From e912b910a2beaaa7cfeb6c991ea327e586892083 Mon Sep 17 00:00:00 2001
From: Houde <LvHHuuaadai@proton.me>
Date: Sat, 20 Jun 2026 19:03:20 +0100
Subject: [PATCH] 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).
---
 baselines/system_info_realvisxl.txt | 55 +++++++++++++++++++++++++++++
 comfy/utils.py                      | 32 +++++++++++++++++
 2 files changed, 87 insertions(+)
 create mode 100644 baselines/system_info_realvisxl.txt

diff --git a/baselines/system_info_realvisxl.txt b/baselines/system_info_realvisxl.txt
new file mode 100644
index 000000000..8a6d3b839
--- /dev/null
+++ b/baselines/system_info_realvisxl.txt
@@ -0,0 +1,55 @@
+=== 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
diff --git a/comfy/utils.py b/comfy/utils.py
index 09d783fff..ceddf7af1 100644
--- a/comfy/utils.py
+++ b/comfy/utils.py
@@ -119,6 +119,32 @@ def load_safetensors(ckpt):
     return sd, header.get("__metadata__", {}),
 
 
+_LARGE_FILE_MMAP_THRESHOLD = 4_000_000_000  # 4 GB
+
+
+def _load_safetensors_no_mmap(ckpt):
+    # Windows + ROCm/CUDA UMA: large mmaps fail after GPU virtual address space is reserved.
+    # Read tensors sequentially from file instead.
+    sd = {}
+    with open(ckpt, "rb") as fh:
+        header_len = struct.unpack("<Q", fh.read(8))[0]
+        header = json.loads(fh.read(header_len).decode("utf-8"))
+        data_start = 8 + header_len
+        for name, info in header.items():
+            if name == "__metadata__":
+                continue
+            start, end = info["data_offsets"]
+            dtype = _TYPES[info["dtype"]]
+            shape = info["shape"]
+            fh.seek(data_start + start)
+            raw = fh.read(end - start)
+            if raw:
+                sd[name] = torch.frombuffer(bytearray(raw), dtype=dtype).reshape(shape).clone()
+            else:
+                sd[name] = torch.empty(shape, dtype=dtype)
+    return sd, header.get("__metadata__", {})
+
+
 def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     if device is None:
         device = torch.device("cpu")
@@ -129,6 +155,12 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
                 sd, metadata = load_safetensors(ckpt)
                 if not return_metadata:
                     metadata = None
+            elif os.path.getsize(ckpt) > _LARGE_FILE_MMAP_THRESHOLD and torch.cuda.is_available():
+                # File > 4 GB with active CUDA/ROCm: mmap would exhaust Windows virtual
+                # address space reserved by UMA GPU init. Use sequential file-read instead.
+                sd, metadata = _load_safetensors_no_mmap(ckpt)
+                if not return_metadata:
+                    metadata = None
             else:
                 with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
                     sd = {}