Skip to content

vllm.model_executor.kernels.linear.nvfp4.flashinfer

Classes:

FlashInferB12xNvFp4LinearKernel

Bases: NvFp4LinearKernel

NVFP4 GEMM via FlashInfer's b12x CuTe DSL warp-level MMA kernel (SM120+).

Source code in vllm/model_executor/kernels/linear/nvfp4/flashinfer.py 
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
273
274
275
276
277
278
279
280
281
282
283
284
285
class FlashInferB12xNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FlashInfer's b12x CuTe DSL warp-level MMA kernel (SM120+)."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if current_platform.has_device_capability(120) and has_flashinfer_b12x_gemm():
            return True, None
        return (
            False,
            "FlashInfer b12x requires SM120+ and FlashInfer "
            "with Sm120BlockScaledDenseGemmKernel",
        )

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        layer.weight_scale = torch.nn.Parameter(
            swizzle_blockscale(layer.weight_scale.data), requires_grad=False
        )
        padded_weight, weights_padding_cols = pad_nvfp4_weight_for_cutlass(
            layer.weight.data
        )
        layer.weight = torch.nn.Parameter(padded_weight, requires_grad=False)
        layer.weights_padding_cols = weights_padding_cols

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="b12x",
        )

        x_fp4 = pad_nvfp4_activation_for_cutlass(
            x_fp4, getattr(layer, "weights_padding_cols", 0)
        )

        out = flashinfer_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
            backend="b12x",
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

FlashInferCudnnNvFp4LinearKernel

Bases: NvFp4LinearKernel

NVFP4 GEMM via FlashInfer's cuDNN wrapper.

Source code in vllm/model_executor/kernels/linear/nvfp4/flashinfer.py 
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
209
210
211
212
213
214
215
216
217
218
class FlashInferCudnnNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FlashInfer's cuDNN wrapper."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if has_flashinfer():
            return True, None
        return False, "FlashInfer required"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # cuDNN uses the same swizzled + padded layout as CUTLASS
        layer.weight_scale = torch.nn.Parameter(
            swizzle_blockscale(layer.weight_scale.data), requires_grad=False
        )
        padded_weight, weights_padding_cols = pad_nvfp4_weight_for_cutlass(
            layer.weight.data
        )
        layer.weight = torch.nn.Parameter(padded_weight, requires_grad=False)
        layer.weights_padding_cols = weights_padding_cols

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]
        weights_padding_bytes = getattr(layer, "weights_padding_cols", 0)

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="flashinfer-cudnn",
            padded_n=x.shape[-1] + weights_padding_bytes * 2,
        )

        out = flashinfer_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
            backend="cudnn",
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

FlashInferCutlassNvFp4LinearKernel

Bases: NvFp4LinearKernel

NVFP4 GEMM via FlashInfer's CUTLASS wrapper.

Source code in vllm/model_executor/kernels/linear/nvfp4/flashinfer.py 
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
class FlashInferCutlassNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FlashInfer's CUTLASS wrapper."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        from vllm.model_executor.layers.quantization.utils.nvfp4_utils import (
            cutlass_fp4_supported,
        )

        if (
            cutlass_fp4_supported()
            and current_platform.has_device_capability(100)
            and has_flashinfer()
        ):
            return True, None
        return False, "FlashInfer + >=sm_100 required"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        layer.weight_scale = torch.nn.Parameter(
            swizzle_blockscale(layer.weight_scale.data), requires_grad=False
        )
        padded_weight, weights_padding_cols = pad_nvfp4_weight_for_cutlass(
            layer.weight.data
        )
        layer.weight = torch.nn.Parameter(padded_weight, requires_grad=False)
        layer.weights_padding_cols = weights_padding_cols

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]
        weights_padding_bytes = getattr(layer, "weights_padding_cols", 0)

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="flashinfer-cutlass",
            padded_n=x.shape[-1] + weights_padding_bytes * 2,
        )

        out = flashinfer_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
            backend="cutlass",
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

FlashInferTrtllmNvFp4LinearKernel

Bases: NvFp4LinearKernel

NVFP4 GEMM via FlashInfer's TensorRT-LLM wrapper.

Source code in vllm/model_executor/kernels/linear/nvfp4/flashinfer.py 
 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
150
151
152
153
154
155
156
class FlashInferTrtllmNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FlashInfer's TensorRT-LLM wrapper."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if has_flashinfer():
            return True, None
        return False, "FlashInfer required"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        from flashinfer import shuffle_matrix_a, shuffle_matrix_sf_a

        weight = layer.weight.data
        weight_scale = layer.weight_scale.data
        epilogue_tile_m = 128

        layer.weight = torch.nn.Parameter(
            shuffle_matrix_a(weight.view(torch.uint8), epilogue_tile_m),
            requires_grad=False,
        )
        layer.weight_scale = torch.nn.Parameter(
            shuffle_matrix_sf_a(weight_scale.view(torch.uint8), epilogue_tile_m)
            .reshape(weight_scale.shape)
            .view(torch.float8_e4m3fn),
            requires_grad=False,
        )

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="flashinfer-trtllm",
        )

        out = flashinfer_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
            backend="trtllm",
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)