[Bug] WAN 2.2 tensor naming weirdness.

[lastrosade]

Git commit

6b675a5

Operating System & Version

Windows 11

GGML backends

HIP

Command-line arguments used

$models = "a:\AI_MODELS\Stable Diffusion\Models"
& 'a:\AI_MODELS\Stable Diffusion\stable-diffusion.cpp.hip.bin\sd-cli.exe' -M vid_gen `
--diffusion-model "$($models)\checkpoints\wan\Remix_v3_low.q5_K.gguf" `
--high-noise-diffusion-model "$($models)\checkpoints\wan\Remix_v3_high.q5_K.gguf" `
--t5xxl "$($models)\text_encoders\NSFW_WAN_umt5-xxl-encoder.q8_0.gguf" `
--tae "$($models)\tae\taew2_1.safetensors" `
--lora-model-dir "$($models)\LoRa\wan" `
-n "" --seed 0 --offload-to-cpu --clip-on-cpu --fa --diffusion-fa --vae-tiling `
--sampling-method euler --cfg-scale 1 --steps 4 --scheduler simple `
--high-noise-sampling-method euler --high-noise-cfg-scale 1 `
--flow-shift 3.0 -W 432 -H 256 --fps 15 --video-frames 61 `
-i "a:\AI_MODELS\Stable Diffusion\stable-diffusion.cpp\assets\with_lcm.png" `
-p "a lovely cat" --verbose

Issue

Using the WAN 2.2 Remix Models

There seems to be an issue with how the tensor names are parsed.
While the low noise model loads correctly, the high noise model does not.

[INFO ] model.cpp:1614 - unknown tensor 'model.high_noise_diffusion_model.model.diffusion_model.blocks.0.cross_attn.norm_q.weight | q5_K | 1 [5120, 1, 1, 1, 1]' in model file
[ERROR] model.cpp:1655 - tensor 'model.high_noise_diffusion_model.blocks.0.cross_attn.norm_q.weight' not in model file

I will spare you the full list.

fulllog.txt

[lastrosade]

I have fixed the issue by stripping model.diffusion_model. out of the tensor names.

Now the model loads correctly, not that it runs, it's wan on hip ( high noise sampling at 478.91s/it, it took 5h for a 61 frames clip and it looks like fog :) ).

[stduhpf]

I think loading it with --model instead of --diffusion-model would have fixed it