vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe ¶

class FlashInferExperts(mk.FusedMoEPermuteExpertsUnpermute):

    def __init__(
        self,
        out_dtype: torch.dtype,
        quant_config: FusedMoEQuantConfig,
        ep_rank: int = 0,
        ep_size: int = 1,
        tp_rank: int = 0,
        tp_size: int = 1,
    ):
        super().__init__(quant_config)
        assert quant_config.quant_dtype in (
            "nvfp4", torch.float8_e4m3fn,
            None), ("Only nvfp4, fp8, bfloat16 and"
                    " float16 quantization are currently supported.")
        self.ep_rank = ep_rank
        self.ep_size = ep_size
        self.tp_rank = tp_rank
        self.tp_size = tp_size
        self.out_dtype = out_dtype

    @property
    def activation_formats(
        self
    ) -> tuple[mk.FusedMoEActivationFormat, mk.FusedMoEActivationFormat]:
        return (mk.FusedMoEActivationFormat.Standard,
                mk.FusedMoEActivationFormat.Standard)

    def supports_expert_map(self) -> bool:
        return False

    def supports_chunking(self) -> bool:
        # This refers to TP chunking; DP chunking is handled separately.
        return True

    def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
        return TopKWeightAndReduceNoOP()

    def workspace_shapes(
        self,
        a: torch.Tensor,
        aq: torch.Tensor,
        M: int,
        N: int,
        K: int,
        topk: int,
        global_num_experts: int,
        local_num_experts: int,
        expert_tokens_meta: Optional[mk.ExpertTokensMetadata],
    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...], torch.dtype]:
        # We use global_num_experts due to how moe_align_block_size handles
        # expert_maps.
        """
        Compute the shapes for the temporary and final outputs of the two gemms
        and activation in the fused expert function.  Since the gemms are
        independent, the workspace for the first gemm can be shared with the
        workspace for the last gemm.

        Returns a tuple of:
        - workspace13 shape tuple: must be large enough to hold the
          result of either expert gemm.
        - workspace2 shape tuple: must be large enough to hold the
          result of the activation function.
        - output shape tuple: must be exact size of the final gemm output.
        - Workspace type: The dtype to use for the workspace tensors.
        - Note: in order for activation chunking to work, the first dimension
          of each tuple must be the number of tokens.
        """
        aq_m, aq_n = aq.shape
        workspace2 = (0, )
        output_shape = (aq_m,
                        aq_n * 2) if self.quant_dtype == "nvfp4" else (aq_m,
                                                                       aq_n)
        workspace_dtype = a.dtype
        workspace1 = output_shape
        # The workspace is determined by `aq`, since it comes after any
        # potential communication op and is involved in the expert computation.
        return (workspace1, workspace2, output_shape, workspace_dtype)

    def apply(
        self,
        output: torch.Tensor,
        hidden_states: torch.Tensor,
        w1: torch.Tensor,
        w2: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        activation: str,
        global_num_experts: int,
        expert_map: Optional[torch.Tensor],
        a1q_scale: Optional[torch.Tensor],
        a2_scale: Optional[torch.Tensor],
        workspace13: Optional[torch.Tensor],
        workspace2: Optional[torch.Tensor],
        expert_tokens_meta: Optional[mk.ExpertTokensMetadata],
        apply_router_weight_on_input: Optional[bool],
    ):

        assert activation == "silu", ("Only activation silu is supported in "
                                      "FlashInferExperts")

        if self.quant_dtype == torch.float8_e4m3fn:
            quant_scales = [
                self.g1_alphas, self.a2_gscale, self.g2_alphas, self.a1_gscale
            ]

            a1q_scale = None  # not passing input_sf in fp8
            fc1_expert_weights = w1
            fc2_expert_weights = w2
        elif self.quant_dtype == "nvfp4":
            # Ensure w1_scale and w2_scale are not None before calling view
            assert self.w1_scale is not None and self.w2_scale is not None, (
                "w1_scale and w2_scale must not "
                "be None for FlashInferExperts")
            # Flashinfer CUTLASS kernel takes scalar global scales,
            # min because inv_scale.
            quant_scales = [
                self.a1_gscale,
                self.w1_scale.view(torch.int32),
                self.g1_alphas,
                self.a2_gscale,
                self.w2_scale.view(torch.int32),
                self.g2_alphas,
            ]
            # FlashInfer API requires weight to be long for nvfp4
            fc1_expert_weights = w1.view(torch.long)
            fc2_expert_weights = w2.view(torch.long)
        else:
            quant_scales = None
            a1q_scale = None
            fc1_expert_weights = w1
            fc2_expert_weights = w2

        _ = flashinfer_cutlass_fused_moe(
            input=hidden_states,
            token_selected_experts=topk_ids.to(torch.int),
            token_final_scales=topk_weights,
            fc1_expert_weights=fc1_expert_weights,
            fc2_expert_weights=fc2_expert_weights,
            output_dtype=self.out_dtype,
            quant_scales=quant_scales,
            input_sf=a1q_scale,
            tp_size=self.tp_size,
            tp_rank=self.tp_rank,
            ep_size=self.ep_size,
            ep_rank=self.ep_rank,
            output=output,
        )