mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-07-03 13:19:23 +08:00
e912b910a2
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).
56 lines
1.9 KiB
Plaintext
56 lines
1.9 KiB
Plaintext
=== ComfyUI RealVisXL no-mmap Stable Baseline ===
|
|
Date: 2026-06-20
|
|
|
|
--- Previous baseline ---
|
|
Tag: rocm-sd15-working-baseline (preserved, not modified)
|
|
|
|
--- This baseline adds ---
|
|
Fix: comfy/utils.py: _load_safetensors_no_mmap() for files >4 GB
|
|
Model: RealVisXL_V4.0.safetensors (6.46 GB) - path only, not in git
|
|
Test: 768x1024, 25 steps, cfg=6, dpmpp_2m, karras -> OK
|
|
|
|
--- Root cause of crash (diagnosed & fixed) ---
|
|
Strix Halo UMA: ROCm init reserves ~14 GB GPU virtual address space.
|
|
safetensors mmap of files >~4 GB then fails (Windows VA space exhausted).
|
|
SD1.5 (3.97 GB) < threshold -> mmap OK.
|
|
SDXL fp16 (~6.5 GB) > threshold -> access violation in safe_open().
|
|
Fix: sequential file-read (open+seek+read) bypasses mmap entirely.
|
|
|
|
--- Patch location ---
|
|
File: comfy/utils.py
|
|
Functions: _load_safetensors_no_mmap(), _LARGE_FILE_MMAP_THRESHOLD = 4_000_000_000
|
|
Branch: load_torch_file() elif os.path.getsize > threshold and cuda available
|
|
|
|
--- Startup command ---
|
|
cd C:\Users\LvHHu\ComfyUI
|
|
.\venv\Scripts\activate
|
|
python main.py --disable-dynamic-vram --disable-mmap
|
|
|
|
--- GPU / ROCm ---
|
|
torch: 2.7.0a0+git3f903c3
|
|
Device: AMD Radeon(TM) 8050S Graphics
|
|
VRAM GB: 14.37
|
|
ROCm: 6.5 / gfx1151 (Strix Halo)
|
|
|
|
--- Models in checkpoints (not in git) ---
|
|
v1-5-pruned-emaonly.safetensors 3.97 GB SD1.5 baseline
|
|
RealVisXL_V4.0.safetensors 6.46 GB SDXL realistic portrait
|
|
|
|
--- Working parameters (RealVisXL) ---
|
|
Resolution: 768x1024
|
|
Steps: 25
|
|
CFG: 6
|
|
Sampler: dpmpp_2m
|
|
Scheduler: karras
|
|
Batch size: 1
|
|
|
|
--- Recovery commands ---
|
|
# 1. Return to this code state:
|
|
git checkout rocm-realvisxl-nommap-working
|
|
|
|
# 2. Re-download RealVisXL if needed (not in git):
|
|
# python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='SG161222/RealVisXL_V4.0', filename='RealVisXL_V4.0.safetensors', local_dir='models/checkpoints')"
|
|
|
|
# 3. Start ComfyUI:
|
|
# python main.py --disable-dynamic-vram --disable-mmap
|