vllm.model_executor.layers.fused_moe ¶

class BatchedDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
    def __init__(
        self,
        max_num_tokens: int,
        num_dispatchers: int,
        quant_config: FusedMoEQuantConfig,
    ):
        """
        max_num_tokens: Maximum number of tokens from a DP Rank
        num_dispatchers: The number of DP dispatchers.
        quant_config: Quantization configuration
        """
        super().__init__(quant_config)
        assert self.block_shape == get_mk_alignment_for_contiguous_layout()
        self.max_num_tokens = max_num_tokens
        self.num_dispatchers = num_dispatchers

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

    def supports_chunking(self) -> bool:
        return False

    def supports_expert_map(self) -> bool:
        return False

    def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
        # Let PrepareAndFinalize::finalize() decide the impl.
        return TopKWeightAndReduceDelegate()

    def workspace_shapes(
        self,
        M: int,
        N: int,
        K: int,
        topk: int,
        global_num_experts: int,
        local_num_experts: int,
        expert_tokens_meta: mk.ExpertTokensMetadata | None,
    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
        # FIXME (varun): We should be able to dispatch only from the leader
        # DP ranks in the case of TP > 1. At the moment, all the Ranks
        # end up sending their tokens. This needs to be fixed.
        num_dispatchers = self.num_dispatchers
        num_experts = local_num_experts
        max_num_tokens = M if self.max_num_tokens is None else self.max_num_tokens
        workspace13 = (num_experts, max_num_tokens * num_dispatchers, max(K, N))
        workspace2 = (num_experts, max_num_tokens * num_dispatchers, (N // 2))
        output = (num_experts, max_num_tokens * num_dispatchers, K)
        return (workspace13, workspace2, output)

    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: torch.Tensor | None,
        a1q_scale: torch.Tensor | None,
        a2_scale: torch.Tensor | None,
        workspace13: torch.Tensor,
        workspace2: torch.Tensor,
        expert_tokens_meta: mk.ExpertTokensMetadata | None,
        apply_router_weight_on_input: bool,
    ):
        assert expert_tokens_meta is not None
        expert_num_tokens = expert_tokens_meta.expert_num_tokens

        assert hidden_states.ndim == 3
        assert self.block_shape is not None

        a1q = hidden_states
        _, N, K = w1.size()

        assert w2.size(1) == K

        E, max_num_tokens, N, K, _ = self.moe_problem_size(
            hidden_states, w1, w2, topk_ids
        )

        workspace1 = _resize_cache(workspace13, (E, max_num_tokens, N))

        # (from deepgemm docs) : A value hint (which is a value on CPU)
        # for the M expectation of each batch, correctly setting this value
        # may lead to better performance.
        expected_m = max_num_tokens
        fp8_m_grouped_gemm_nt_masked(
            (a1q, a1q_scale),
            (w1, self.w1_scale),
            workspace1,
            expert_num_tokens,
            expected_m,
        )

        a2q, a2q_scale = persistent_masked_m_silu_mul_quant(
            workspace1, expert_num_tokens
        )

        fp8_m_grouped_gemm_nt_masked(
            (a2q, a2q_scale), (w2, self.w2_scale), output, expert_num_tokens, expected_m
        )

class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
    """
    A Triton based MoE expert class that operates on expert batched format,
    i.e. E x max_num_tokens x K.  This is the format that the pplx
    dispatch/combine kernels use.
    """

    def __init__(
        self,
        max_num_tokens: int,
        num_dispatchers: int,
        quant_config: FusedMoEQuantConfig,
    ):
        super().__init__(quant_config)
        assert not self.quant_config.use_int8_w8a8, "NYI"
        assert not self.quant_config.use_int8_w8a16, "NYI"
        assert not self.quant_config.use_int4_w4a16, "NYI"
        assert self.quant_config.ocp_mx_scheme is None, "NYI"
        assert max_num_tokens > 0
        assert num_dispatchers > 0
        self.max_num_tokens = max_num_tokens
        self.num_dispatchers = num_dispatchers

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

    def supports_chunking(self) -> bool:
        return False

    def supports_expert_map(self) -> bool:
        return False

    def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
        # Let PrepareAndFinalize::finalize() decide the impl.
        return TopKWeightAndReduceDelegate()

    def workspace_shapes(
        self,
        M: int,
        N: int,
        K: int,
        topk: int,
        global_num_experts: int,
        local_num_experts: int,
        expert_tokens_meta: mk.ExpertTokensMetadata | None,
    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
        num_dp = self.num_dispatchers
        num_experts = local_num_experts
        max_num_tokens = self.max_num_tokens
        workspace13 = (num_experts, max_num_tokens * num_dp, max(K, N))
        workspace2 = (num_experts, max_num_tokens * num_dp, (N // 2))
        output = (num_experts, max_num_tokens * num_dp, K)
        return (workspace13, workspace2, output)

    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: torch.Tensor | None,
        a1q_scale: torch.Tensor | None,
        a2_scale: torch.Tensor | None,
        workspace13: torch.Tensor,
        workspace2: torch.Tensor,
        expert_tokens_meta: mk.ExpertTokensMetadata | None,
        apply_router_weight_on_input: bool,
    ):
        # Check constraints.
        if self.quant_config.use_int4_w4a16:
            assert hidden_states.size(-1) // 2 == w1.size(2), "Hidden size mismatch"
        else:
            assert hidden_states.size(-1) == w1.size(2), (
                f"Hidden size mismatch {hidden_states.size(-1)} != {w1.size(2)}"
            )

        assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
        assert w1.stride(-1) == 1, "Stride of last dimension must be 1"
        assert w2.stride(-1) == 1, "Stride of last dimension must be 1"
        assert hidden_states.dtype in [
            torch.float32,
            torch.float16,
            torch.bfloat16,
            torch.float8_e4m3fn,
        ]
        assert expert_tokens_meta is not None

        expert_num_tokens = expert_tokens_meta.expert_num_tokens

        E, max_num_tokens, N, K, top_k_num = self.moe_problem_size(
            hidden_states, w1, w2, topk_ids
        )

        assert w1.size(0) == E
        assert w2.size(0) == E

        config_dtype = self.quant_config.config_name(hidden_states.dtype)

        config = try_get_optimal_moe_config(
            w1.size(),
            w2.size(),
            top_k_num,
            config_dtype,
            max_num_tokens,
            block_shape=self.block_shape,
        )

        if hidden_states.dtype == torch.bfloat16:
            compute_type = tl.bfloat16
        elif hidden_states.dtype == torch.float16:
            compute_type = tl.float16
        elif hidden_states.dtype == torch.float32:
            compute_type = tl.float32
        elif hidden_states.dtype == torch.float8_e4m3fn:
            compute_type = tl.bfloat16
        else:
            raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")

        # We can reuse the memory between these because by the time we need
        # cache3, we're done with cache1
        intermediate_cache1 = _resize_cache(workspace13, (E, max_num_tokens, N))
        intermediate_cache2 = _resize_cache(workspace2, (E, max_num_tokens, N // 2))

        # TODO(bnell): should this be done for any quantized type?
        if self.quant_config.use_fp8_w8a8:
            intermediate_cache1.fill_(0)

        a1q_scale = normalize_batched_scales_shape(a1q_scale, E)

        # MM1
        invoke_moe_batched_triton_kernel(
            A=hidden_states,
            B=w1,
            C=intermediate_cache1,
            expert_num_tokens=expert_num_tokens,
            compute_type=compute_type,
            A_scale=a1q_scale,
            B_scale=self.w1_scale,
            B_zp=self.w1_zp,
            use_fp8_w8a8=self.quant_config.use_fp8_w8a8,
            use_int8_w8a16=self.quant_config.use_int8_w8a16,
            use_int4_w4a16=self.quant_config.use_int4_w4a16,
            config=config,
            per_act_token_quant=self.per_act_token_quant,
            block_shape=self.block_shape,
        )

        intermediate_cache2.fill_(0)

        # TODO (bnell): use triton utility from batched deep gemm.
        self.activation(
            activation,
            intermediate_cache2.view(-1, N // 2),
            intermediate_cache1.view(-1, N),
        )

        qintermediate_cache2, a2q_scale = batched_moe_kernel_quantize_input(
            intermediate_cache2,
            a2_scale,
            max_num_tokens,
            E,
            N,
            expert_num_tokens,
            self.quant_dtype,
            self.per_act_token_quant,
            self.block_shape,
        )

        invoke_moe_batched_triton_kernel(
            A=qintermediate_cache2,
            B=w2,
            C=output,
            expert_num_tokens=expert_num_tokens,
            compute_type=compute_type,
            A_scale=a2q_scale,
            B_scale=self.w2_scale,
            B_zp=self.w2_zp,
            use_fp8_w8a8=self.quant_config.use_fp8_w8a8,
            use_int8_w8a16=self.quant_config.use_int8_w8a16,
            use_int4_w4a16=self.quant_config.use_int4_w4a16,
            config=config,
            per_act_token_quant=self.per_act_token_quant,
            block_shape=self.block_shape,
        )

class BatchedTritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
    def __init__(
        self,
        max_num_tokens: int,
        num_dispatchers: int,
        quant_config: FusedMoEQuantConfig,
        allow_deep_gemm: bool = False,
    ):
        super().__init__(quant_config)

        self.batched_triton_experts = BatchedTritonExperts(
            max_num_tokens=max_num_tokens,
            num_dispatchers=num_dispatchers,
            quant_config=self.quant_config,
        )

        self.allow_deep_gemm = (
            allow_deep_gemm
            and self.quant_config.use_fp8_w8a8
            and self.block_shape == get_mk_alignment_for_contiguous_layout()
        )

        self.batched_deep_gemm_experts = (
            BatchedDeepGemmExperts(
                max_num_tokens=max_num_tokens,
                num_dispatchers=num_dispatchers,
                quant_config=self.quant_config,
            )
            if self.allow_deep_gemm
            else None
        )

        assert (
            self.batched_deep_gemm_experts is not None
            or self.batched_triton_experts is not None
        )

    @property
    def activation_formats(
        self,
    ) -> tuple[mk.FusedMoEActivationFormat, mk.FusedMoEActivationFormat]:
        if self.batched_triton_experts is not None:
            assert (
                self.batched_deep_gemm_experts is None
                or self.batched_deep_gemm_experts.activation_formats
                == self.batched_triton_experts.activation_formats
            )
            return self.batched_triton_experts.activation_formats
        else:
            assert self.batched_deep_gemm_experts is not None
            return self.batched_deep_gemm_experts.activation_formats

    def supports_chunking(self) -> bool:
        bdge = self.batched_deep_gemm_experts
        bte = self.batched_triton_experts
        return (bdge is None or bdge.supports_chunking()) and (
            bte is None or bte.supports_chunking()
        )

    def supports_expert_map(self) -> bool:
        bdge = self.batched_deep_gemm_experts
        bte = self.batched_triton_experts
        return (bdge is None or bdge.supports_expert_map()) and (
            bte is None or bte.supports_expert_map()
        )

    def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
        bdge = self.batched_deep_gemm_experts
        bte = self.batched_triton_experts
        bdge_war = bdge.finalize_weight_and_reduce_impl() if bdge else None
        bte_war = bte.finalize_weight_and_reduce_impl() if bte else None
        is_bdge_war = bdge_war is not None
        is_bte_war = bte_war is not None

        if is_bdge_war and is_bte_war:
            assert bdge_war == bte_war, (
                "Both implementations should agree on WeightAndReduce impls. "
                f"Got bdge_war: {bdge_war}, and bte_war: {bte_war}"
            )

        if bdge_war is not None:
            return bdge_war

        assert bte_war is not None
        return bte_war

    def workspace_dtype(self, act_dtype: torch.dtype) -> torch.dtype:
        return act_dtype

    def workspace_shapes(
        self,
        M: int,
        N: int,
        K: int,
        topk: int,
        global_num_experts: int,
        local_num_experts: int,
        expert_tokens_metadata: mk.ExpertTokensMetadata | None,
    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
        # Note: the deep gemm workspaces are strictly larger than the triton
        # workspaces so we can be pessimistic here and allocate for DeepGemm
        # even if we fall back to triton later, e.g. if expert maps are set.
        if self.allow_deep_gemm:
            assert self.batched_deep_gemm_experts is not None
            return self.batched_deep_gemm_experts.workspace_shapes(
                M,
                N,
                K,
                topk,
                global_num_experts,
                local_num_experts,
                expert_tokens_metadata,
            )
        else:
            assert self.batched_triton_experts is not None
            return self.batched_triton_experts.workspace_shapes(
                M,
                N,
                K,
                topk,
                global_num_experts,
                local_num_experts,
                expert_tokens_metadata,
            )

    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: torch.Tensor | None,
        a1q_scale: torch.Tensor | None,
        a2_scale: torch.Tensor | None,
        workspace13: torch.Tensor,
        workspace2: torch.Tensor,
        expert_tokens_meta: mk.ExpertTokensMetadata | None,
        apply_router_weight_on_input: bool,
    ):
        experts = (
            self.batched_deep_gemm_experts
            if self.allow_deep_gemm
            else self.batched_triton_experts
        )
        assert experts is not None
        experts.apply(
            output,
            hidden_states,
            w1,
            w2,
            topk_weights,
            topk_ids,
            activation,
            global_num_experts,
            expert_map,
            a1q_scale,
            a2_scale,
            workspace13,
            workspace2,
            expert_tokens_meta,
            apply_router_weight_on_input,
        )