Load mmproj vision tower for qwen3vl text encoders (fixes Qwen-Image-2.1 12288 error)

[pottokao-dotcom]

Using a Qwen3-VL GGUF as the Qwen-Image-2.1 text encoder (CLIPLoaderGGUF, type qwen_image → TextEncodeQwenImage21) fails in the sampler with:

RuntimeError: Given normalized_shape=[4096], expected input with shape [*, 4096], but got input of size [1, 512, 12288]

Cause

gguf_clip_loader only loads the mmproj vision tower for qwen2vl. For qwen3vl the vision tower is missing, so ComfyUI doesn't detect the state dict as Qwen3-VL (it looks for model.visual.deepstack_merger_list.*) and builds the wrong text encoder, which returns 12288-wide (FFN) hidden states.

Fix

• Load the mmproj for qwen3vl too.
• Rename the tensors gguf_mmproj_loader returns to ComfyUI's Qwen3-VL layout (model.visual.*): attn_qkv → attn.qkv, up/down_proj → linear_fc1/fc2, v.deepstack.{8,16,24} → deepstack_merger_list.{0,1,2}, merger ln_q/mlp.0/mlp.2 → norm/linear_fc1/linear_fc2, v.position_embd → pos_embed.

qwen2vl is unchanged.

Tested

ComfyUI 0.36.0 with a Qwen3-VL-8B Q4_K_M GGUF + f16 mmproj (llama.cpp conversion):

• The loaded state dict has the same 750 keys and shapes as Comfy-Org's bf16 safetensors text encoder; the 351 vision tensors match it numerically (rel. error ≤ 1.1e-7).
• Qwen-Image-2.1 text-to-image and reference-image editing both run, and match the bf16 encoder's output for the same seed up to Q4 noise.

🤖 Generated with Claude Code