Share encoder-only BF16 and validated INT8 ConvRot weights; replace Quanto variants

Qwen3-VL-8B-Instruct/ENCODER_VARIANTS.md

@@ -0,0 +1,7 @@
+# Shared Qwen3-VL-8B-Instruct encoder variants
+
+BF16 and INT8 ConvRot variants accompany the existing FP8 and NF4 files. All use the encoder-only Qwen3VLModel layout: `language_model.*` and `visual.*`, without the unused `lm_head`.
+
+The BF16 encoder was repacked from Qwen/Qwen-Image-2.1 revision b3179ad355be050328e483a9dfdd9e60cd62adfa. All 749 retained tensors were verified bit-for-bit against that source. Against the existing FP8/NF4 encoders, all 381 unquantized tensors match exactly; sampled values from all 368 quantized tensors show quantization-scale differences. The RoPE/config differences are schema changes, not different model dimensions.
+
+INT8 ConvRot was generated with WanGP's shared converter through a headless load-only queue. 360 repeated-block linear weights are quantized, and remaining tensors retain BF16. WanGP Qwen Image 2.1 generation and reference editing were validated with these files. Existing Ideogram FP8/NF4 defaults are unchanged. Source weights remain subject to their upstream license; see https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE.

Qwen3-VL-8B-Instruct/Qwen3-VL-8B-Instruct_bf16.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d0d0b9aa0ef42627800c6d3f25ed8fe82c9dd909f6c3a6cd21609cdc6756f41
+size 16289676320

Qwen3-VL-8B-Instruct/Qwen3-VL-8B-Instruct_int8_convrot.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36e5886b28d08dda11cab85223ea38b475e3e027795ec2c783914ff86f838096
+size 8939652364