Skip to content

vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize

FlashInferAllGatherMoEPrepareAndFinalize

Bases: FlashInferCutlassMoEPrepareAndFinalize

Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
class FlashInferAllGatherMoEPrepareAndFinalize(FlashInferCutlassMoEPrepareAndFinalize):
    def __init__(
        self,
        use_dp: bool,
        num_dispatchers: int = 1,
    ):
        super().__init__(use_dp, num_dispatchers)

    def prepare(
        self,
        a1: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        num_experts: int,
        expert_map: torch.Tensor | None,
        apply_router_weight_on_input: bool,
        quant_config: FusedMoEQuantConfig,
    ) -> mk.PrepareResultType:
        self._apply_router_weight_on_input(
            a1, topk_weights, topk_ids, apply_router_weight_on_input
        )

        a1q, a1q_scale = moe_kernel_quantize_input(
            a1,
            quant_config.a1_gscale,
            quant_config.quant_dtype,
            quant_config.per_act_token_quant,
            quant_config.block_shape,
            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(),
            )
            if quant_config.quant_dtype == "nvfp4":
                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:
        assert isinstance(weight_and_reduce_impl, TopKWeightAndReduceNoOP)

        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)

__init__ 

__init__(use_dp: bool, num_dispatchers: int = 1)
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
153
154
155
156
157
158
def __init__(
    self,
    use_dp: bool,
    num_dispatchers: int = 1,
):
    super().__init__(use_dp, num_dispatchers)

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 
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
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:
    assert isinstance(weight_and_reduce_impl, TopKWeightAndReduceNoOP)

    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)

prepare 

prepare(
    a1: Tensor,
    topk_weights: Tensor,
    topk_ids: Tensor,
    num_experts: int,
    expert_map: Tensor | None,
    apply_router_weight_on_input: bool,
    quant_config: FusedMoEQuantConfig,
) -> PrepareResultType
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
def prepare(
    self,
    a1: torch.Tensor,
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    num_experts: int,
    expert_map: torch.Tensor | None,
    apply_router_weight_on_input: bool,
    quant_config: FusedMoEQuantConfig,
) -> mk.PrepareResultType:
    self._apply_router_weight_on_input(
        a1, topk_weights, topk_ids, apply_router_weight_on_input
    )

    a1q, a1q_scale = moe_kernel_quantize_input(
        a1,
        quant_config.a1_gscale,
        quant_config.quant_dtype,
        quant_config.per_act_token_quant,
        quant_config.block_shape,
        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(),
        )
        if quant_config.quant_dtype == "nvfp4":
            a1q_scale = nvfp4_block_scale_interleave(a1q_scale)

    return a1q, a1q_scale, None, topk_ids, topk_weights

FlashInferAllToAllMoEPrepareAndFinalize

Bases: FlashInferCutlassMoEPrepareAndFinalize

FlashInfer implementation using AllToAll communication.

Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
 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
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
class FlashInferAllToAllMoEPrepareAndFinalize(FlashInferCutlassMoEPrepareAndFinalize):
    """FlashInfer implementation using AllToAll communication."""

    def __init__(
        self,
        use_dp: bool,
        num_dispatchers: int = 1,
    ):
        super().__init__(use_dp, num_dispatchers)
        self.alltoall_info = None

        # Initialize all2all_manager only for DP case
        self.all2all_manager = None
        if self.use_dp:
            self.all2all_manager = get_ep_group().device_communicator.all2all_manager

    def prepare(
        self,
        a1: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        num_experts: int,
        expert_map: torch.Tensor | None,
        apply_router_weight_on_input: bool,
        quant_config: FusedMoEQuantConfig,
    ) -> mk.PrepareResultType:
        self._apply_router_weight_on_input(
            a1, topk_weights, topk_ids, apply_router_weight_on_input
        )

        if not self.use_dp:
            # Non-DP case: standard quantization
            a1q, a1q_scale = moe_kernel_quantize_input(
                a1,
                quant_config.a1_gscale,
                quant_config.quant_dtype,
                quant_config.per_act_token_quant,
                quant_config.block_shape,
                is_fp4_scale_swizzled=not self.use_dp,
            )
        else:
            # DP case: use FlashInfer AllToAll
            global_num_tokens_cpu = get_local_sizes()
            top_k = topk_ids.size(1)

            (self.alltoall_info, topk_ids, topk_weights, a1q, a1q_scale) = (
                flashinfer_alltoall_dispatch(
                    self.all2all_manager,
                    global_num_tokens_cpu,
                    a1,
                    quant_config.a1_gscale,
                    topk_ids,
                    topk_weights,
                    top_k,
                    num_experts,
                    quant_config,
                )
            )

        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:
            top_k = topk_ids.size(1)
            token_count = output.shape[0]
            fused_expert_output = flashinfer_alltoall_combine(
                self.all2all_manager,
                fused_expert_output,
                top_k=top_k,
                token_count=token_count,
                alltoall_info=self.alltoall_info,
            )
        output.copy_(fused_expert_output)

all2all_manager instance-attribute 

all2all_manager = None

alltoall_info instance-attribute 

alltoall_info = None

__init__ 

__init__(use_dp: bool, num_dispatchers: int = 1)
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
72
73
74
75
76
77
78
79
80
81
82
83
def __init__(
    self,
    use_dp: bool,
    num_dispatchers: int = 1,
):
    super().__init__(use_dp, num_dispatchers)
    self.alltoall_info = None

    # Initialize all2all_manager only for DP case
    self.all2all_manager = None
    if self.use_dp:
        self.all2all_manager = get_ep_group().device_communicator.all2all_manager

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 
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
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:
        top_k = topk_ids.size(1)
        token_count = output.shape[0]
        fused_expert_output = flashinfer_alltoall_combine(
            self.all2all_manager,
            fused_expert_output,
            top_k=top_k,
            token_count=token_count,
            alltoall_info=self.alltoall_info,
        )
    output.copy_(fused_expert_output)

prepare 

prepare(
    a1: Tensor,
    topk_weights: Tensor,
    topk_ids: Tensor,
    num_experts: int,
    expert_map: Tensor | None,
    apply_router_weight_on_input: bool,
    quant_config: FusedMoEQuantConfig,
) -> PrepareResultType
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
def prepare(
    self,
    a1: torch.Tensor,
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    num_experts: int,
    expert_map: torch.Tensor | None,
    apply_router_weight_on_input: bool,
    quant_config: FusedMoEQuantConfig,
) -> mk.PrepareResultType:
    self._apply_router_weight_on_input(
        a1, topk_weights, topk_ids, apply_router_weight_on_input
    )

    if not self.use_dp:
        # Non-DP case: standard quantization
        a1q, a1q_scale = moe_kernel_quantize_input(
            a1,
            quant_config.a1_gscale,
            quant_config.quant_dtype,
            quant_config.per_act_token_quant,
            quant_config.block_shape,
            is_fp4_scale_swizzled=not self.use_dp,
        )
    else:
        # DP case: use FlashInfer AllToAll
        global_num_tokens_cpu = get_local_sizes()
        top_k = topk_ids.size(1)

        (self.alltoall_info, topk_ids, topk_weights, a1q, a1q_scale) = (
            flashinfer_alltoall_dispatch(
                self.all2all_manager,
                global_num_tokens_cpu,
                a1,
                quant_config.a1_gscale,
                topk_ids,
                topk_weights,
                top_k,
                num_experts,
                quant_config,
            )
        )

    return a1q, a1q_scale, None, topk_ids, topk_weights

FlashInferCutlassMoEPrepareAndFinalize

Bases: FusedMoEPrepareAndFinalize

Base class for FlashInfer MoE prepare and finalize operations.

Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
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
class FlashInferCutlassMoEPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
    """Base class for FlashInfer MoE prepare and finalize operations."""

    def __init__(
        self,
        use_dp: bool,
        num_dispatchers: int = 1,
    ):
        super().__init__()
        self.num_dispatchers_ = num_dispatchers
        self.use_dp = use_dp
        self.local_tokens = None

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

    def max_num_tokens_per_rank(self) -> int | None:
        return None

    def topk_indices_dtype(self) -> torch.dtype | None:
        return None

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

    def output_is_reduced(self) -> bool:
        return False

    def _apply_router_weight_on_input(
        self,
        a1: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        apply_router_weight_on_input: bool,
    ) -> None:
        """Apply router weight on input if needed."""
        if apply_router_weight_on_input:
            topk = topk_ids.size(1)
            assert topk == 1, (
                "apply_router_weight_on_input is only implemented for topk=1"
            )
            a1.mul_(topk_weights.to(a1.dtype))

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, num_dispatchers: int = 1)
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
27
28
29
30
31
32
33
34
35
def __init__(
    self,
    use_dp: bool,
    num_dispatchers: int = 1,
):
    super().__init__()
    self.num_dispatchers_ = num_dispatchers
    self.use_dp = use_dp
    self.local_tokens = None

_apply_router_weight_on_input 

_apply_router_weight_on_input(
    a1: Tensor,
    topk_weights: Tensor,
    topk_ids: Tensor,
    apply_router_weight_on_input: bool,
) -> None

Apply router weight on input if needed.

Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
53
54
55
56
57
58
59
60
61
62
63
64
65
66
def _apply_router_weight_on_input(
    self,
    a1: torch.Tensor,
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    apply_router_weight_on_input: bool,
) -> None:
    """Apply router weight on input if needed."""
    if apply_router_weight_on_input:
        topk = topk_ids.size(1)
        assert topk == 1, (
            "apply_router_weight_on_input is only implemented for topk=1"
        )
        a1.mul_(topk_weights.to(a1.dtype))

max_num_tokens_per_rank 

max_num_tokens_per_rank() -> int | None
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
41
42
def max_num_tokens_per_rank(self) -> int | None:
    return None

num_dispatchers 

num_dispatchers() -> int
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
47
48
def num_dispatchers(self) -> int:
    return self.num_dispatchers_

output_is_reduced 

output_is_reduced() -> bool
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
50
51
def output_is_reduced(self) -> bool:
    return False

topk_indices_dtype 

topk_indices_dtype() -> dtype | None
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
44
45
def topk_indices_dtype(self) -> torch.dtype | None:
    return None

create_flashinfer_prepare_finalize 

create_flashinfer_prepare_finalize(
    use_dp: bool,
    use_nvfp4: bool = False,
    enable_alltoallv: bool = False,
) -> FlashInferCutlassMoEPrepareAndFinalize

Factory function to create the appropriate FlashInfer implementation.

Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
298
299
300
301
302
303
304
305
306
307
308
309
310
def create_flashinfer_prepare_finalize(
    use_dp: bool,
    use_nvfp4: bool = False,
    enable_alltoallv: bool = False,
) -> FlashInferCutlassMoEPrepareAndFinalize:
    """Factory function to create the appropriate FlashInfer implementation."""
    if use_nvfp4:
        if enable_alltoallv:
            return FlashInferAllToAllMoEPrepareAndFinalize(use_dp)
        else:
            return FlashInferAllGatherMoEPrepareAndFinalize(use_dp)
    # Fp8 only supports AllGather
    return FlashInferAllGatherMoEPrepareAndFinalize(use_dp)

flashinfer_alltoall_combine 

flashinfer_alltoall_combine(
    all2all_manager: All2AllManagerBase,
    output: Tensor,
    top_k: int,
    token_count: int,
    alltoall_info,
)
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
def flashinfer_alltoall_combine(
    all2all_manager: All2AllManagerBase,
    output: torch.Tensor,
    top_k: int,
    token_count: int,
    alltoall_info,
):
    from flashinfer.comm.trtllm_alltoall import MnnvlMoe

    assert all2all_manager.ensure_alltoall_workspace_initialized(), (
        "FlashInfer AllToAll workspace not available"
    )
    return MnnvlMoe.mnnvl_moe_alltoallv_combine(
        output,
        alltoall_info,
        all2all_manager.workspace_tensor,
        ep_rank=all2all_manager.rank,
        ep_size=all2all_manager.world_size,
        top_k=top_k,
        token_count=token_count,
    )

flashinfer_alltoall_dispatch 

flashinfer_alltoall_dispatch(
    all2all_manager: All2AllManagerBase,
    global_num_tokens_cpu: list[int],
    x: Tensor,
    gs: Tensor,
    topk_ids: Tensor,
    topk_weights: Tensor,
    top_k: int,
    num_experts: int,
    quant_config: FusedMoEQuantConfig,
)
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
def flashinfer_alltoall_dispatch(
    all2all_manager: All2AllManagerBase,
    global_num_tokens_cpu: list[int],
    x: torch.Tensor,
    gs: torch.Tensor,
    topk_ids: torch.Tensor,
    topk_weights: torch.Tensor,
    top_k: int,
    num_experts: int,
    quant_config: FusedMoEQuantConfig,
):
    from flashinfer.comm.trtllm_alltoall import MnnvlMoe

    assert all2all_manager.ensure_alltoall_workspace_initialized(), (
        "FlashInfer AllToAll workspace not available"
    )

    ep_rank = all2all_manager.rank
    ep_size = all2all_manager.world_size
    max_num_token = (
        max(global_num_tokens_cpu) if global_num_tokens_cpu is not None else x.shape[0]
    )
    alltoall_info, topk_ids, topk_weights, _ = (
        MnnvlMoe.mnnvl_moe_alltoallv_prepare_without_allgather(
            topk_ids,
            topk_weights,
            None,
            all2all_manager.prepare_workspace,
            max_num_token,
            ep_rank,
            ep_size,
            num_experts,
            num_experts,
            top_k,
        )
    )

    x, x_sf = moe_kernel_quantize_input(
        x,
        gs,
        quant_config.quant_dtype,
        quant_config.per_act_token_quant,
        quant_config.block_shape,
        is_fp4_scale_swizzled=False,  # delay swizzle to after comm
    )
    x = MnnvlMoe.mnnvl_moe_alltoallv(
        x,
        alltoall_info,
        all2all_manager.workspace_tensor,
        ep_rank,
        ep_size,
    )

    x_sf = MnnvlMoe.mnnvl_moe_alltoallv(
        x_sf,
        alltoall_info,
        all2all_manager.workspace_tensor,
        ep_rank,
        ep_size,
    )
    x_sf = nvfp4_block_scale_interleave(x_sf)
    return alltoall_info, topk_ids, topk_weights, x, x_sf

get_local_sizes 

get_local_sizes()
Source code in vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py 
20
21
def get_local_sizes():
    return get_forward_context().dp_metadata.get_chunk_sizes_across_dp_rank()