torch._grouped_mm crashes (0xC0000005) on gfx1200 (RX 9060 XT) Windows ROCm 7.2.1

[PialGhosh2233]

🐛 Describe the bug

Describe the bug

torch._grouped_mm crashes with a fatal access violation (Exception Code 0xC0000005) on Windows with an AMD Radeon RX 9060 XT (gfx1200 / RDNA 4) using ROCm 7.2.1 PyTorch wheels. The process terminates immediately with no catchable Python exception — try/except cannot catch it.

Basic torch.mm on CUDA works fine. Only _grouped_mm crashes.

Minimal reproducible example

import torch

# This works fine
x = torch.randn(4, 4, device='cuda')
y = torch.randn(4, 4, device='cuda')
z = torch.mm(x, y)
print('basic cuda mm OK:', z.shape)

# This crashes the entire Python process (0xC0000005)
result = torch._grouped_mm(x.unsqueeze(0), y.unsqueeze(0))
print('grouped_mm OK')  # never reached

Error output

basic cuda mm OK
Exception Code: 0xC0000005
  #0 0x00007fffc9250000 (torch_hip.dll+0x0)
  #1 at::cuda::_grouped_mm(...) (torch_hip.dll+0x270c4a9)
  #2 at::cuda::_fused_adagrad_(...) (torch_hip.dll+0x288e711)
  #3 at::_ops::_grouped_mm::redispatch(...) (torch_cpu.dll+0x1309a8f)
  #4 torch::autograd::impl::JitDecompRegisterer::JitDecompRegisterer(...) (torch_cpu.dll+0x3b9518b)
  #5 at::_ops::_grouped_mm::call(...) (torch_cpu.dll+0x1308a32)

System information

• OS: Windows 11 64-bit
• GPU: AMD Radeon RX 9060 XT
• GPU Architecture: gfx1200 (RDNA 4)
• VRAM: 16 GB
• PyTorch version: 2.9.1+rocm7.2.1
• ROCm version: 7.2.1
• HIP SDK version: 7.1
• Python version: 3.12
• PyTorch installed via: AMD official pip wheels from https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/windows/install-pytorch.html

Expected behavior

torch._grouped_mm should either work correctly on gfx1200 Windows, or raise a catchable Python exception instead of fatally crashing the process.

Additional context

• torch.mm, torch.cuda.is_available(), and basic tensor operations all work correctly on this GPU
• The crash is not catchable with try/except - it terminates the entire Python process
• This also causes import unsloth to crash, since Unsloth calls _grouped_mm during module initialization
• hipInfo correctly detects the GPU and reports gfx1200 architecture
• Training of CNN and ViT models works correctly using standard PyTorch operations

Versions

PyTorch version: 2.9.1+rocm7.2.1
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 7.2.53211-158bd99533

OS: Microsoft Windows 11 Pro (10.0.26200 64-bit)
GCC version: Could not collect
Clang version: 21.0.0git (git@github.com:Compute-Mirrors/llvm-project 5dcc622b51ecd499912c1062ce2b0ecda60d8e93)
CMake version: Could not collect
Libc version: N/A

Python version: 3.12.2 (tags/v3.12.2:6abddd9, Feb 6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] (64-bit runtime)
Python platform: Windows-11-10.0.26200-SP0
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to:
GPU models and configuration: AMD Radeon RX 9060 XT (gfx1200)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: 7.2.53211
MIOpen runtime version: 3.5.1
Is XNNPACK available: True

CPU:
Name: Intel(R) Core(TM) i5-9400F CPU @ 2.90GHz
Manufacturer: GenuineIntel
Family: 205
Architecture: 9
ProcessorType: 3
DeviceID: CPU0
CurrentClockSpeed: 2904
MaxClockSpeed: 2904
L2CacheSize: 1536
L2CacheSpeed: None
Revision: None

Versions of relevant libraries:
[pip3] numpy==2.4.6
[pip3] torch==2.9.1+rocm7.2.1
[pip3] torchao==0.17.0
[pip3] torchaudio==2.9.1+rocm7.2.1
[pip3] torchvision==0.24.1+rocm7.2.1
[pip3] triton-windows==3.7.0.post26
[conda] Could not collect

[mapatel-amd]

Hey @nkulshre-amd, I saw that this issue has the error code 0xC0000005 as one of the issues I was working on before. If it helps, check out the most recent comment put by me on this issue. It has details on what the problem might be and what's happening right now for it. If you have any questions let me know.

[nkulshre-amd]

Able to reproduce issue on above GPU configuration and latest ROCm release (7.13)

[PialGhosh2233]

Hey @nkulshre-amd, I saw that this issue has the error code 0xC0000005 as one of the issues I was working on before. If it helps, check out the most recent comment put by me on this issue. It has details on what the problem might be and what's happening right now for it. If you have any questions let me know.

I have 2 current recommendations for you to try:

Turn off integrated graphics in BIOS (I recommend this one) ----> My CPU don't have any igpu
Revert your adrenalin driver version to 26.3.1 (Might not have as good performance or may have other bugs, but should work for this case) ---> How much performance will I loss if i rollback to old driver ?

[nkulshre-amd]

Hi @PialGhosh2233, I can confirm that this is a separate issue since like you I have tried with only dGPU and rolling back driver isn't a fix. I'm continuing to triage it.

[nkulshre-amd]

Issue seems to be patched in latest nightly (7.14)
@PialGhosh2233, please try upgrading via pip install --index-url https://rocm.nightlies.amd.com/whl-multi-arch/ "rocm[libraries,device-gfx1200]" "torch[device-gfx1200]"

pip show torch should display

Name: torch
Version: 2.12.0+rocm7.14.0a20260615