[MoE] Plumb gemm1_alpha/beta/clamp_limit into TRT-LLM FP8 MoE (#45723)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
parent
d3229431f9
commit
fa248139a0
+2
@@ -153,6 +153,8 @@ class CompressedTensorsW8A8Mxfp8MoEMethod(CompressedTensorsMoEMethod):
|
||||
a2_scale=layer.w2_input_scale,
|
||||
block_shape=self.weight_block_size,
|
||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||
)
|
||||
|
||||
def maybe_make_prepare_finalize(
|
||||
|
||||
@@ -224,6 +224,8 @@ class Mxfp8OnlineMoEMethod(OnlineMoEMethodBase):
|
||||
w2_bias=getattr(layer, "w2_bias", None),
|
||||
block_shape=self.weight_block_size,
|
||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||
)
|
||||
|
||||
def process_weights_after_loading(self, layer: Module) -> None:
|
||||
|
||||
@@ -33,6 +33,8 @@ def activation_to_flashinfer_type(activation: MoEActivation) -> "ActivationType"
|
||||
MoEActivation.SILU_NO_MUL: ActivationType.Silu,
|
||||
MoEActivation.GELU_NO_MUL: ActivationType.Gelu,
|
||||
MoEActivation.SILU: ActivationType.Swiglu,
|
||||
# SwiGLU-OAI uses Swiglu; the OAI alpha/beta/clamp come from gemm1_* args.
|
||||
MoEActivation.SWIGLUOAI_UNINTERLEAVE: ActivationType.Swiglu,
|
||||
MoEActivation.GELU: ActivationType.Geglu,
|
||||
MoEActivation.GELU_TANH: ActivationType.Geglu,
|
||||
MoEActivation.RELU2_NO_MUL: ActivationType.Relu2,
|
||||
|
||||
Reference in New Issue
Block a user