comfykitchen fatal crash at torch.cuda.is_available()

[germzx]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

ComfyUI desktop or portable windows AMD version to launch at all.

Actual Behavior

Attempting to install ComfyUI for Windows 11, failed midway through due to comfykitchen crash.

Switched to ComfyUI Windows portable AMD, fail to launch due to the same.

Steps to Reproduce

This bug is fatal before ComfyUI even loads. It's part of the starting code with comfykitchen.

Debug Logs

N/A

Other

Environment:
GPU: AMD Radeon RX 6800 XT (gfx1030, RDNA2)
Driver: AMD Adrenalin 26.3.1
AMD HIP SDK: 7.1
OS: Windows 11
ComfyUI: Desktop app + AMD portable (both affected)
torch: 2.9.1+rocm7.2.1

Both the desktop app and the AMD portable crash on startup with a fatal Windows access violation. The crash occurs in comfy_kitchen\backends\cuda_init_.py when it calls torch.cuda.is_available() or torch.cuda._lazy_init() — these throw an unrecoverable access violation on ROCm Windows instead of returning False or raising a catchable exception.

Workaround: uninstall comfy-kitchen and comfy-aimdo, patch the hard imports in main.py to try/except. However model_management.py then crashes similarly at torch.cuda.current_device().

Request: torch.cuda.is_available() calls should be wrapped in try/except throughout, and comfy_kitchen should gracefully skip CUDA backend registration on AMD/ROCm rather than fatally crashing the process.

Additionally, of note, ROCm is starting to ship with amdhip64_6.dll or amdhip64_7.dll (basically often there but 'invisible' due to specificity of targeting) as well as NO amdhsa64.dll and now uses amd_comgr.dll instead. Anyways they keep changing the naming schema on those. Those issues were revealed after I uninstalled comfykitchen which wouldn't stop it's cuda crash.

[chejh-amd]

Hi @germzx This might be happening below normal Python exception handling, not only in ComfyUI startup code.

Could you run this in the same env where it crashes?

python -X faulthandler -c "import torch; print('torch', torch.__version__, 'hip', torch.version.hip); print('cuda_available', torch.cuda.is_available())"

And from Windows Event Viewer (python.exe / ComfyUI), could you share:

• Faulting module name
• Exception code
• Fault offset

If this already crashes on plain torch.cuda.is_available(), it likely points more to HIP/PyTorch runtime than ComfyUI logic.

[germzx]

Hey @chejh-amd thanks for the fast response. I managed to get it working after some intensive patching. To be honest I can follow the logic but don't know coding well myself. I provided overarcing logic and Claude helped me via powershell and many .py edits.

So, the following is a summarized post. (I had installed AMD HIP SDK but that ended up being totally irrelevant). The actionable solution is at the bottom. Sorry it's a long post, I went through a lot to get it to work. I also don't know if the solutions at the bottom other than the try/else are things you'd want to implement.
.
.
.
.
Title: AMD RX 6000 series (gfx1030) on Windows — comfy_aimdo/comfy_kitchen crash on startup, workaround documented
Environment:

GPU: AMD Radeon RX 6800 XT (gfx1030, RDNA2, 16GB VRAM)
OS: Windows 11
AMD Adrenalin driver: 26.3.1
torch used in workaround: 2.8.0a0+rocm6.4.4
Python: 3.12.10
ComfyUI: 0.19.3 (AMD portable)
.
.
Problem:
The ComfyUI AMD portable crashes on startup with two distinct failure modes on Windows with an RX 6800 XT.

Crash 1 — Fatal access violation (comfy_kitchen):
comfy_kitchen\backends\cuda_init_.py calls torch.cuda.is_available() at import time. With torch 2.9.1+rocm7.2.1 and an RX 6800 XT, this produces a fatal Windows access violation (exception code 0xC0000005) rather than a catchable Python exception, killing the process instantly. Since comfy_kitchen is a hard import with no fallback, there is no recovery.

Crash 2 — No HIP GPUs available (driver/torch version mismatch):
After removing comfy_kitchen, model_management.py crashes at torch.cuda.current_device() with RuntimeError: No HIP GPUs are available. Root cause: AMD Adrenalin 26.3.1 ships amdhip64_6.dll into System32 — exposing HIP 6 at the driver level — while the portable bundles torch 2.9.1+rocm7.2.1 which requires HIP 7. These are incompatible.
.
.
Root cause details:
The Adrenalin driver ships its HIP runtime directly into System32 as versioned DLLs. This is what torch uses at runtime — no separately installed SDK is involved. On Adrenalin 26.3.1, these are present in System32:

• amdhip64_6.dll — HIP 6 runtime
• amd_comgr_2.dll — code object manager

Note: AMD no longer ships amdhip64.dll or amdhsa64.dll as unversioned filenames. Detection logic must use wildcard patterns (amdhip64_.dll, amd_comgr_.dll) and extract the version number from the filename. For amdhsa64 specifically — it may still exist on some systems as a legacy name, so a safe fallback would be: check for amd_comgr_.dll first, then fall back to amdhsa64.dll or amdhsa64_.dll before failing.
Despite HIP SDK 7.1.1 officially listing gfx1030 as supported, the Adrenalin 26.3.1 driver only exposes HIP 6 for RDNA2 hardware. The additionally installed SDK version is irrelevant — only the driver-provided System32 DLLs matter for runtime.
.
.
.
Workaround that works:

Uninstall comfy-kitchen and comfy-aimdo from python_embeded
Downgrade torch to match the driver's HIP 6 exposure:

pip install https://repo.radeon.com/rocm/windows/rocm-rel-6.4.4/torch-2.8.0a0+gitfc14c65-cp312-cp312-win_amd64.whl
pip install https://repo.radeon.com/rocm/windows/rocm-rel-6.4.4/torchvision-0.24.0a0+c85f008-cp312-cp312-win_amd64.whl
pip install https://repo.radeon.com/rocm/windows/rocm-rel-6.4.4/torchaudio-2.6.0a0+1a8f621-cp312-cp312-win_amd64.whl

Wrap all hard import comfy_aimdo.* statements in try/except ImportError in:

• main.py — import comfy_aimdo.control and comfy_aimdo.control.init()
• comfy/windows.py — import comfy_aimdo.control
• execution.py — import comfy_aimdo.model_vbar
• comfy/pinned_memory.py — import comfy_aimdo.host_buffer, import comfy_aimdo.torch
• comfy/ops.py — import comfy_aimdo.model_vbar, import comfy_aimdo.torch
• comfy/model_patcher.py — import comfy_aimdo.model_vbar
• comfy/utils.py — import comfy_aimdo.model_mmap

Create a launcher.py that calls os.add_dll_directory() for torch\lib before importing torch — Python 3.8+ no longer uses PATH for DLL dependency resolution
Set HSA_OVERRIDE_GFX_VERSION=10.3.0 and HIP_VISIBLE_DEVICES=0 in the environment

After these steps ComfyUI boots successfully:

• AMD arch: gfx1030
• ROCm version: (6, 4)
• Device: cuda:0 AMD Radeon RX 6800 XT : native
• Total VRAM 16368 MB

Suggested fixes:

ComfyUI's responsibility (straightforward):
Wrap all import comfy_aimdo.* statements in try/except ImportError. These are currently hard imports that fatally crash on any system without comfy_aimdo. This single change would allow AMD users to at least get a meaningful error rather than a fatal crash.

For portable installer (more involved):
Detect the driver's exposed HIP version by scanning System32 for amdhip64_.dll and extracting the version number N. Select the matching ROCm torch wheel automatically. Do not rely on any installed HIP SDK — only the driver-provided System32 DLLs matter. Account for AMD's current naming conventions: amdhip64_N.dll (not amdhip64.dll), and amd_comgr_N.dll with a fallback to amdhsa64.dll/amdhsa64_.dll for older systems.
To get the portable working out of the box for users in similar environments, implementing runtime detection of the driver's exposed HIP version and automatically selecting the correctly matched ROCm torch version would solve it without any manual intervention from the user.

Re: running the faulthandler command: The issue no longer reproduces after the workaround and the original environment has been substantially patched. The access violation occurred specifically with torch 2.9.1+rocm7.2.1 when the driver only exposes HIP 6 — downgrading to torch 2.8.0+rocm6.4.4 resolved it entirely.

[chejh-amd]

Thanks for writing all this up, splitting crash 1 (fatal AV in torch.cuda init) from crash 2 (No HIP GPUs after dropping comfy_kitchen) makes the picture much easier to follow.

The driver HIP major vs bundled ROCm torch angle is a solid lead for anyone else seeing the same combo — sucks that it took that much patching on your side.

Re: gating comfy_aimdo behind try/`except ImportError, that’s the same idea as making those imports optional so the process doesn’t die immediately when something’s missing.

No worries on not being able to re-run the minimal faulthandler line after the downgrade, thanks for spelling out what actually fixed it on your box.