From efd14cfb66a316cd1e028d179947a4dcedd5ef41 Mon Sep 17 00:00:00 2001 From: Ziyi Xu <76582120+DarkSharpness@users.noreply.github.com> Date: Sun, 11 Oct 2026 19:03:19 +0800 Subject: [PATCH] [DeepSeek-V4.1] Use num_stages=1 for the hc_mix stats partial kernel (#43070) Co-authored-by: Claude Opus 5.5 Co-authored-by: Yuan Luo <4908075+yuan-luo@users.noreply.github.com> --- python/sglang/kernels/ops/layernorm/mhc.py | 17 +++++------------ 1 file changed, 5 insertions(+), 12 deletions(-) diff --git a/python/sglang/kernels/ops/layernorm/mhc.py b/python/sglang/kernels/ops/layernorm/mhc.py index dae2556e70..bc7c7035c3 100644 --- a/python/sglang/kernels/ops/layernorm/mhc.py +++ b/python/sglang/kernels/ops/layernorm/mhc.py @@ -2303,9 +2303,9 @@ _HC_MIX_BLOCK_K = 64 _HC_MIX_NUM_WARPS = 4 # Triton on AMD has no tf32x3 (gfx942 adds only tf32); "ieee" is the fp32 MFMA _HC_MIX_DOT_PRECISION = "ieee" if _is_hip else "tf32x3" -# num_stages only reorders memory issue, not arithmetic; 2 is enough to cover the -# short k_per_slice loop (K=20480 gives 80 slices, i.e. 4 BLOCK_K tiles per CTA). -_HC_MIX_NUM_STAGES = 2 +# num_stages only reorders memory issue, not arithmetic; AMD keeps 2, not measured. +# Measured on B300 and H200: 1 stage beats 2-4 by 2-5x at every M in 1..4096. +_HC_MIX_NUM_STAGES = 2 if _is_hip else 1 # BLOCK_M 8/16/32 preserve each row's K reduction order; thresholds were measured on GB300. # Keep BLOCK_M below 64, where Triton lowers tf32x3 to plain TF32 and changes rounding. @@ -2326,13 +2326,6 @@ def _block_m_for(m: int) -> int: return _HC_MIX_BLOCK_M -def _num_stages_for(m: int, k: int) -> int: - # GB300 verify batches benefit from a smaller shared-memory footprint. - if get_platform().is_blackwell and k == 20480 and 64 <= m <= 384: - return 1 - return _HC_MIX_NUM_STAGES - - def _num_slices_for(k: int) -> int: """Slice count depends only on K, never on batch size M.""" blocks = k // _HC_MIX_BLOCK_K @@ -2383,7 +2376,7 @@ def hc_mix_stats(x_flat: torch.Tensor, hc_fn: torch.Tensor, eps: float) -> torch BLOCK_K=_HC_MIX_BLOCK_K, DOT_PRECISION=_HC_MIX_DOT_PRECISION, num_warps=_HC_MIX_NUM_WARPS, - num_stages=_num_stages_for(m, k), + num_stages=_HC_MIX_NUM_STAGES, ) _hc_mix_stats_reduce_kernel[(grid_m,)]( part_mix, @@ -2515,7 +2508,7 @@ def hc_mix_stats_sinkhorn( BLOCK_K=_HC_MIX_BLOCK_K, DOT_PRECISION=_HC_MIX_DOT_PRECISION, num_warps=_HC_MIX_NUM_WARPS, - num_stages=_num_stages_for(m, k), + num_stages=_HC_MIX_NUM_STAGES, ) if _is_hip: hc_mix_reduce_sinkhorn_vec(