nodes/pollockjj-ComfyUI-MultiGPU
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

feat: patch load_models_gpu for accurate memory calculations; unpatch load_models_gpu

Browse Source 
Refactor memory management in distorch_2.py to patch load_models_gpu instead of LoadedModel.model_memory_required. Implement correct memory reporting based on model flags (eject_models and is_distorch_model), ensuring proper eviction logic and improved handling of virtual VRAM. This drives behavior purely by either comfy core matching or DisTorch flag, fixing potential issues in multi-GPU setups.
This commit is contained in:
John Pollock
2025-10-04 17:16:51 -05:00
parent e6d19951d7
commit 72a20338ef
19 changed files with 158 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files
web/docs/TripleCLIPLoaderDisTorch2MultiGPU.md
+1 -1
View File
@@ -15,7 +15,7 @@ This node automatically detects models located in the `ComfyUI/models/clip` fold
| `virtual_vram_gb` | `FLOAT` | Amount of virtual VRAM in gigabytes to allocate for distributed tensor management (default: 4.0, range: 0.0-128.0). |
| `donor_device` | `STRING` | Device to donate VRAM from when allocating virtual memory (default: 'cpu'). |
| `expert_mode_allocations` | `STRING` | Advanced allocation string for expert users to manually specify device/ratio distributions (e.g., 'cuda:0,50%;cpu,*'). |
| `keep_loaded` | `BOOLEAN` | Whether to keep the model loaded when triggering memory cleanup operations (default: true). |
| `eject_models` | `BOOLEAN` | Whether to unload ALL models from the target device before loading this model, enabling deterministic model eviction for testing and memory management (default: false for CLIP loaders). |
## Outputs