From 49bafb58cd6eb3a4feff035ce33dd72bb73addfd Mon Sep 17 00:00:00 2001 From: Chunan Zeng Date: Tue, 6 Oct 2026 07:47:47 -0700 Subject: [PATCH] perf(mla): use native SATFINITE FP8 conversion in MLA KV/Q preparation (#42659) Co-authored-by: Claude Opus 5.5 Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> --- .../kernels/jit/csrc/elementwise/set_mla_kv_concat_q.cuh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/kernels/jit/csrc/elementwise/set_mla_kv_concat_q.cuh b/python/sglang/kernels/jit/csrc/elementwise/set_mla_kv_concat_q.cuh index 4bbf245967..4c0f8ae2cb 100644 --- a/python/sglang/kernels/jit/csrc/elementwise/set_mla_kv_concat_q.cuh +++ b/python/sglang/kernels/jit/csrc/elementwise/set_mla_kv_concat_q.cuh @@ -325,10 +325,10 @@ struct SetMlaKVConcatQFp8Params { int32_t qo_stride_1; }; -// 2x bf16 -> 2x fp8 e4m3, float-mediated cvt.rn NOSAT (matches aten: overflow -> NaN). +// Native packed E4M3 conversion: clamp overflow/infinity to +/-448, preserve NaN. SGL_DEVICE uint16_t bf16x2_to_fp8x2(const bf16x2_t v) { const float2 f = __bfloat1622float2(v); - return __nv_cvt_float2_to_fp8x2(f, __NV_NOSAT, __NV_E4M3); + return __nv_cvt_float2_to_fp8x2(f, __NV_SATFINITE, __NV_E4M3); } // Convert 8 bf16 (one int4 load) to 8 fp8 packed in a uint2.