onnx::MatMul_8479_45_1_90": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -65024,7 +65024,7 @@ "file_name": "cache/pos_embedprojConv_583.const", "file_size": 10764288 }, - "onnx::MatMul_8480": { + "onnx::MatMul_8480_45_1_92": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -65050,7 +65050,7 @@ "file_name": "cache/pos_embedprojConv_585.const", "file_size": 2691072 }, - "/transformer_blocks.23/norm1_context/norm/Constant_output_0": { + "/transformer_blocks.23/norm1_context/norm/Constant_output_0_14_92": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65063,7 +65063,7 @@ "file_name": "cache/pos_embedprojConv_586.const", "file_size": 3072 }, - "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0": { + "/transformer_blocks.23/norm1_context/norm/Constant_1_output_0_14_92": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65076,7 +65076,7 @@ "file_name": "cache/pos_embedprojConv_587.const", "file_size": 3072 }, - "/transformer_blocks.23/norm1/norm/Constant_output_0": { + "/transformer_blocks.23/norm1/norm/Constant_output_0_14_93": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65089,7 +65089,7 @@ "file_name": "cache/pos_embedprojConv_588.const", "file_size": 3072 }, - "/transformer_blocks.23/norm1/norm/Constant_1_output_0": { + "/transformer_blocks.23/norm1/norm/Constant_1_output_0_14_93": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65167,7 +65167,7 @@ "file_name": "cache/pos_embedprojConv_594.const", "file_size": 5382144 }, - "onnx::MatMul_8497": { + "onnx::MatMul_8497_45_5_46": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -65193,7 +65193,7 @@ "file_name": "cache/pos_embedprojConv_596.const", "file_size": 2691072 }, - "/transformer_blocks.23/norm2/Constant_output_0": { + "/transformer_blocks.23/norm2/Constant_output_0_14_94": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65206,7 +65206,7 @@ "file_name": "cache/pos_embedprojConv_597.const", "file_size": 3072 }, - "/transformer_blocks.23/norm2/Constant_1_output_0": { + "/transformer_blocks.23/norm2/Constant_1_output_0_14_94": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65232,7 +65232,7 @@ "file_name": "cache/pos_embedprojConv_599.const", "file_size": 5382144 }, - "onnx::MatMul_8498": { + "onnx::MatMul_8498_45_1_93": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -65245,7 +65245,7 @@ "file_name": "cache/pos_embedprojConv_600.const", "file_size": 10764288 }, - "onnx::MatMul_8499": { + "onnx::MatMul_8499_45_1_94": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -65271,7 +65271,7 @@ "file_name": "cache/pos_embedprojConv_602.const", "file_size": 2691072 }, - "/norm_out/norm/Constant_output_0": { + "/norm_out/norm/Constant_output_0_14_95": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65284,7 +65284,7 @@ "file_name": "cache/pos_embedprojConv_603.const", "file_size": 3072 }, - "/norm_out/norm/Constant_1_output_0": { + "/norm_out/norm/Constant_1_output_0_14_95": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfloat16", @@ -65323,7 +65323,7 @@ "file_name": "cache/pos_embedprojConv_606.const", "file_size": 128 }, - "onnx::MatMul_8500": { + "onnx::MatMul_8500_45_0_4": { "packed_buffer_label": "const", "xrt_arg_id": 3, "dtype": "bfp16ebs8", @@ -65341,83 +65341,203 @@ "dynamic_shape_list": [ { "state_dim1": 1024, - "floor(h/2)": 32, - "batch_size": 2, - "h": 64, - "max_length + state_dim1": 1184, "max_length": 160, "w": 64, - "floor(w/2)": 32 + "h": 64, + "floor(h/2)": 32, + "floor(w/2)": 32, + "batch_4": 2, + "max_length + state_dim1": 1184 }, { - "state_dim1": 1536, - "floor(h/2)": 32, - "batch_size": 2, - "h": 64, - "max_length + state_dim1": 1696, + "state_dim1": 1600, "max_length": 160, - "w": 96, - "floor(w/2)": 48 + "w": 80, + "h": 80, + "floor(h/2)": 40, + "floor(w/2)": 40, + "batch_4": 2, + "max_length + state_dim1": 1760 }, { - "state_dim1": 1536, + "state_dim1": 2304, + "max_length": 160, + "w": 96, + "h": 96, "floor(h/2)": 48, - "batch_size": 2, + "floor(w/2)": 48, + "batch_4": 2, + "max_length + state_dim1": 2464 + }, + { + "state_dim1": 4096, + "max_length": 160, + "w": 128, + "h": 128, + "floor(h/2)": 64, + "floor(w/2)": 64, + "batch_4": 2, + "max_length + state_dim1": 4256 + }, + { + "state_dim1": 768, + "max_length": 160, + "w": 48, + "h": 64, + "floor(h/2)": 32, + "floor(w/2)": 24, + "batch_4": 2, + "max_length + state_dim1": 928 + }, + { + "state_dim1": 1200, + "max_length": 160, + "w": 60, + "h": 80, + "floor(h/2)": 40, + "floor(w/2)": 30, + "batch_4": 2, + "max_length + state_dim1": 1360 + }, + { + "state_dim1": 1728, + "max_length": 160, + "w": 72, "h": 96, - "max_length + state_dim1": 1696, + "floor(h/2)": 48, + "floor(w/2)": 36, + "batch_4": 2, + "max_length + state_dim1": 1888 + }, + { + "state_dim1": 3072, + "max_length": 160, + "w": 96, + "h": 128, + "floor(h/2)": 64, + "floor(w/2)": 48, + "batch_4": 2, + "max_length + state_dim1": 3232 + }, + { + "state_dim1": 768, "max_length": 160, "w": 64, - "floor(w/2)": 32 + "h": 48, + "floor(h/2)": 24, + "floor(w/2)": 32, + "batch_4": 2, + "max_length + state_dim1": 928 }, { - "state_dim1": 2304, - "floor(h/2)": 36, - "batch_size": 2, + "state_dim1": 1200, + "max_length": 160, + "w": 80, + "h": 60, + "floor(h/2)": 30, + "floor(w/2)": 40, + "batch_4": 2, + "max_length + state_dim1": 1360 + }, + { + "state_dim1": 1728, + "max_length": 160, + "w": 96, "h": 72, - "max_length + state_dim1": 2464, + "floor(h/2)": 36, + "floor(w/2)": 48, + "batch_4": 2, + "max_length + state_dim1": 1888 + }, + { + "state_dim1": 3072, "max_length": 160, "w": 128, - "floor(w/2)": 64 + "h": 96, + "floor(h/2)": 48, + "floor(w/2)": 64, + "batch_4": 2, + "max_length + state_dim1": 3232 + }, + { + "state_dim1": 576, + "max_length": 160, + "w": 36, + "h": 64, + "floor(h/2)": 32, + "floor(w/2)": 18, + "batch_4": 2, + "max_length + state_dim1": 736 + }, + { + "state_dim1": 960, + "max_length": 160, + "w": 48, + "h": 80, + "floor(h/2)": 40, + "floor(w/2)": 24, + "batch_4": 2, + "max_length + state_dim1": 1120 + }, + { + "state_dim1": 1344, + "max_length": 160, + "w": 56, + "h": 96, + "floor(h/2)": 48, + "floor(w/2)": 28, + "batch_4": 2, + "max_length + state_dim1": 1504 }, { "state_dim1": 2304, - "floor(h/2)": 64, - "batch_size": 2, - "h": 128, - "max_length + state_dim1": 2464, "max_length": 160, "w": 72, - "floor(w/2)": 36 + "h": 128, + "floor(h/2)": 64, + "floor(w/2)": 36, + "batch_4": 2, + "max_length + state_dim1": 2464 }, { - "state_dim1": 3072, - "floor(h/2)": 48, - "batch_size": 2, - "h": 96, - "max_length + state_dim1": 3232, + "state_dim1": 576, "max_length": 160, - "w": 128, - "floor(w/2)": 64 + "w": 64, + "h": 36, + "floor(h/2)": 18, + "floor(w/2)": 32, + "batch_4": 2, + "max_length + state_dim1": 736 }, { - "state_dim1": 3072, - "floor(h/2)": 64, - "batch_size": 2, - "h": 128, - "max_length + state_dim1": 3232, + "state_dim1": 960, + "max_length": 160, + "w": 80, + "h": 48, + "floor(h/2)": 24, + "floor(w/2)": 40, + "batch_4": 2, + "max_length + state_dim1": 1120 + }, + { + "state_dim1": 1344, "max_length": 160, "w": 96, - "floor(w/2)": 48 + "h": 56, + "floor(h/2)": 28, + "floor(w/2)": 48, + "batch_4": 2, + "max_length + state_dim1": 1504 }, { - "state_dim1": 4096, - "floor(h/2)": 64, - "batch_size": 2, - "h": 128, - "max_length + state_dim1": 4256, + "state_dim1": 2304, "max_length": 160, "w": 128, - "floor(w/2)": 64 + "h": 72, + "floor(h/2)": 36, + "floor(w/2)": 64, + "batch_4": 2, + "max_length + state_dim1": 2464 } ], "aux_info": {} diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle new file mode 100644 index 0000000000000000000000000000000000000000..6a20eff9d876b83d743f7e876f506524e4a12cc4 --- /dev/null +++ b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.ddbundle @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6979e1df0b587d234309798a7fed4fe51ff572dc0a32ddf2273e1c96c4840391 +size 2340803940 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state index c0edb99e01a585da42316f7ee0481bb1cebf51b8..2a141ee22032e5420060b008bb02b7b9bfcc2591 100644 Binary files a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state and b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv.state differ diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.ctrlpkt deleted file mode 100644 index ec8f8395dcb2a6f51d626da94d2381044224d072..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:e2a4fdc1f2614b0614e8610c6107fc61a235f231e33ec434d8516c543d5c359e -size 175875888 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.meta deleted file mode 100644 index 51ed69c2e1b76ca3af379153f8ee2ae9e2257b9e..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a96c479e7ee44c8fc7fe5f67c0cf052473d275e2a714d759a98410834c96c86d -size 26933138 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_128_state_dim1_4096_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.ctrlpkt deleted file mode 100644 index fc8cf16498fc964d19411b3a0a6c1e93ebf5f7bc..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:d27e8a21100793befd71640449ace80abd60cf3f37a0987a975e0e62523a1918 -size 111397296 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.meta deleted file mode 100644 index 16500892ab3f93c5f635fb13fd8aa277217e1ba9..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:dd5a009d238dd9582b02783cf88ecf4afb3fc820dcd933291157def6f424732e -size 16450661 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_72_state_dim1_2304_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.ctrlpkt deleted file mode 100644 index 571ea61284a7553045195eff222303f82b936a23..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a49377c8e7149386ec627bebae1ae32c0218f08853a96f6facd57bee0adc90c8 -size 140762712 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.meta deleted file mode 100644 index 806419d2eb61551ce1e1179c3471178e95eca6f3..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:c993f965da2b2d2d7ceffa904df6b97236bf38639ba52655375c7cd3e1a6eb1b -size 21455934 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_128_h_96_state_dim1_3072_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.ctrlpkt deleted file mode 100644 index 30516586e1d591aaa7b38e23224918b7279949d3..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:3531f32f83aceb62d00ec5d3bc5d524996f58622b2e4cde9c6b44f9db9f9f5f7 -size 69159096 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.meta deleted file mode 100644 index d498602342b6f83d6b8da998c87393515e792b1e..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a049b4a47692482223a9b2a42022b6499c08455854924b2d943ef807c1ea4f8a -size 9487085 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_64_state_dim1_1024_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.ctrlpkt deleted file mode 100644 index 30aa6bde3bb0465ba38bd83f676a1ed079d12e81..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:a382ab957669604427cd6dd3c5ba1f5d084fead7dc592d2c5dd2475530452d37 -size 87129432 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.meta deleted file mode 100644 index a48fd0d09d124bcb34d8993d6a7924651b125d9b..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:8945011b02b8a4d8f4609b44cebceec92f02baf93e0e3c906fdc30e1977dfb6b -size 12299135 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_64_h_96_state_dim1_1536_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.ctrlpkt deleted file mode 100644 index e9086b61f76d8fbd91df94ad287d6aa8f56a496b..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:117b48d0851e520eacc05ee8a979730bbe39c317baccf743acad76718b68d09e -size 111340464 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.meta deleted file mode 100644 index 3068dab0d31a8f1cda670604917e394224238a4c..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:4cb2a588b7c3f50aee80d4c421c8df497becf2cec446dfb70a842c6d6305a6e9 -size 16447712 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_72_h_128_state_dim1_2304_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.ctrlpkt deleted file mode 100644 index fdf3702cdb48be07b4ab131f5af84781fde2ef4b..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:50d6dab11ce89325ad28bcc84b916abd4c3c36c25b6477125cefa984908d8f46 -size 140737656 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.meta deleted file mode 100644 index ecb202157dcea5c86e94bf24a56cf5ba83a7b968..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:68f4db8119c67a6785cba8755b01713c86476c63f7df33b563b8dc5e776d267c -size 21462315 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_128_state_dim1_3072_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.ctrlpkt b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.ctrlpkt deleted file mode 100644 index 07875f31397db7637922c606e6fdd432676f581c..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.ctrlpkt +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:72baa7f100f5c8098fddf1b37da227952a3db3a9043ef34cd4b34f1aee5ebcd6 -size 87129432 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.meta b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.meta deleted file mode 100644 index e8cd0b0a1c5a0bbbc0d219fa5f59ec2e45476c88..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.meta +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:9db54334ae068b1e9cd00f6fafbbf984c82ec8a7a0aac4daec1d3c1817e8e441 -size 12299135 diff --git a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.super b/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.super deleted file mode 100644 index 578425527c883a86ac5c9ab779f4920f04f09a90..0000000000000000000000000000000000000000 --- a/normal/transformer/dynamic/dd/dd_metastate_Sd30_Mmdit_pos_embedprojConv_batch_size_2_w_96_h_64_state_dim1_1536_max_length_160_.super +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:71d091eddadfce78342a8a3e523b021f7e650bec3fd0af7034a255c911d579d1 -size 902984 diff --git a/normal/transformer/dynamic/dd/replaced.onnx b/normal/transformer/dynamic/dd/replaced.onnx index e7b4637c6a1292ff4d1d3886d746c7d9e772d260..442201b6a2dc1bd2a48a4cbfe18cf1d9ad152d6a 100644 --- a/normal/transformer/dynamic/dd/replaced.onnx +++ b/normal/transformer/dynamic/dd/replaced.onnx @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:451c5d977f66191d7c7d21d82d227b48872c51d57d52293bf93f4f15a78c67a6 -size 937072 +oid sha256:5219f90fd0ba7cdaea77bf5a2293bf9e9ae0a6f1b195bb8332b6e430c708c52a +size 920140