Quality seems a little worse than the Lightning model, with ~30% less peak VRAM usage, and similar speed gains.

@leejet , looks likea123e25 from the qwen_image_edit branch is enough to support the '13b' pruned model. Thanks!

The '12b' variant still doesn't work, though, maybe because it has non-contiguous layers. I guess they're keeping the non-pruned layers with the same number as they have on the original model.

@wbruna I think it might be a problem with the GGUF quant, not the model. Look at the GGUFhttps://huggingface.co/wsbagnsv1/Qwen-Image-Pruning-GGUF/tree/main?show_file_info=Qwen-Image-Pruning-12b-Q4_0.gguf Transformer block 18,

versus the originalhttps://huggingface.co/OPPOer/Qwen-Image-Pruning/tree/main/Qwen-Image-12B/transformer?show_file_info=Qwen-Image-12B%2Ftransformer%2Fdiffusion_pytorch_model-00002-of-00003.safetensors

I've reported it to the GGUF quant guy but he doesn't seem to get what i meanhttps://huggingface.co/wsbagnsv1/Qwen-Image-Pruning-GGUF/discussions/1

@leejet , looks likea123e25 from the qwen_image_edit branch is enough to support the '13b' pruned model. Thanks!

Not working anymore afterd21d1aa (didn't check which revision). I'll retest and update this PR.

What do you mean "Not working anymore", is it still generating an image? It seems to work for me

What do you mean "Not working anymore", is it still generating an image? It seems to work for me

This PR still works; I was referring to my previous comment, which mentioned thata123e25 (from theqwen_image branch itself) made this PR unnecessary.

The support for dynamic number of Qwen image transformer blocks is available in the qwen_image_edit branch.

The support for dynamic number of Qwen image transformer blocks is available in the qwen_image_edit branch.

I tested it with the same model I use to test this PR (the 13b, they renamed it afterwards):

[DEBUG] model.cpp:2088 - loading tensors from /opt/sdif/models/SD/Qwen-Image-Pruning-Q4_K_M.gguf  |===================================>              | 1293/1825 - 45.33it/s[DEBUG] model.cpp:2088 - loading tensors from /opt/llm/Qwen2.5-VL-7B-Instruct-IQ4_XS.gguf  |============================================>     | 1631/1825 - 43.55it/s[DEBUG] model.cpp:2088 - loading tensors from /opt/sdif/models/VAE/Qwen_Image-VAE.safetensors  |============================================>     | 1634/1825 - 43.63it/s[INFO ] model.cpp:2358 - unknown tensor 'first_stage_model.conv1.bias | bf16 | 1 [32, 1, 1, 1, 1]' in model file[INFO ] model.cpp:2358 - unknown tensor 'first_stage_model.conv1.weight | bf16 | 4 [1, 1, 1, 1024, 1]' in model file  |==================================================| 1825/1825 - 48.21it/s[INFO ] model.cpp:2326 - loading tensors completed, taking 37.90s (process: 0.05s, read: 24.94s, memcpy: 0.00s, convert: 0.47s, copy_to_backend: 12.08s)[ERROR] model.cpp:2399 - tensor 'model.diffusion_model.transformer_blocks.40.attn.add_k_proj.bias' not in model file[ERROR] model.cpp:2399 - tensor 'model.diffusion_model.transformer_blocks.40.attn.add_k_proj.weight' not in model file[ERROR] model.cpp:2399 - tensor 'model.diffusion_model.transformer_blocks.40.attn.add_q_proj.bias' not in model file[ERROR] model.cpp:2399 - tensor 'model.diffusion_model.transformer_blocks.40.attn.add_q_proj.weight' not in model file[ERROR] model.cpp:2399 - tensor 'model.diffusion_model.transformer_blocks.40.attn.add_v_proj.bias' not in model file(...)

Please use the latest code from this PR#877.

Oh, I see: I forgot that rev was on theqwen_image_edit branch 🤦

@LostRuins , btw: if you're already syncing with that one, feel free to drop my changes, they shouldn't be necessary.

I noticed that this PR was automatically closed after the qwen_image branch was deleted.
If the related changes are still relevant, please feel free to reopen it for further review.

I noticed that this PR was automatically closed after the qwen_image branch was deleted. If the related changes are still relevant, please feel free to reopen it for further review.

Thanks. I was keeping this open just to further investigate about the 12b variant, but I could just open another one instead.