Skip to content

vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize

FlashInferCutlassMoEPrepareAndFinalize

Bases: FusedMoEPrepareAndFinalize

Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
class FlashInferCutlassMoEPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):

    def __init__(
        self,
        use_dp: bool,
        a1_gscale: Optional[torch.Tensor],
        num_dispatchers: int = 1,
    ):
        super().__init__()
        self.num_dispatchers_ = num_dispatchers
        self.use_dp = use_dp
        self.a1_gscale = a1_gscale
        self.local_tokens = None

    @property
    def activation_format(self) -> mk.FusedMoEActivationFormat:
        return mk.FusedMoEActivationFormat.Standard

    def max_num_tokens_per_rank(self) -> Optional[int]:
        return None

    def topk_indices_dtype(self) -> Optional[torch.dtype]:
        return None

    def num_dispatchers(self) -> int:
        return self.num_dispatchers_

    def prepare(
        self,
        a1: torch.Tensor,
        a1_scale: Optional[torch.Tensor],  # Not used
        a2_scale: Optional[torch.Tensor],  # Not used
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        num_experts: int,
        expert_map: Optional[torch.Tensor],
        apply_router_weight_on_input: bool,
        # TODO(bnell): use quant_config + scales instead of ctor args
        quant_config: FusedMoEQuantConfig,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor],
               Optional[mk.ExpertTokensMetadata], Optional[torch.Tensor],
               Optional[torch.Tensor]]:

        if apply_router_weight_on_input:
            topk = topk_ids.size(1)
            # TODO: this only works for topK=1, will need to update for topK>1
            assert topk == 1, \
                "apply_router_weight_on_input is only implemented for topk=1"
            a1.mul_(topk_weights.to(a1.dtype))

        a1q, a1q_scale = moe_kernel_quantize_input(
            a1,
            self.a1_gscale,
            quant_config.quant_dtype,
            quant_config.per_act_token_quant,
            quant_config.block_shape,
            # Swizzling after communication
            is_fp4_scale_swizzled=not self.use_dp,
        )
        if self.use_dp:
            topk_weights, topk_ids, a1q, a1q_scale = \
                get_dp_group().all_gatherv(
                    [topk_weights, topk_ids, a1q, a1q_scale],
                    dim=0,
                    sizes=get_local_sizes(),
                )
            a1_m, a1_n = a1q.shape
            a1q_scale = nvfp4_block_scale_interleave(a1q_scale)

        return a1q, a1q_scale, None, topk_ids, topk_weights

    def finalize(self, output: torch.Tensor, fused_expert_output: torch.Tensor,
                 topk_weights: torch.Tensor, topk_ids: torch.Tensor,
                 apply_router_weight_on_input: bool,
                 weight_and_reduce_impl: mk.TopKWeightAndReduce) -> None:

        if self.use_dp:
            fused_expert_output = get_dp_group().reduce_scatterv(
                fused_expert_output, dim=0, sizes=get_local_sizes())
        output.copy_(fused_expert_output)

a1_gscale instance-attribute 

a1_gscale = a1_gscale

activation_format property 

activation_format: FusedMoEActivationFormat

local_tokens instance-attribute 

local_tokens = None

num_dispatchers_ instance-attribute 

num_dispatchers_ = num_dispatchers

use_dp instance-attribute 

use_dp = use_dp

__init__ 

__init__(
    use_dp: bool,
    a1_gscale: Optional[Tensor],
    num_dispatchers: int = 1,
)
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
22
23
24
25
26
27
28
29
30
31
32
def __init__(
    self,
    use_dp: bool,
    a1_gscale: Optional[torch.Tensor],
    num_dispatchers: int = 1,
):
    super().__init__()
    self.num_dispatchers_ = num_dispatchers
    self.use_dp = use_dp
    self.a1_gscale = a1_gscale
    self.local_tokens = None

finalize 

finalize(
    output: Tensor,
    fused_expert_output: Tensor,
    topk_weights: Tensor,
    topk_ids: Tensor,
    apply_router_weight_on_input: bool,
    weight_and_reduce_impl: TopKWeightAndReduce,
) -> None
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
91
92
93
94
95
96
97
98
99
def finalize(self, output: torch.Tensor, fused_expert_output: torch.Tensor,
             topk_weights: torch.Tensor, topk_ids: torch.Tensor,
             apply_router_weight_on_input: bool,
             weight_and_reduce_impl: mk.TopKWeightAndReduce) -> None:

    if self.use_dp:
        fused_expert_output = get_dp_group().reduce_scatterv(
            fused_expert_output, dim=0, sizes=get_local_sizes())
    output.copy_(fused_expert_output)

max_num_tokens_per_rank 

max_num_tokens_per_rank() -> Optional[int]
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
38
39
def max_num_tokens_per_rank(self) -> Optional[int]:
    return None

num_dispatchers 

num_dispatchers() -> int
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
44
45
def num_dispatchers(self) -> int:
    return self.num_dispatchers_

prepare 

prepare(
    a1: Tensor,
    a1_scale: Optional[Tensor],
    a2_scale: Optional[Tensor],
    topk_weights: Tensor,
    topk_ids: Tensor,
    num_experts: int,
    expert_map: Optional[Tensor],
    apply_router_weight_on_input: bool,
    quant_config: FusedMoEQuantConfig,
) -> tuple[
    Tensor,
    Optional[Tensor],
    Optional[ExpertTokensMetadata],
    Optional[Tensor],
    Optional[Tensor],
]
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
def prepare(
    self,
    a1: torch.Tensor,
    a1_scale: Optional[torch.Tensor],  # Not used
    a2_scale: Optional[torch.Tensor],  # Not used
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    num_experts: int,
    expert_map: Optional[torch.Tensor],
    apply_router_weight_on_input: bool,
    # TODO(bnell): use quant_config + scales instead of ctor args
    quant_config: FusedMoEQuantConfig,
) -> tuple[torch.Tensor, Optional[torch.Tensor],
           Optional[mk.ExpertTokensMetadata], Optional[torch.Tensor],
           Optional[torch.Tensor]]:

    if apply_router_weight_on_input:
        topk = topk_ids.size(1)
        # TODO: this only works for topK=1, will need to update for topK>1
        assert topk == 1, \
            "apply_router_weight_on_input is only implemented for topk=1"
        a1.mul_(topk_weights.to(a1.dtype))

    a1q, a1q_scale = moe_kernel_quantize_input(
        a1,
        self.a1_gscale,
        quant_config.quant_dtype,
        quant_config.per_act_token_quant,
        quant_config.block_shape,
        # Swizzling after communication
        is_fp4_scale_swizzled=not self.use_dp,
    )
    if self.use_dp:
        topk_weights, topk_ids, a1q, a1q_scale = \
            get_dp_group().all_gatherv(
                [topk_weights, topk_ids, a1q, a1q_scale],
                dim=0,
                sizes=get_local_sizes(),
            )
        a1_m, a1_n = a1q.shape
        a1q_scale = nvfp4_block_scale_interleave(a1q_scale)

    return a1q, a1q_scale, None, topk_ids, topk_weights

topk_indices_dtype 

topk_indices_dtype() -> Optional[dtype]
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
41
42
def topk_indices_dtype(self) -> Optional[torch.dtype]:
    return None

get_local_sizes 

get_local_sizes()
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
16
17
def get_local_sizes():
    return get_forward_context().dp_metadata.get_chunk_sizes_across_dp_rank()