April 2026: Delivered a critical ROCm SDPA dropout handling bug fix in PyTorch, re-applying and stabilizing the original dropout logic across forward and backward paths, restoring correct seed/offset propagation and ensuring compatibility with CK-specific dropout mask logic for testing. Re-enabled CK-parametrized SDPA tests and updated testing workflows to exercise backend selection and AOTriton paths. The changes improved ROCm SDPA reliability and reduced test flakiness, enabling more predictable experimentation and production workflows on AMD hardware.

March 2026 monthly summary for pytorch/pytorch. Focused on stabilizing the ROCm backend for CK SDPA dropout and delivering a concise, business-value driven improvement across the codebase. Implemented a targeted memory access fix to GPU memory handling for dropout, while maintaining Dynamo compatibility in output handling. Result is increased training stability and reliability on ROCm GPUs, reducing runtime errors and enabling broader hardware coverage for production workloads.

January 2026: Delivered a critical stability improvement in the PyTorch SDPA dropout path, fixing a device-side memory access fault and aligning tensor lifecycles and RNG handling. This results in more reliable attention computations on GPUs (ROCm) and reduces crashes during training and inference. Change tracked in PR #154864, with code contributions that enhance ROCm compatibility and overall GPU performance.

Month: 2025-09 — Summary of key features delivered, major improvements, and value realized in graphcore/pytorch-fork. Focused on ROCm optimization and kernel enhancements to boost stability and performance on ROCm-enabled platforms. Delivered build-time optimizations for CK SDPA, updated CK integration, and integrated AITER Fav3 forward kernels to accelerate tensor operations. No explicit bugs fixed this month; emphasis on performance, compatibility, and build reliability improvements.

August 2025 — ROCm/pytorch: Key features delivered and bugs fixed focused on performance, stability, and backend reliability. Highlights include the Composable Kernel (CK) kernel generation optimization to reduce kernel proliferation and the device-side memory access fix for SDPA with dropout on ROCm, improving attention stability and backend reliability.

July 2025 ROCm/pytorch: Delivered initial AITER-based optimization for ROCm backward assembly kernels in multi-head attention, enabling improved throughput for transformer workloads on ROCm devices. Key commit: b5ce77c1f5964293299eb1366f341872a4e47fa6. No major user-facing features beyond kernel optimization; no documented bug fixes this month. Foundations laid for further kernel-level performance gains and future work on mha_bwd optimizations.

February 2025 monthly summary for ROCm/FBGEMM focusing on feature enhancements in fused MoE and kernel optimization. Delivered fused MoE enhancements with local expert masking and optimized sorting dispatch; updated CK version; re-implemented kernel generation for fused MoE operations; refined dispatch mechanisms for fused MoE sorting kernels to boost flexibility, throughput, and scalability of MoE models.