ViperEkura f9efb705b8 perf: output fp8 gemm in bf16 instead of fp32 ...

- cublasLt C layout and buffer switched to CUDA_R_16BF, halving output bandwidth
- downstream ops (RMSNorm etc.) keep matching bf16 dtype, fused kernels stay
- numeric error unchanged (0.19% vs fp32 ref on quantized inputs)

2026-08-14 01:08:26 +08:00

..

attn_common.h

perf: flatten paged prefill tile dispatch

2026-08-09 23:12:53 +08:00

attn_decode_split_kv_mma.cuh

refactor: unify attention param field names

2026-08-09 20:23:58 +08:00

attn_decode_split_kv.cuh

refactor: tidy attention params and launcher interfaces

2026-08-09 20:52:06 +08:00

attn_decode.cu

refactor: tidy attention params and launcher interfaces

2026-08-09 20:52:06 +08:00

attn_dispatchers.cuh

perf: flatten paged prefill tile dispatch

2026-08-09 23:12:53 +08:00

attn_entry_utils.cuh

perf: flatten paged prefill tile dispatch

2026-08-09 23:12:53 +08:00

attn_kv_source.cuh

refactor: streamline Q block mapping

2026-08-10 08:40:18 +08:00

attn_mma_utils.cuh

refactor: unify attention param field names

2026-08-09 20:23:58 +08:00

attn_paged_decode.cu

refactor: tidy attention params and launcher interfaces

2026-08-09 20:52:06 +08:00

attn_paged_prefill.cu

perf: flatten paged prefill tile dispatch

2026-08-09 23:12:53 +08:00

attn_prefill_split_q_mma.cuh

refactor: streamline Q block mapping

2026-08-10 08:40:18 +08:00

attn_prefill_split_q.cuh

refactor: streamline Q block mapping

2026-08-10 08:40:18 +08:00

attn_prefill.cu

refactor: tidy attention params and launcher interfaces

2026-08-09 20:52:06 +08:00

attn_warp_utils.cuh

perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

2026-07-27 00:35:34 +08:00

fp8_mm.cu

perf: output fp8 gemm in bf16 instead of fp32

2026-08-14 01:08:26 +08:00

rotary_emb.cu

perf: launch CUDA kernels on torch's current stream

2026-08-02 13:20:14 +08:00