mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-20 22:46:50 +00:00
- Q6_K dp4a decode (perf/moe-fused-decode, base): ~4.9x kernel, 28.2->33.0 T/s eager decode - Q4_K/Q6_K native prefill GEMM (perf/q4k-native-prefill): SWAR weight-decode + SROW=48 b128 LDS - WMMA flash-attention (perf/rocm-flash-attn, uncommitted worktree): flash.hip + ops/kernel/lib reg + test - HIP launch-arg ABI fix (fix/hip-launch-arg-abi): launch_binary/unary/pow pass arg ADDRESSES quant.hip merged cleanly (decode qdp4a/moe_qmatvec_dp4a_core vs prefill decode_w_half/SROW are disjoint regions). ops.rs merged cleanly (launcher fix in launch_binary/unary/pow; flash ADDS launch_flash_attn + FlashAttnShape).