mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-05-16 03:57:27 +08:00
61235fc35a
Improves on the previous directml commit with three research-based refinements: 1. model_management.py — module_mmap_residency() and cast_to_gathered() Replace broad try/except NotImplementedError with an explicit `t.device.type == 'privateuseone'` guard. Checking device type is faster in a hot loop and makes the intent self-documenting. Fixes: github.com/Comfy-Org/ComfyUI/issues/8347 2. attention.py — attention_split() Replace the "assume 4 GB free" heuristic with `steps = 64`. 64-slice chunking is safe and correct regardless of card size; the 4 GB assumption was fragile on cards with less or more VRAM. 3. diffusionmodules/model.py — slice_attention() Apply the identical `steps = 64` guard to the second call site for the same ZeroDivisionError (was missed in the previous commit). Fixes: github.com/comfyanonymous/ComfyUI/issues/1518 Tested end-to-end on AMD RX 5600 XT (6 GB VRAM), Windows 11, torch-directml 0.2.5, ComfyUI 0.21.1, DreamShaper 8 (SD 1.5). Full 20-step txt2img pipeline completes and returns a valid PNG. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| diffusionmodules | ||
| distributions | ||
| encoders | ||
| attention.py | ||
| ema.py | ||
| sdpose.py | ||
| sub_quadratic_attention.py | ||
| temporal_ae.py | ||