Search results

Bfloat16 large numerical discrepancy on Blackwell

module: bfloat16

module: correctness (silent)issue that returns an incorrect result silentlyissue that returns an incorrect result silently

module: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in general

module: numerical-stabilityProblems related to numerical stability of operationsProblems related to numerical stability of operations

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#163574 In pytorch/pytorch;

· drisspg opened on Sep 22, 2025

<code>torch._grouped_gemm</code> has poor performance for large batch sizes with dtype BF16

module: bfloat16

module: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in general

module: linear algebraIssues related to specialized linear algebra operations in PyTorch; includes matrix multiply matmulIssues related to specialized linear algebra operations in PyTorch; includes matrix multiply matmul

module: performanceIssues related to performance, either of kernel code or framework glueIssues related to performance, either of kernel code or framework glue

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#163425 In pytorch/pytorch;

· Skylion007 opened on Sep 20, 2025

Incorrect bf16 results in Conv

module: bfloat16

module: correctness (silent)issue that returns an incorrect result silentlyissue that returns an incorrect result silently

module: mkldnnRelated to Intel IDEEP or oneDNN (a.k.a. mkldnn) integrationRelated to Intel IDEEP or oneDNN (a.k.a. mkldnn) integration

needs reproductionEnsure you have actionable steps to reproduce the issue. Someone else needs to confirm the repro.Ensure you have actionable steps to reproduce the issue. Someone else needs to confirm the repro.

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#163347 In pytorch/pytorch;

· gvartany-amd opened on Sep 19, 2025

[Track] Optimizing Groupwise convolution for aarch64

module: armRelated to ARM architectures builds of PyTorch. Includes Apple M1Related to ARM architectures builds of PyTorch. Includes Apple M1

module: bfloat16

module: convolutionProblems related to convolutions (THNN, THCUNN, CuDNN)Problems related to convolutions (THNN, THCUNN, CuDNN)

module: performanceIssues related to performance, either of kernel code or framework glueIssues related to performance, either of kernel code or framework glue

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#158923 In pytorch/pytorch;

· aditew01 opened on Jul 23, 2025

sparse_csr bfloat16 matrix multiplication backward is 10x slower than float16

module: bfloat16

module: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in general

module: performanceIssues related to performance, either of kernel code or framework glueIssues related to performance, either of kernel code or framework glue

module: sparseRelated to torch.sparseRelated to torch.sparse

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#157808 In pytorch/pytorch;

· zbh2047 opened on Jul 8, 2025

crash in torch.histc

module: bfloat16

module: crashProblem manifests as a hard crash, as opposed to a RuntimeErrorProblem manifests as a hard crash, as opposed to a RuntimeError

module: edge casesAdversarial inputs unlikely to occur in practiceAdversarial inputs unlikely to occur in practice

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#155393 In pytorch/pytorch;

· cx104906 opened on Jun 7, 2025

bfloat16 Conv2d slower than float16 on 4090

module: bfloat16

module: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in general

module: halfRelated to float16 half-precision floatsRelated to float16 half-precision floats

module: performanceIssues related to performance, either of kernel code or framework glueIssues related to performance, either of kernel code or framework glue

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#154351 In pytorch/pytorch;

· Seeker98 opened on May 26, 2025

[cuBLAS] Strange behavior with <code>out_dtype</code> argument in <code>torch.addmm</code>

matrix multiplication

module: bfloat16

module: cublasProblem related to cublas supportProblem related to cublas support

module: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in general

module: halfRelated to float16 half-precision floatsRelated to float16 half-precision floats

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#154086 In pytorch/pytorch;

· eqy opened on May 22, 2025

missalignment with differenet shape in F.linear with bf16 dtype

module: bfloat16

module: linear algebraIssues related to specialized linear algebra operations in PyTorch; includes matrix multiply matmulIssues related to specialized linear algebra operations in PyTorch; includes matrix multiply matmul

module: padding

needs reproductionEnsure you have actionable steps to reproduce the issue. Someone else needs to confirm the repro.Ensure you have actionable steps to reproduce the issue. Someone else needs to confirm the repro.

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#153033 In pytorch/pytorch;

· likelyzhao opened on May 7, 2025

<code>torch.bmm</code> is slow on non-contiguous BF16 CPU tensors

matrix multiplication

module: bfloat16

module: cpuCPU specific problem (e.g., perf, algorithm)CPU specific problem (e.g., perf, algorithm)

module: halfRelated to float16 half-precision floatsRelated to float16 half-precision floats

module: performanceIssues related to performance, either of kernel code or framework glueIssues related to performance, either of kernel code or framework glue

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#151934 In pytorch/pytorch;

· ar0ck opened on Apr 22, 2025

<code>torch.mul</code> uses <code>OpMathType</code> for computation.

module: bfloat16

module: cudaRelated to torch.cuda, and CUDA support in generalRelated to torch.cuda, and CUDA support in general

module: halfRelated to float16 half-precision floatsRelated to float16 half-precision floats

module: python frontendFor issues relating to PyTorch's Python frontendFor issues relating to PyTorch's Python frontend

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#147134 In pytorch/pytorch;

· ysiraichi opened on Feb 13, 2025

inconsistency in <code>torch.special.xlogy</code> on CPU and GPU

module: bfloat16

module: numerical-stabilityProblems related to numerical stability of operationsProblems related to numerical stability of operations

triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Status: Open.

#140892 In pytorch/pytorch;

· usmonali4 opened on Nov 17, 2024