minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Moe] Fix flashinfer_trtllm silently dropping swiglu_limit clamped SwiGLU activation (#39920)

Browse Source
This commit is contained in:
Shijin Zhang
2026-09-17 00:25:17 -07:00
committed by GitHub
parent f3c0256771
commit aebae58b8c
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/fp8.py
+6 -1
View File
@@ -2358,10 +2358,15 @@ class Fp8MoEMethod(FusedMoEMethodBase):
"""Materialize optional TRT-LLM SwiGLU parameters once per expert.""" """Materialize optional TRT-LLM SwiGLU parameters once per expert."""
num_experts = int(layer.num_local_experts) num_experts = int(layer.num_local_experts)
device = layer.w13_weight.device device = layer.w13_weight.device
clamp_limit = (
self.moe_runner_config.gemm1_clamp_limit
if self.moe_runner_config.gemm1_clamp_limit is not None
else self.moe_runner_config.swiglu_limit
)
for name, value in ( for name, value in (
("gemm1_alpha", self.moe_runner_config.gemm1_alpha), ("gemm1_alpha", self.moe_runner_config.gemm1_alpha),
("gemm1_beta", self.moe_runner_config.gemm1_beta), ("gemm1_beta", self.moe_runner_config.gemm1_beta),
("gemm1_clamp_limit", self.moe_runner_config.gemm1_clamp_limit), ("gemm1_clamp_limit", clamp_limit),
): ):
tensor = ( tensor = (
None None