April 2026 monthly summary for ROCm/composable_kernel: Delivered native SwigluStep activation support in the MoE blockscale path of the composable kernel, enabling this activation for specific model architectures and AITER-based workflows. Implemented the corresponding blockscale epilogue path and preserved existing SiLU/GELU behavior to avoid regressions. Scoped changes to two CK files; validated with classic CK 2-stage MoE runtime and AITER MoE enabled, ensuring stability. This work expands model versatility, maintains performance, and positions the project for broader activation-function support in future CK releases.

March 2026 monthly summary for unsloth. Focused on delivering ROCm RDNA GPU support, stability improvements, and performance optimizations to accelerate training workloads on AMD GPUs while preserving compatibility across CDNA and RDNA generations. Implemented GPU-detection and selective compilation controls, performed targeted kernel optimizations, and cleaned up erroneous error handling paths to reduce false positives. Achieved measurable improvements on ROCm 7.1 test hardware and hardened the repository against misconfigurations and unsupported hardware.

February 2026 (unslothai/unsloth): Delivered a critical stability fix for Triton kernels on gfx950 by updating the is_cdna() thread-limit checks to include gfx950, aligning with the 1024-thread workgroup limit used by gfx942. This prevents OutOfResources crashes and ensures consistent performance for GPU-accelerated workloads.