July 2026 ROCm/aiter monthly summary focused on stability, performance, and extended hardware support for GEMM2 decode. Key work stabilized GEMM2 decode on Kimi-K2.5/K2.6 by aligning block sizes and re-tuning kernel configurations, addressing output corruption, and porting BM16 kernels across token buckets 8–256 on MI355X GPUs. The BM16 path was extended to decode token buckets 1/2/4 as well, with coordinated tuning. Validation on vLLM nightly and GSM8K demonstrated correctness parity and resilience, while end-to-end performance improved across representative workloads. Key results include substantial decode-path improvements (up to 21% throughput for specific token counts) and notable end-to-end gains (approximately +15% in c32 workloads under realistic prompts; about +2.4% E2E gain for BM16 vs SB32 in tested scenarios). The changes also reduced decode output anomalies and increased throughput consistency across token bucket configurations. Co-authored commits with Claude to ensure robust coverage and reproducibility, with extensive interleaved benchmarking to manage drift and validate decisions. Technologies and skills demonstrated include GPU kernel tuning (BM16, GEMM2, tile_m vs block_m coordination), FP4 activation quantization integration, MI355X hardware tuning, vLLM nightly validation, performance measurement, and cross-team collaboration for large-scale kernel porting and configuration management.