From 16e353d93e4e6d870623bee3fe1ea62936475886 Mon Sep 17 00:00:00 2001 From: iridiumine <42236072+iridiumine@users.noreply.github.com> Date: Wed, 23 Sep 2026 21:02:54 +0800 Subject: [PATCH] [NPU] Skip fused gmm1+swiglu for swiglu_limit (SiLU-with-clamp) checkpoints (#40438) --- .../srt/layers/moe/moe_runner/ascend.py | 26 ++++++++++++++++--- 1 file changed, 22 insertions(+), 4 deletions(-) diff --git a/python/sglang/srt/layers/moe/moe_runner/ascend.py b/python/sglang/srt/layers/moe/moe_runner/ascend.py index 82ebb0ffc0..082276784d 100644 --- a/python/sglang/srt/layers/moe/moe_runner/ascend.py +++ b/python/sglang/srt/layers/moe/moe_runner/ascend.py @@ -28,11 +28,29 @@ from sglang.srt.hardware_backend.npu.quantization.moe_methods import ( ) -def _uses_fused_gmm1(kernel) -> bool: +def _uses_fused_gmm1(kernel, config: MoeRunnerConfig) -> bool: """Whether gmm1 runs matmul+swiglu+requant fused (no separate activation).""" + # The fused gmm1 kernel (npu_grouped_matmul_swiglu_quant_v2) applies plain + # SiLU with no clamp. Models whose activation is SiLU-with-clamp + # (swiglu_limit, e.g. DSV4) must keep the unfused path so the clamp is + # applied (NPUSwigluMxfp8Quant / NPUSwigluStepAndMul below). + has_clamp = config.swiglu_limit is not None and config.swiglu_limit > 0 + + if not isinstance(kernel, (NPUMXFP8MoEMethod, NPUW4A8MXFP4MoEMethod)): + return False + if has_clamp: + # MXFP8 has no unfused gmm1 path at all, so a swiglu_limit + # checkpoint cannot be served by it. + if isinstance(kernel, NPUMXFP8MoEMethod): + raise NotImplementedError( + "NPUMXFP8MoEMethod has no unfused gmm1 path, and the fused " + "gmm1 kernel applies SiLU without clamp — a swiglu_limit " + "checkpoint cannot be served by this method." + ) + return False # W4A8MXFP4: fall back to unfused to preserve the clamp if isinstance(kernel, NPUMXFP8MoEMethod): return True - return isinstance(kernel, NPUW4A8MXFP4MoEMethod) and kernel.use_fused_gmm1 + return kernel.use_fused_gmm1 from sglang.srt.layers.moe.moe_runner.base import ( @@ -102,7 +120,7 @@ class AscendRunnerCore(MoeRunnerCore): kernel = config.layer.w2_kernel - if _uses_fused_gmm1(kernel): + if _uses_fused_gmm1(kernel, config): # Fused methods (MXFP8; MXFP4 W4A8 via use_fused_gmm1) fold # gate/up + swiglu + requant into gmm1, so there is no separate # activation step — run() skips it. Left None on purpose so that @@ -201,7 +219,7 @@ class AscendRunnerCore(MoeRunnerCore): w13_kernel = self.config.layer.w13_kernel - if _uses_fused_gmm1(w13_kernel): + if _uses_fused_gmm1(w13_kernel, self.config): # --- w13 projection + activation, fused into one kernel --- # The fused gmm1 returns activations already requantised for gmm2, # so there is no separate activation step to run.