[Fix] [NPU] kimi-k3: skip fused gmm1 for situ activation in Ascend runner (#43837)

This commit is contained in:
zhaozx-cn
2026-10-19 19:30:41 +08:00
committed by GitHub
parent 0025f93944
commit ed0b8f433f
2 changed files with 5 additions and 0 deletions
@@ -12,11 +12,13 @@ from sglang.kernels.jit.utils import (
load_jit,
make_cpp_args,
)
from sglang.srt.utils import is_npu
if TYPE_CHECKING:
from tvm_ffi.module import Module
_THREADS: int = 256
_is_npu = is_npu()
@cache_once
@@ -35,6 +37,7 @@ def covered(x: torch.Tensor, gate: torch.Tensor) -> bool:
return (
x.is_cuda
and gate.is_cuda
and not _is_npu
and x.dtype == torch.bfloat16
and gate.dtype == torch.bfloat16
and x.shape == gate.shape
@@ -36,6 +36,8 @@ def _uses_fused_gmm1(kernel, config: MoeRunnerConfig) -> bool:
# applied (NPUSwigluMxfp8Quant / NPUSwigluStepAndMul below).
has_clamp = config.swiglu_limit is not None and config.swiglu_limit > 0
if config.activation == "situ":
return False
if not isinstance(kernel, (NPUMXFP8MoEMethod, NPUW4A8MXFP4MoEMethod)):
return False
if has_clamp: