[Issue]: ComfyUI vae stage using cudnn will get error on gfx1103

[LuXuxue]

Problem Description

i'm testing torch-2.10.0a0+rocm7.11.0a20251211-cp312-cp312-win_amd64.whl on gfx1103(780M)
using comfyui with torch.backends.cudnn.enabled(tweak by custom_nodes) will cause black screen
and if using rdp to login system, we can see error log like this:

MIOpen(HIP): Error [GenericSearch] Error: Exception encountered : DESKTOP-5V05ODG:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/hip/handlehip.cpp:161: Error setting device 0: unspecified launch failure

MIOpen(HIP): Error [GenericSearch] #11 (12)  Failed rc=1
MIOpen(HIP): Error [FindSolutionImpl] Search Failed: 2x640x48x32x1x3x3x1x1280x2x1x1x0x1x1x0x1x1x0x0x1xNCHWxFP16xFConvBinWinogradRxSf3x2: DESKTOP-5V05ODG:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/generic_search.hpp:718: Search failed
MIOpen(HIP): Warning [Find] DESKTOP-5V05ODG:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/hip/handlehip.cpp:735: : unspecified launch failure
MIOpen(HIP): Error [EvaluateInvokers] DESKTOP-5V05ODG:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/hipoc/hipoc_program.cpp:368: Empty code object path.
MIOpen Error: DESKTOP-5V05ODG:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/hipoc/hipoc_program.cpp:368: Empty code object path.
C:\Portable\python_embeded-new\Lib\site-packages\torch\nn\modules\conv.py:548: UserWarning: HIP warning: unspecified launch failure (Triggered internally at B:/src/torch/aten/src\ATen/hip/impl/HIPGuardImplMasqueradingAsCUDA.h:83.)
  return F.conv2d(

without cudnn, speed of sdxl@1024x1536 is about 6s/it
and with cudnn, speed of sdxl@1024x1536 is about 3.7s/it(scottt‘s build)
maybe fix conv2d issue can speedup comfy

Operating System

Windows 11 24H2

CPU

AMD Ryzen 7 8845HS

GPU

Radeon 780M Graphics

ROCm Version

rocm7.11.0a20251211

ROCm Component

MIOpen

Steps to Reproduce

1. pip install rocm-7.11.0a20251211.tar.gz rocm_sdk_core-7.11.0a20251211-py3-none-win_amd64.whl rocm_sdk_devel-7.11.0a20251211-py3-none-win_amd64.whl rocm_sdk_libraries_gfx110x_all-7.11.0a20251211-py3-none-win_amd64.whl torch-2.10.0a0+rocm7.11.0a20251211-cp312-cp312-win_amd64.whl torchvision-0.25.0a0+rocm7.11.0a20251211-cp312-cp312-win_amd64.whl torchaudio-2.10.0a0+rocm7.11.0a20251211-cp312-cp312-win_amd64.whl
2. install comfy
3. install https://github.com/patientx/ComfyUI-Zluda/blob/master/cfz/nodes/cfz_cudnn.toggle.py
4. turn on enable_cudnn and cudnn_benchmark
5. error

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[ScottTodd]

Does this reproduce on torch 2.9.1? I'm setting up pytorch unit test coverage on Windows (#2258) and I see similar HIP error: unspecified launch failure on torch 2.10.0a but not on torch 2.9.1, at least on gfx1151.

[astrelsky]

I think this might be the same as #2488 and #2489. For which I get the problem using 2.9.1+rocm7.11.0a20251211

[LuXuxue]

Does this reproduce on torch 2.9.1? I'm setting up pytorch unit test coverage on Windows (#2258) and I see similar HIP error: unspecified launch failure on torch 2.10.0a but not on torch 2.9.1, at least on gfx1151.

same error, and 2.9.1 have not enabled aotriton, will get worse speed, about 7.5s/it

[astrelsky]

Does this reproduce on torch 2.9.1? I'm setting up pytorch unit test coverage on Windows (#2258) and I see similar HIP error: unspecified launch failure on torch 2.10.0a but not on torch 2.9.1, at least on gfx1151.

same error, and 2.9.1 have not enabled aotriton, will get worse speed, about 7.5it/s

7.5it/s is really good. Perhaps you meant 7.5s/it?.

[LuXuxue]

test report on torch-2.10.0a0+rocm7.11.0a20251213-cp312-cp312-win_amd64.whl:
after i upgrade torch, i run comfy directly with cudnn. after a warning: xnack 'Off' was requested for a processor that does not support it! log and a long time(about 11min) benchmark(or something else), it generated a blank image.
a normal image was generated on the second run, at speed of about 5.3s/it, which is a bit faster than the speed without cudnn, but is slower than scottt‘s build.
maybe it is the backend difference, so i closed comfy, and run a test script(validate_pytorch_vroom.py), everything passed and it said backend is Cublas, while scottt‘s build's backend is Cublaslt.
i run comfy again, and i meet the black screen and the error log. delete .miopen folder and reboot system can not solve this problem.
comfy dead, but the test script is still alive, showing a better performance number in theory.

[astrelsky]

@LuXuxue your symptoms are suspiciously similar to behavior I was seeing and hopefully ROCm/rocm-libraries#3364 will fix it.

[LuXuxue]

test on torch-2.9.1+rocm7.11.0a20251219-cp312-cp312-win_amd64.whl: fixed, speed is about 5.2s/it with or without cudnn(with cudnn is a little faster, but it takes a long time before the first run)
there is no torch 2.10a whl today, i will test on 2.10 when new whl released