update: 549 fp8 (Linear-only), Conv weights stay bf16

README.md

@@ -91,8 +91,9 @@ native names, run a key remap (see `diffusers_to_native_wan` in
 
 ## What's in the file
 
-- **551 FP8 tensors** (14.5 GB): the big linear weights - self-attn Q/K/V/O,
-  FFN projections, controlnet linears
+- **549 FP8 tensors** (14.5 GB): rank-2 Linear weights only (self-attn
+  Q/K/V/O, FFN projections, controlnet linears). Conv weights stay BF16:
+  no scaled-fp8 Conv kernel exists in PyTorch / ComfyUI
 - **1800 BF16 tensors** (5.84 GB):
   - 362 `norm` tensors (RMSNorm gains): 1.26 GB
   - 80 `add_*` tensors (I2V cross-attention to image embeddings): 4.19 GB