Fix ideogram crash: mm: dont reset cast buffers in cleanup_models_gc()

[rattus128]

#14289

cleanup_models_gc can be called once per load_models_gpu via free_memory, which in turn can de-activate an active model via this reset_cast_buffers.

cleanup_models_gc() could also come via obscure garbage collector paths so limit reset_cast_buffers to the post-node callsite instead.

This was affecting low RAM users of ideogram, where the independent load calls meant one model was marked inactivate and could get cause a bad state stealing its own pins.

Example test conditions:
Windows, RTX5060, 16GB RAM
Ideogram4 fp8

Before:

...
  File "C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 321, in cast_bias_weight
    offload_stream = cast_modules_with_vbar([s], dtype, device, bias_dtype, non_blocking)
  File "C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 188, in cast_modules_with_vbar
    handle_pin(s, pin, xfer_source, xfer_dest, size=dest_size)
    ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 186, in handle_pin
    cast_maybe_lowvram_patch(source, pin, offload_stream, xfer_dest2=dest)
    ~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 177, in cast_maybe_lowvram_patch
    comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=stream, r2=xfer_dest2)
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\model_management.py", line 1431, in cast_to_gathered
    dest_views = comfy.memory_management.interpret_gathered_like(tensors, r) if r is not None else [None] * len(tensors)
                 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
  File "C:\Users\rattus\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\memory_management.py", line 162, in interpret_gathered_like
    raise ValueError(f"Buffer too small: needs {offset + size} bytes, but only has {gathered.numel()}. ")
ValueError: Buffer too small: needs 56623104 bytes, but only has 0.

[INFO] Prompt executed in 29.06 seconds

After:

[INFO] Requested to load Ideogram4
[INFO] 0 models unloaded.
[INFO] Model Ideogram4 prepared for dynamic VRAM loading. 8849MB Staged. 0 patches attached. Force pre-loaded 204 weights: 1261 KB.
[INFO] Requested to load Ideogram4
[INFO] 0 models unloaded.
[INFO] Model Ideogram4 prepared for dynamic VRAM loading. 8849MB Staged. 0 patches attached. Force pre-loaded 204 weights: 1261 KB.
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [01:55<00:00,  5.76s/it]
[INFO] Requested to load AutoencoderKL
[INFO] 0 models unloaded.
[INFO] Model AutoencoderKL prepared for dynamic VRAM loading. 160MB Staged. 0 patches attached. Force pre-loaded 110 weights: 187 KB.
[INFO] Prompt executed in 134.77 seconds

Regression tests:

Linux, 5090, 96GB, ideogram4 ✅
Linux, 5090, 96GB, stable cascade ->Flux2 ✅
Linux, 5090, 96GB, ace step 1.5 XL ✅
Linux, 5090, 96GB, LTX2.3 ✅
Linux, 5090, 96GB, Wan 2.2 2x14B ✅
Linux, 5090, 96GB, qwen ✅
Windows, 5060, 16GB, LTX2.3 ✅

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 427e71e4-4fa7-4072-8269-ab621e96868f

📥 Commits

Reviewing files that changed from the base of the PR and between 95cb7df and 46a28f9.

📒 Files selected for processing (1)

• comfy/model_management.py

🚧 Files skipped from review as they are similar to previous changes (1)

• comfy/model_management.py

---

📝 Walkthrough

Walkthrough

This PR removes a call to reset_cast_buffers() at the start of cleanup_models_gc in comfy/model_management.py. The function now immediately scans current_loaded_models for dead models and only calls gc.collect()/soft_empty_cache() when dead models are detected; no other logic was changed.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically describes the main change: removing reset_cast_buffers() call from cleanup_models_gc() to fix an Ideogram crash issue.
Description check	✅ Passed	The description is directly related to the changeset, explaining the motivation behind removing the reset_cast_buffers() call and providing detailed evidence of the fix working.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}