From 71bcc5b204a418b5c1bea2a769bbb44807dfa7bb Mon Sep 17 00:00:00 2001 From: Shao Date: Wed, 19 Nov 2025 14:59:27 -0700 Subject: [PATCH] Add PyTorch 2.9 release notes for ROCm --- .../pytorch-compatibility.rst | 22 ++++++++++++++----- 1 file changed, 16 insertions(+), 6 deletions(-) diff --git a/docs/compatibility/ml-compatibility/pytorch-compatibility.rst b/docs/compatibility/ml-compatibility/pytorch-compatibility.rst index 3e900ac3e..e5634d188 100644 --- a/docs/compatibility/ml-compatibility/pytorch-compatibility.rst +++ b/docs/compatibility/ml-compatibility/pytorch-compatibility.rst @@ -401,15 +401,25 @@ with ROCm. Key features and enhancements for PyTorch 2.9 with ROCm 7.1.1 ================================================================================ -- Added OCP Micro-scaling Format (mx-fp8/mx-fp4) support for advanced precision training. +- Scaled Dot Product Attention (SDPA) upgraded to use AOTriton version 0.11b -- `torch.backends.miopen.immediate` flag to toggle MIOpen Immediate Mode independently of - deterministic and benchmark settings, providing finer control over convolution execution. +- Default hipBLASLt support enabled for gfx908 architecture on ROCm 6.3 and later -- rocSOLVER now used for Cholesky inversion operations, providing improved numerical stability - and performance for linear algebra workloads. +- MIOpen now supports channels last memory format for 3D convolutions and batch normalization -- MI355X GPU testing enabled in CI. +- NHWC convolution operations in MIOpen optimized by eliminating unnecessary transpose operations + +- Improved tensor.item() performance by removing redundant synchronization + +- Enhanced performance for elementwise operations and reduction kernels + +- Added support for grouped GEMM operations through fbgemm_gpu generative AI components + +- Resolved device error in Inductor when using CUDA graph trees with HIP + +- Corrected logsumexp scaling in AOTriton-based SDPA implementation + +- Added stream graph capture status validation in memory copy synchronization functions Key features and enhancements for PyTorch 2.7/2.8 with ROCm 7.1.1 ================================================================================