xFormers-amd-rocm-gfx1151-cp312-torch.2.11

xFormers / PyTorch / ROCm / Windows Diagnosis

This document was generated automatically. The diagnostic script does not modify any files or environment variables.

Diagnosis Timestamp

System

Installed Package Versions

Relevant Environment Variables

PyTorch

xFormers

xFormers Build Metadata

{
  "version": {
    "cuda": null,
    "hip": "7.1.51803",
    "torch": "2.11.0",
    "python": "3.12.10",
    "flash": "0.0.0",
    "use_torch_flash": false
  },
  "env": {}
}

Output of python -m xformers.info

Exit code: 0

xFormers 0.0.32+local
memory_efficient_attention.ckF:                    available
memory_efficient_attention.ckB:                    available
memory_efficient_attention.ck_splitKF:             available
memory_efficient_attention.cutlassF-pt:            unavailable
memory_efficient_attention.cutlassB-pt:            unavailable
[email protected]:             unavailable
[email protected]:             unavailable
[email protected]:             unavailable
[email protected]:             unavailable
[email protected]:     unavailable
memory_efficient_attention.triton_splitKF:         available
indexing.scaled_index_addF:                        available
indexing.scaled_index_addB:                        available
indexing.index_select:                             available
sp24.sparse24_sparsify_both_ways:                  available
sp24.sparse24_apply:                               available
sp24.sparse24_apply_dense_output:                  available
sp24._sparse24_gemm:                               available
[email protected]:                 available
[email protected]:                        available
swiglu.dual_gemm_silu:                             available
swiglu.gemm_fused_operand_sum:                     available
swiglu.fused.p.cpp:                                available
is_triton_available:                               True
pytorch.version:                                   2.11.0+rocm7.13.0
pytorch.cuda:                                      available
gpu.compute_capability:                            11.5
gpu.name:                                          AMD Radeon(TM) 8060S Graphics
dcgm_profiler:                                     unavailable
build.info:                                        available
build.cuda_version:                                None
build.hip_version:                                 7.1.51803
build.python_version:                              3.12.10
build.torch_version:                               2.11.0
source.privacy:                                    open source

Detected xFormers Operators

Attention Function Tests

xFormers memory_efficient_attention uses the layout (Batch, Sequence, Heads, HeadDim). PyTorch SDPA uses (Batch, Heads, Sequence, HeadDim). Comparisons account for this difference.

ATTENTION: The CK kernel from ROCM Pytorch seems to has a bug when dealing with extremely short sequence lengths and very few heads. In regular Use-Cases it doesn't matter.

SDPA comparison: OK

small edge case: (1, 4, 1, 64)

typical small attention: (1, 128, 8, 64)

typical larger attention: (1, 512, 8, 64)

Random Data Test

PyTorch Build Configuration

PyTorch built with:
  - C++ Version: 201703
  - clang 23.0.0
  - MSVC 194435224
  - Intel(R) MKL-DNN v3.10.2 (Git Hash f1d471933dc852f956fd05389f9313c7148783d5)
  - OpenMP 202011
  - LAPACK is enabled (usually provided by MKL)
  - CPU capability usage: AVX512
  - HIP Runtime 7.13.99004
  - MIOpen 3.5.1
  - Build settings: BLAS_INFO=open, BUILD_TYPE=Release, COMMIT_SHA=71601f273f8d82178b99832e984a059e77d970c5, CXX_COMPILER=C:/home/runner/_work/_tool/Python/3.12.10/x64/Lib/site-packages/_rocm_sdk_devel/lib/llvm/bin/clang-cl.exe, CXX_FLAGS=/DWIN32 /D_WINDOWS /EHsc /Zc:__cplusplus /bigobj /FS /utf-8 -DUSE_PTHREADPOOL -DNDEBUG -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE /wd4624 /wd4068 /wd4067 /wd4267 /wd4661 /wd4717 /wd4244 /wd4804 /wd4273, LAPACK_INFO=open, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, TORCH_VERSION=2.11.0, USE_CUDA=OFF, USE_CUDNN=OFF, USE_CUSPARSELT=OFF, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_GLOO=OFF, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=OFF, USE_NNPACK=OFF, USE_OPENMP=ON, USE_ROCM=ON, USE_ROCM_KERNEL_ASSERT=OFF, USE_XCCL=OFF, USE_XPU=OFF,

Relevant Files

Optional Export / Dependency Information

dumpbin.exe was not found or _C.pyd is not available. This section is optional.

pip Package Information

pip show torch

Exit code: 0

Name: torch
Version: 2.11.0+rocm7.13.0
Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration
Home-page: https://pytorch.org
Author: 
Author-email: PyTorch Team <[email protected]>
License: BSD-3-Clause
Location: C:\Users\xxx\AppData\Local\Programs\Python\Python312\Lib\site-packages
Requires: filelock, fsspec, jinja2, networkx, rocm, setuptools, sympy, typing-extensions
Required-by: accelerate, auto-round, bitsandbytes, clean-fid, clip-anytorch, compressed-tensors, controlnet_aux, cut-cross-entropy, descript-audio-codec, descript-audiotools, docling-ibm-models, easyocr, fairscale, flair, invisible-watermark, julius, kornia, lightning, lpips, lycoris_lora, open_clip_torch, optimum-quanto, peft, pytorch-fid, pytorch-lightning, pytorch-wavelets, pytorch_revgrad, SAM-2, sentence-transformers, silero-vad, spandrel, timm, torch-stoi, torchdiffeq, torchmetrics, torchsde, torchvision, transformer-smaller-training-vocab, turboquant, ultralytics, ultralytics-thop, unsloth, unsloth_zoo, xformers

pip show torchvision

Exit code: 0

Name: torchvision
Version: 0.26.0+rocm7.13.0
Summary: image and video datasets and models for torch deep learning
Home-page: https://github.com/pytorch/vision
Author: PyTorch Core Team
Author-email: [email protected]
License: BSD
Location: C:\Users\xxx\AppData\Local\Programs\Python\Python312\Lib\site-packages
Requires: numpy, pillow, torch
Required-by: clean-fid, clip-anytorch, controlnet_aux, docling-ibm-models, easyocr, lpips, open_clip_torch, pytorch-fid, SAM-2, spandrel, timm, ultralytics, unsloth

pip show torchaudio

Exit code: 0

Name: torchaudio
Version: 2.11.0+rocm7.13.0
Summary: An audio package for PyTorch
Home-page: https://github.com/pytorch/audio
Author: Soumith Chintala, David Pollack, Sean Naren, Peter Goldsborough, Moto Hira, Caroline Chen, Jeff Hwang, Zhaoheng Ni, Xiaohui Zhang
Author-email: [email protected]
License: 
Location: C:\Users\xxx\AppData\Local\Programs\Python\Python312\Lib\site-packages
Requires: 
Required-by: descript-audio-codec, descript-audiotools, silero-vad, torch-stoi

pip show xformers

Exit code: 0

Name: xformers
Version: 0.0.32+6b467648.d20260812
Summary: XFormers: A collection of composable Transformer building blocks.
Home-page: https://facebookresearch.github.io/xformers/
Author: Facebook AI Research
Author-email: [email protected]
License: 
Location: C:\Users\xxx\AppData\Local\Programs\Python\Python312\Lib\site-packages
Requires: numpy, torch
Required-by: unsloth

pip show triton

Exit code: 1

WARNING: Package(s) not found: triton

Automatic Summary

• ✅ The script was not started from a recognized xFormers source tree.
• ✅ PyTorch reports HIP/ROCm: 7.13.99004
• ✅ PyTorch detects at least one GPU.
• ✅ xFormers could be imported.