hires with a model-backed upscaler aborts on Qwen-Image-2.1 (GGML_ASSERT ne[2] in ggml_im2col)

[nbeerbower]

What happens

With Qwen-Image-2.1 loaded, an img_gen request whose hires.upscaler is a
model-backed upscaler (ESRGAN) aborts the process:

[INFO ] image.cpp:913  - hires fix: upscaling to 1024x1024
[INFO ] image.cpp:921  - hires fix: loading model upscaler from '.../RealESRGAN_x4plus_anime_6B.pth'
[INFO ] image.cpp:945  - hires fix: scheduler_steps=4, denoising_strength=0.50, sigma_sched_size=4
ggml/src/ggml.c:4557: GGML_ASSERT(a->ne[2] == b->ne[2]) failed

That assertion is the is_2D branch of ggml_im2col. The server does not
return an error, it dies, so every other queued job goes with it.

A latent upscaler on the same model is fine, which is what makes this look
like a shape mismatch rather than a bad model file.

Reproduction

Same server, same request, only hires.upscaler differs.

Server:

sd-server --diffusion-model qwen_image_2.1-Q8_0.gguf \
          --vae qwen_image_2.1_vae_bf16.safetensors \
          --llm Qwen3VL-8B-Instruct-Q8_0.gguf \
          --llm_vision mmproj-Qwen3VL-8B-Instruct-F16.gguf \
          --hires-upscalers-dir <dir with RealESRGAN_x4plus_anime_6B.pth> \
          --diffusion-fa

Works — completes in about 7 s:

{"prompt":"a cat","width":512,"height":512,"seed":1,
 "sample_params":{"sample_method":"euler","sample_steps":4},
 "hires":{"enabled":true,"upscaler":"Latent","scale":2.0,
          "steps":2,"denoising_strength":0.5}}

Aborts the server:

{"prompt":"a cat","width":512,"height":512,"seed":1,
 "sample_params":{"sample_method":"euler","sample_steps":4},
 "hires":{"enabled":true,"upscaler":"RealESRGAN_x4plus_anime_6B","scale":2.0,
          "steps":2,"denoising_strength":0.5}}

Guess at the cause

I have not fixed this, so this is only where I would start looking. The two
paths diverge at image.cpp:708: a model-backed upscaler decodes the latent
to an image, runs the upscaler, and re-encodes, where a latent upscaler stays
in latent space. Qwen Image 2.1 carries qwen_image_layers + 1 in the layer
dimension (image.cpp:221, generate_init_latent(..., request->qwen_image_layers + 1, true),
default 3, so 4), and a latent re-encoded from a plain RGB image would have 1
there. That would fit an ne[2] mismatch reaching ggml_im2col.

Environment

• master at 6dcb5bb, CUDA build, RTX A6000, Linux
• Qwen-Image-2.1 Q8_0 (leejet/Qwen-Image-2.1-GGUF) with
  qwen_image_2.1_vae_bf16.safetensors and Qwen3-VL-8B-Instruct Q8_0
• RealESRGAN_x4plus_anime_6B.pth

Standalone upscaling with the same model file is fine (sd-cli -M upscale,
512x512 to 2048x2048 in ~2 s), so the upscaler itself loads and runs.