[BUG] SDXL KSampler produces pure noise on RTX 4060 Ti — reproduced on brand-new, zero-custom-node install (v0.29.0), full elimination chain included
Potential Bug
### Custom Node Testing
- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
### Expected Behavior
A minimal SDXL txt2img workflow (no ControlNet, no LoRA, no custom nodes — just CheckpointLoaderSimple → CLIPTextEncode ×2 → EmptyLatentImage → KSampler → VAEDecode → SaveImage) should produce a coherent image matching the prompt.
### Actual Behavior
KSampler completes with zero exceptions, but the decoded output is pure RGB noise — not a malformed or artifacted image, but noise indistinguishable from random static. This reproduces 100% of the time, including on a brand-new install with no custom nodes whatsoever.
<img width="1024" height="1024" alt="Image" src="https://github.com/user-attachments/assets/eeedbcf1-2582-4463-b76c-7dc95ab71f8c" />
### Steps to Reproduce
1. Fresh install of ComfyUI v0.29.0 portable (official Windows build), zero custom nodes.
2. Load the workflow below (also attached as JSON).
3. Checkpoint: sd_xl_base_1.0.safetensors (official StabilityAI release, integrity verified via safetensors header — 2515 tensors, official sha256, dated 2023-07-26).
4. VAE: sdxl.vae.safetensors (madebyollin/sdxl-vae-fp16-fix, used to rule out the known fp16-NaN bug in the checkpoint's bundled VAE).
5. Run the prompt. KSampler finishes in ~10s with no errors, but VAEDecode/SaveImage output is pure noise.
Minimal reproduction workflow (JSON):
```json
{
"last_node_id": 8, "last_link_id": 9,
"nodes": [
{"id":1,"type":"CheckpointLoaderSimple","widgets_values":["sd_xl_base_1.0.safetensors"]},
{"id":2,"type":"CLIPTextEncode","widgets_values":["professional studio photograph of a generic young adult woman fashion model, plain unremarkable generic facial features, simple neutral hairstyle, wearing a plain fitted black long-sleeve top, black fitted leggings, black high heels, plain even grey studio background, soft even studio lighting, full body visible head to feet, photorealistic, sharp focus"]},
{"id":3,"type":"CLIPTextEncode","widgets_values":["distorted, deformed, extra limbs, extra fingers, missing fingers, bad anatomy, blurry, low quality, watermark, text, cartoon, illustration, worst quality"]},
{"id":4,"type":"EmptyLatentImage","widgets_values":[1024,1024,1]},
{"id":5,"type":"KSampler","widgets_values":[999888,"randomize",25,6.5,"dpmpp_2m","karras",1.0]},
{"id":6,"type":"VAELoader","widgets_values":["sdxl.vae.safetensors"]},
{"id":7,"type":"VAEDecode"},
{"id":8,"type":"SaveImage","widgets_values":["Kalon_TESTE_LIMPO_SDXL_Puro"]}
],
"links": [
[1,1,0,5,0,"MODEL"],[2,1,1,2,0,"CLIP"],[3,1,1,3,0,"CLIP"],
[4,2,0,5,1,"CONDITIONING"],[5,3,0,5,2,"CONDITIONING"],
[6,4,0,5,3,"LATENT"],[7,5,0,7,0,"LATENT"],[8,6,0,7,1,"VAE"],[9,7,0,8,0,"IMAGE"]
]
}
```
### Debug Logs
```powershell
Windows PowerShell
Copyright (C) Microsoft Corporation. All rights reserved.
PS C:\Users\bobga> cd "D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable"
PS D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable> .\run_nvidia_gpu.bat --port 8189
D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
[INFO] setup plugin alembic.autogenerate.schemas
[INFO] setup plugin alembic.autogenerate.tables
[INFO] setup plugin alembic.autogenerate.types
[INFO] setup plugin alembic.autogenerate.constraints
[INFO] setup plugin alembic.autogenerate.defaults
[INFO] setup plugin alembic.autogenerate.comments
[INFO] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
[INFO] Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Checkpoint files will always be loaded safely.
[INFO] Total VRAM 8188 MB, total RAM 31821 MB
[INFO] pytorch version: 2.13.0+cu130
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 NVIDIA GeForce RTX 4060 Ti : cudaMallocAsync
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 12728.0
[INFO] Using pytorch attention
aimdo: src-win/cuda-detour.c:38:INFO:aimdo_setup_hooks: installing 6 hooks
aimdo: src-win/cuda-detour.c:28:DEBUG:install_hook_entries: hooks successfully installed
aimdo: src-win/shmem-detect.c:80:INFO:comfy-aimdo WDDM adapter match: NVIDIA GeForce RTX 4060 Ti runtime_luid=00000000:0000db25 dxgi_luid=00000000:0000db25
aimdo: src/control.c:248:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 4060 Ti (VRAM: 8187 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.13.14 (tags/v3.13.14:fd17997, Jun 10 2026, 13:03:48) [MSC v.1944 64 bit (AMD64)]
[INFO] ComfyUI version: 0.29.0
[INFO] comfy-aimdo version: 0.4.10
[INFO] comfy-kitchen version: 0.2.22
[INFO] comfyui-frontend-package version: 1.47.10
[INFO] comfyui-workflow-templates version: 0.11.19
[INFO] comfyui-embedded-docs version: 0.5.9
[INFO] comfy-kitchen version: 0.2.22
[INFO] comfy-aimdo version: 0.4.10
[INFO] [Prompt Server] web root: D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_frontend_package\static
[INFO] Asset seeder disabled
[INFO] No OpenGL_accelerate module loaded: No module named 'OpenGL_accelerate'
[INFO]
Import times for custom nodes:
[INFO] 0.0 seconds: D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable\ComfyUI\custom_nodes\websocket_image_save.py
[INFO]
[INFO] Context impl SQLiteImpl.
[INFO] Will assume non-transactional DDL.
[INFO] Using RAM pressure cache.
[INFO] Starting server
[INFO] To see the GUI go to: http://127.0.0.1:8188
[INFO]
Stopped server
Deseja finalizar o arquivo em lotes (S/N)? s
PS D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable> cd "D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable"
PS D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable> .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --port 8189
[INFO] setup plugin alembic.autogenerate.schemas
[INFO] setup plugin alembic.autogenerate.tables
[INFO] setup plugin alembic.autogenerate.types
[INFO] setup plugin alembic.autogenerate.constraints
[INFO] setup plugin alembic.autogenerate.defaults
[INFO] setup plugin alembic.autogenerate.comments
[INFO] Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
[INFO] Checkpoint files will always be loaded safely.
[INFO] Total VRAM 8188 MB, total RAM 31821 MB
[INFO] pytorch version: 2.13.0+cu130
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 NVIDIA GeForce RTX 4060 Ti : cudaMallocAsync
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 12728.0
[INFO] Using pytorch attention
aimdo: src-win/cuda-detour.c:38:INFO:aimdo_setup_hooks: installing 6 hooks
aimdo: src-win/cuda-detour.c:28:DEBUG:install_hook_entries: hooks successfully installed
aimdo: src-win/shmem-detect.c:80:INFO:comfy-aimdo WDDM adapter match: NVIDIA GeForce RTX 4060 Ti runtime_luid=00000000:0000db25 dxgi_luid=00000000:0000db25
aimdo: src/control.c:248:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 4060 Ti (VRAM: 8187 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.13.14 (tags/v3.13.14:fd17997, Jun 10 2026, 13:03:48) [MSC v.1944 64 bit (AMD64)]
[INFO] ComfyUI version: 0.29.0
[INFO] comfy-aimdo version: 0.4.10
[INFO] comfy-kitchen version: 0.2.22
[INFO] comfyui-frontend-package version: 1.47.10
[INFO] comfyui-workflow-templates version: 0.11.19
[INFO] comfyui-embedded-docs version: 0.5.9
[INFO] comfy-kitchen version: 0.2.22
[INFO] comfy-aimdo version: 0.4.10
[INFO] [Prompt Server] web root: D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_frontend_package\static
[INFO] Asset seeder disabled
[INFO] No OpenGL_accelerate module loaded: No module named 'OpenGL_accelerate'
[INFO]
Import times for custom nodes:
[INFO] 0.0 seconds: D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable\ComfyUI\custom_nodes\websocket_image_save.py
[INFO]
[INFO] Context impl SQLiteImpl.
[INFO] Will assume non-transactional DDL.
[INFO] Using RAM pressure cache.
[INFO] Starting server
[INFO] To see the GUI go to: http://127.0.0.1:8189
[INFO] got prompt
[INFO] Using pytorch attention in VAE
[INFO] Using pytorch attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] model weight dtype torch.float16, manual cast: None
[INFO] model_type EPS
[INFO] Using pytorch attention in VAE
[INFO] Using pytorch attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load SDXLClipModel
[INFO] Model SDXLClipModel prepared for dynamic VRAM loading. 1560MB Staged. 0 patches attached. Force pre-loaded 180 weights: 400 KB.
[INFO] Model SDXLClipModel prepared for dynamic VRAM loading. 1560MB Staged. 0 patches attached. Force pre-loaded 180 weights: 400 KB.
[INFO] Requested to load SDXL
[INFO] Model SDXL prepared for dynamic VRAM loading. 4896MB Staged. 0 patches attached. Force pre-loaded 512 weights: 1197 KB.
100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [00:09<00:00, 2.56it/s]
[INFO] Requested to load AutoencoderKL
[INFO] 0 models unloaded.
[INFO] Model AutoencoderKL prepared for dynamic VRAM loading. 159MB Staged. 0 patches attached. Force pre-loaded 112 weights: 182 KB.
[INFO] Prompt executed in 79.51 seconds
[INFO]
Stopped server
PS D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable> .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --port 8189 --disable-dynamic-vram
[INFO] setup plugin alembic.autogenerate.schemas
[INFO] setup plugin alembic.autogenerate.tables
[INFO] setup plugin alembic.autogenerate.types
[INFO] setup plugin alembic.autogenerate.constraints
[INFO] setup plugin alembic.autogenerate.defaults
[INFO] setup plugin alembic.autogenerate.comments
[INFO] Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
[INFO] Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Checkpoint files will always be loaded safely.
[INFO] Total VRAM 8188 MB, total RAM 31821 MB
[INFO] pytorch version: 2.13.0+cu130
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 NVIDIA GeForce RTX 4060 Ti : cudaMallocAsync
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 12728.0
[INFO] Using pytorch attention
[INFO] Python version: 3.13.14 (tags/v3.13.14:fd17997, Jun 10 2026, 13:03:48) [MSC v.1944 64 bit (AMD64)]
[INFO] ComfyUI version: 0.29.0
[INFO] comfy-aimdo version: 0.4.10
[INFO] comfy-kitchen version: 0.2.22
[WARNING] Dynamic vram disabled with argument. If you have any issues with dynamic vram enabled please give us a detailed reports as this argument will be removed soon. If you use gguf we recommend keeping dynamic vram enabled and using native ComfyUI model formats instead. ComfyUI native formats like fp8 will be faster even if they are larger than your memory.
[INFO] comfyui-frontend-package version: 1.47.10
[INFO] comfyui-workflow-templates version: 0.11.19
[INFO] comfyui-embedded-docs version: 0.5.9
[INFO] comfy-kitchen version: 0.2.22
[INFO] comfy-aimdo version: 0.4.10
[INFO] [Prompt Server] web root: D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_frontend_package\static
[INFO] Asset seeder disabled
[INFO] No OpenGL_accelerate module loaded: No module named 'OpenGL_accelerate'
[INFO]
Import times for custom nodes:
[INFO] 0.0 seconds: D:\AI\ComfyUI_CLEAN_TEST\ComfyUI_windows_portable\ComfyUI\custom_nodes\websocket_image_save.py
[INFO]
[INFO] Context impl SQLiteImpl.
[INFO] Will assume non-transactional DDL.
[INFO] Using RAM pressure cache.
[INFO] Starting server
[INFO] To see the GUI go to: http://127.0.0.1:8189
[INFO] got prompt
[INFO] Using pytorch attention in VAE
[INFO] Using pytorch attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] model weight dtype torch.float16, manual cast: None
[INFO] model_type EPS
[INFO] Using pytorch attention in VAE
[INFO] Using pytorch attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] Requested to load SDXLClipModel
[INFO] loaded completely; 1560.80 MB loaded, full load: True
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16
[INFO] Requested to load SDXL
[INFO] Unloaded partially: 1560.80 MB freed, 0.00 MB remains loaded, 446.76 MB buffer reserved, lowvram patches: 0
[INFO] loaded completely; 5607.67 MB usable, 4897.05 MB loaded, full load: True
100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [00:10<00:00, 2.49it/s]
[INFO] Requested to load AutoencoderKL
[INFO] Unloaded partially: 3106.92 MB freed, 1790.17 MB remains loaded, 75.06 MB buffer reserved, lowvram patches: 0
[INFO] loaded completely; 272.70 MB usable, 159.56 MB loaded, full load: True
[INFO] Prompt executed in 162.61 seconds
```
### Other
Full elimination chain tested (each with direct evidence, not assumption):
1. VAE fp16-NaN bug — fixed via sdxl-vae-fp16-fix, but noise persisted after.
2. LoRA/IPAdapter-FaceID/InsightFace/CLIPVision — bypassed entirely, noise persisted.
3. ControlNet (OpenPose-SDXL) — bypassed entirely; file audited via safetensors header as genuine.
4. torch/torchvision/xformers versions — confirmed unchanged via pip show.
5. Checkpoint file integrity — verified via safetensors header, official StabilityAI release.
6. NVIDIA driver — updated 560.94 → 610.88 Studio, clean reinstall + reboot, noise persisted.
7. Entire ComfyUI environment/custom nodes — ruled out via this exact from-scratch, zero-custom-node install.
8. Raw GPU/CUDA/PyTorch fp16 numerics — standalone script (matmul, sdpa, chained ops) all NaN=False/Inf=False. Note: this was run under a cu126 PyTorch build on a different (unaffected) install; this clean install itself uses cu130 — a discrepancy worth flagging.
9. Partial state_dict key mismatch during checkpoint load — full log inspected, zero "missing/unexpected/left over keys" warnings.
10. comfy-aimdo / dynamic VRAM CUDA hooks — re-ran with --disable-dynamic-vram to fully disable the aimdo CUDA-detour hooks visible in the startup log. Noise persisted, identical.
Related: #8749 (closed as stale, no resolution) reports the identical symptom on a different RTX 40-series GPU (RTX 4070 Laptop). Two independent reports on different RTX 40-series cards with the same zero-exception, pure-noise symptom.
4 条评论