Pexqman/Wan2.2-TI2V-5B-hexpack

Wan 2.2 TI2V-5B for the Snapdragon X NPU (PulseX Wan)

The run files of PulseX Wan: Wan-AI's Wan 2.2 TI2V-5B (text to video and image to video) converted to run on the Hexagon NPU of a Snapdragon X laptop - the transformer in Q4_0 (ffn.2 in Q8_0) repacked in the NPU's own layout, and the VAE as compiled QNN graphs. About 30 minutes for a 720p 5 s video on a Snapdragon X Plus with 16 GB. See the program's README for how to install and run it.

Not here (take them unchanged from their own repositories): the umT5-XXL text encoder (city96/umt5-xxl-encoder-gguf, umt5-xxl-encoder-Q8_0.gguf) and spiece.model (Wan-AI/Wan2.2-TI2V-5B, google/umt5-xxl/spiece.model).

The QNN context binaries are compiled for the Hexagon v73 NPU (Snapdragon X Plus X1P-42-100); other chips are untested. They are run with the Qualcomm AI Runtime 2.45.0 (build 2.45.0.260326154327, all five QNN files from that one build - the setup file of PulseX Wan ships exactly that set); vae_npu_step_w8a8/golden_y2.npy was recorded with it.

License

Apache License 2.0 - see LICENSE.txt (Wan 2.2's own) and NOTICE: these files are modified derivatives of Wan 2.2 TI2V-5B (Wan-AI); taehv/ is TAEHV by madebyollin under the MIT License. Not affiliated with or endorsed by Wan-AI, Alibaba or Qualcomm.

Files (SHA-256)

File	What	Bytes	SHA-256
Wan2.2-TI2V-5B-Q4_0-ffn2q8-ggml-slim.gguf	transformer: slim GGUF (metadata + the few tensors the CPU needs)	182462208	87eb793ec2a14d1f8506d1ddff7874cc3cbac03ed49845b7433564d54eae6d1b
Wan2.2-TI2V-5B-Q4_0-ffn2q8-ggml-slim.hexpack	transformer: the 30 blocks in the NPU's own layout	3428376576	4e2576d5103dfc1bf8917e4071bed497bc7ee71e9abfed3b97dd09eb02b7be73
Wan2.2-TI2V-5B-npu-extras.gguf	text / time embeddings, patch embedding and head	182461856	cfba2c11aa8722802380464c8519384d99d6afe9eeeba7ed851324b1d2d3b27d
Wan2.2_VAE-Q4_0.gguf	the Wan 2.2 VAE, Q4_0	398514304	12b3d533a70160c0f27d4152a43debb09895eee3c2acce6e935ff1e4cc74b1e1
Wan2.2_VAE-Q4_0.post_quant_conv.npz	the VAE's post_quant_conv, cached (keyed to the VAE file's content)	10444	12f835266cc445205fd29b69bc03dab12f7e4e68e1757e9417aedd95907cbc6b
Wan2.2_VAE-Q4_0.npu_golden_tile.npz	a fixed control tile the NPU decoder is checked against at every run	7152892	247e5d5669e4fc3a9a1c8c2eb0b80708c36371323cd2f3647ccab80726009c80
vae_config/config.json	the VAE's configuration (latent mean / std)	1201	0cadc3ee1bb700c6c73de51738dc0f112c542901f2fc93a98f6bd9020e324618
vae_npu_first_p3/model.onnx	Sharp decoder: frames 0-1, QNN context (ONNX wrapper)	3388	fa817c74b9751c93b2c179ac2492cd473386b89994cc7a477bc9d99c883f9059
vae_npu_first_p3/model.bin	Sharp decoder: frames 0-1, QNN context binary	1165393920	5e261fd79e73f9719a04b3a35e3727b244e22d64918dce4a5a765c65cf4a7917
vae_npu_step_w8a8/model.onnx	Sharp decoder: one frame step, w8a8 QNN context (ONNX wrapper)	3134	b8ce64f8e6eb523b51333c67c7d436e5b958a053b15a076f2e5a468f94d866aa
vae_npu_step_w8a8/model_qnn.bin	Sharp decoder: one frame step, w8a8 QNN context binary	566190080	53acf01607e266874014f9a32efbaeaeefa5e19369447c0e7510c1b6ce1f233f
vae_npu_step_w8a8/golden_y2.npy	the step graph's own recorded output for the control tile	3145856	27521ea3e825e07672e8f9e831e446805fa3b9fc08bf542b99c0baaa70a36100
vae_npu_enc/model.onnx	picture encoder 832x480 (image to video), ONNX wrapper	351	376091c9ccaa48ad72dffcb2be24423a687bf07fb0dde1cef54e60a8ca7b843f
vae_npu_enc/model.bin	picture encoder 832x480, QNN context binary	305913856	ccaa3424f78254860b209b097ecc68e30bf1d0778b0cc42b5df642f4430a702e
vae_npu_enc_704x1280/model.onnx	picture encoder 1280x704, ONNX wrapper	351	7df0be7b98bc4d01ffb0783745bf708c36931f81c6e49822b160dc48bc466d86
vae_npu_enc_704x1280/model.bin	picture encoder 1280x704, QNN context binary	330588160	6fd3b616a91802ca8ac4ab21f5acaaf2d10d7981f69076a083afabb81515ee2f
taew_npu_ctx/model.onnx	Fast decoder (TAEHV) on the NPU, 480p 49 frames, ONNX wrapper	348	2074976bdf20178d362e4b28bfe2ec5aed98349f56894673e48300dc608eb555
taew_npu_ctx/model.bin	Fast decoder (TAEHV), QNN context binary	43798528	a60fe925303ec044e3b2aee6e2e93b1d6eaf8e9c3e9cfe49ac95886a919dccbd
taehv/taew2_2_super.pth	TAEHV weights for Wan 2.2 (madebyollin/taehv, MIT)	41011171	60ae3d27df47b1b7380a165745571f34df73103e153b89c0e74e9381b266b7f6
taehv/LICENSE	TAEHV's MIT license	1073	532f9e394518ffddecd294a517d5b41d79d3d3866c3fb95a6cb0e8bcc02370bf