2 Commits

Author	SHA1	Message	Date
Houde	0df0b0d613	fix: address CodeRabbit review comments on PR #14587 ... - utils.py: add device param to _load_safetensors_no_mmap, move tensors to target device instead of always returning CPU tensors - utils.py: validate read length == expected bytes; raise RuntimeError on partial/corrupt reads instead of silently creating empty tensors - utils.py: scope no-mmap fallback to sys.platform == win32 to avoid unnecessary overhead on Linux/Mac CUDA systems; add sys import - baselines: replace hardcoded LvHHu username with %USERPROFILE% in startup commands for portability Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-06-23 14:51:22 +01:00
Houde	e912b910a2	fix: add no-mmap safetensors loader for >4GB files on Windows ROCm/UMA ... Root cause: Strix Halo UMA ROCm init reserves ~14 GB of Windows virtual address space for GPU. This prevents safetensors from mmap-ing files larger than ~4 GB (SDXL fp16 ~6.5 GB), causing access violations. SD1.5 (3.97 GB) is below the threshold and unaffected. Fix in comfy/utils.py: - Add _LARGE_FILE_MMAP_THRESHOLD = 4_000_000_000 - Add _load_safetensors_no_mmap(): reads tensors via open()+seek()+read() instead of mmap, then clones each tensor for independent ownership - In load_torch_file(): route files >4 GB with CUDA active through _load_safetensors_no_mmap() automatically Tested: RealVisXL_V4.0.safetensors (6.46 GB) loads and generates 768x1024 portrait images at ~5 it/s on AMD Radeon 8050S (gfx1151). SD1.5 baseline unaffected (still uses original mmap path).	2026-06-22 18:52:42 +01:00