Fix model path in Qwen-Image-Distill-DMD2 examples

[rakhimovv]

Thanks for this project — I've been working through the examples/qwen_image/ scripts and hit a few small things; this is the most clear-cut one.

Both Qwen-Image-Distill-DMD2.py examples download the distill LoRA into one directory layout and read it back from another, so the load fails.

model_inference/Qwen-Image-Distill-DMD2.py:18-19:

snapshot_download("MusePublic/Qwen-Image-Distill", allow_file_pattern="qwen_image_distill_3step.safetensors", cache_dir="models")
lora_state_dict = load_state_dict("models/MusePublic/Qwen-Image-Distill/qwen_image_distill_3step.safetensors")

With modelscope 1.39.1, cache_dir= writes the hub cache layout, so the file lands at:

models/models/MusePublic--Qwen-Image-Distill/snapshots/master/qwen_image_distill_3step.safetensors

while the next line reads the local_dir= layout, models/MusePublic/Qwen-Image-Distill/.... The download reports success and the read then fails:

Downloading: 100%|██████████| 1/1 [00:03<00:00,  3.52s/file]
Traceback (most recent call last):
  File ".../Qwen-Image-Distill-DMD2.py", line 19, in <module>
    lora_state_dict = load_state_dict("models/MusePublic/Qwen-Image-Distill/qwen_image_distill_3step.safetensors")
FileNotFoundError: No such file or directory: models/MusePublic/Qwen-Image-Distill/qwen_image_distill_3step.safetensors

I first hit this during a CDN timeout and assumed that was the cause, but a clean, fully successful download fails the same way. Since the cache layout has changed across modelscope versions, I'd guess this worked when it was written.

These are the only two call sites in the repo using cache_dir=: of 321 snapshot_download(...) calls under examples/ at 4dbf980, 319 pass local_dir= and these 2 pass cache_dir=. The patch switches them to local_dir=, matching the sibling Qwen-Image-Distill-LoRA.py:5.

Checking

Both scripts fail at 4dbf980 and produce an image with this patch, on 1×H100 downloading from ModelScope:

• model_inference/Qwen-Image-Distill-DMD2.py — 53s
• model_inference_low_vram/Qwen-Image-Distill-DMD2.py — 69s