[BUG] "Could not detect model type"

[eliteprox]

Describe the bug

ComfyStream server memory leak during failed workflow

To Reproduce

After submitting an invalid workflow, usually where a model is missing, the server does not recover and cannot process further streams.

1. Download a workflow like sd15_multi_cnet_depthanything_face.json.
2. Modify the engine name for depthanything to be the wrong name.
3. Submit the broken workflow to comfystream.
4. The SD 1.5 models will be loaded, but the workflow will fail to run
5. Submit a correct version of the workflow sd15_multi_cnet_depthanything_face.json to ComfyStream UI.
6. The SD 1.5 model is unable to load (presumably twice, due to still being loaded from failed workflow start), and the following error occurs:

  File "/workspace/miniconda3/envs/comfystream/lib/python3.11/site-packages/comfy/sd.py", line 892, in load_checkpoint_guess_config
    raise RuntimeError("Could not detect model type of: {}".format(ckpt_path))
RuntimeError: Could not detect model type of: /workspace/ComfyUI/models/checkpoints/dreamshaper_8.safetensors

• Restarting the Pod via the Runpod UI always works (when I load the workflow for the first time after restarting)
• Reloading the same workflow (default SD1.5 workflow) sometime works, sometime doesn't work

Expected behavior

ComfyStream server should free cuda memory, free memory and restart the appropriate thread/process queue in EmbeddedComfyClient, then return a 500 error response

Screenshots

Desktop (please complete the following information)

No response

Smartphone (please complete the following information)

No response

Additional context

No response

[eliteprox]

This issue is ready to be closed, will give some time to verify downstream resolution in staging