Skip to content

vllm.model_executor.layers.mamba.ops.ssu_dispatch

Dispatch module for Mamba selective state update (SSU) backends.

Provides a unified selective_state_update function that dispatches to either the Triton or FlashInfer backend based on the configured MambaBackendEnum. Follows SGLang's dispatch pattern adapted for vLLM.

FlashInferSSUBackend

Bases: MambaSSUBackend

FlashInfer-based SSU backend.

Source code in vllm/model_executor/layers/mamba/ops/ssu_dispatch.py 
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
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
class FlashInferSSUBackend(MambaSSUBackend):
    """FlashInfer-based SSU backend."""

    def __init__(self, mamba_config: MambaConfig):
        super().__init__(mamba_config)
        try:
            from flashinfer.mamba import selective_state_update as _fi_ssu
        except ImportError as e:
            raise ImportError(
                "FlashInfer is required for the flashinfer Mamba SSU backend. "
                "Please install flashinfer (>= 0.6.4): "
                "pip install flashinfer-python"
            ) from e
        self._kernel = _fi_ssu

    @property
    def name(self) -> str:
        return "flashinfer"

    def __call__(
        self,
        state: torch.Tensor,
        x: torch.Tensor,
        dt: torch.Tensor,
        A: torch.Tensor,
        B: torch.Tensor,
        C: torch.Tensor,
        D: torch.Tensor,
        dt_bias: torch.Tensor,
        z: torch.Tensor | None = None,
        dt_softplus: bool = False,
        state_batch_indices: torch.Tensor | None = None,
        dst_state_batch_indices: torch.Tensor | None = None,
        null_block_id: int = NULL_BLOCK_ID,
        out: torch.Tensor | None = None,
        num_accepted_tokens: torch.Tensor | None = None,
        cu_seqlens: torch.Tensor | None = None,
        is_blackwell: bool = False,
    ) -> None:
        rand_seed = (
            torch.randint(0, 2**32, (1,), device=state.device)
            if self._mamba_config.enable_stochastic_rounding
            else None
        )

        self._kernel(
            state,
            x,
            dt,
            A,
            B,
            C,
            D=D,
            z=z,
            dt_bias=dt_bias,
            dt_softplus=dt_softplus,
            state_batch_indices=state_batch_indices,
            dst_state_batch_indices=dst_state_batch_indices,
            cu_seqlens=cu_seqlens,
            num_accepted_tokens=num_accepted_tokens,
            cache_steps=state_batch_indices.size(-1)
            if cu_seqlens is not None and state_batch_indices is not None
            else 0,
            pad_slot_id=null_block_id,
            out=out,
            rand_seed=rand_seed,
            philox_rounds=self._mamba_config.stochastic_rounding_philox_rounds or 10,
        )

MambaSSUBackend

Bases: ABC

Abstract base class for Mamba SSU backends.

Source code in vllm/model_executor/layers/mamba/ops/ssu_dispatch.py 
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
class MambaSSUBackend(ABC):
    """Abstract base class for Mamba SSU backends."""

    def __init__(self, mamba_config: MambaConfig):
        self._mamba_config = mamba_config

    @property
    @abstractmethod
    def name(self) -> str: ...

    @abstractmethod
    def __call__(
        self,
        state: torch.Tensor,
        x: torch.Tensor,
        dt: torch.Tensor,
        A: torch.Tensor,
        B: torch.Tensor,
        C: torch.Tensor,
        D: torch.Tensor,
        dt_bias: torch.Tensor,
        z: torch.Tensor | None = None,
        dt_softplus: bool = False,
        state_batch_indices: torch.Tensor | None = None,
        dst_state_batch_indices: torch.Tensor | None = None,
        null_block_id: int = NULL_BLOCK_ID,
        out: torch.Tensor | None = None,
        num_accepted_tokens: torch.Tensor | None = None,
        cu_seqlens: torch.Tensor | None = None,
        is_blackwell: bool = False,
    ) -> None: ...

TritonSSUBackend

Bases: MambaSSUBackend

Triton-based SSU backend (vLLM's default).

Source code in vllm/model_executor/layers/mamba/ops/ssu_dispatch.py 
 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
100
101
102
103
104
105
106
107
108
109
110
class TritonSSUBackend(MambaSSUBackend):
    """Triton-based SSU backend (vLLM's default)."""

    def __init__(self, mamba_config: MambaConfig):
        super().__init__(mamba_config)
        from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
            selective_state_update as _triton_selective_state_update,
        )

        self._kernel = _triton_selective_state_update

    @property
    def name(self) -> str:
        return "triton"

    def __call__(
        self,
        state: torch.Tensor,
        x: torch.Tensor,
        dt: torch.Tensor,
        A: torch.Tensor,
        B: torch.Tensor,
        C: torch.Tensor,
        D: torch.Tensor,
        dt_bias: torch.Tensor,
        z: torch.Tensor | None = None,
        dt_softplus: bool = False,
        state_batch_indices: torch.Tensor | None = None,
        dst_state_batch_indices: torch.Tensor | None = None,
        null_block_id: int = NULL_BLOCK_ID,
        out: torch.Tensor | None = None,
        num_accepted_tokens: torch.Tensor | None = None,
        cu_seqlens: torch.Tensor | None = None,
        is_blackwell: bool = False,
    ) -> None:
        self._kernel(
            state,
            x,
            dt,
            A,
            B,
            C,
            D=D,
            z=z,
            dt_bias=dt_bias,
            dt_softplus=dt_softplus,
            state_batch_indices=state_batch_indices,
            dst_state_batch_indices=dst_state_batch_indices,
            null_block_id=null_block_id,
            out=out,
            num_accepted_tokens=num_accepted_tokens,
            cu_seqlens=cu_seqlens,
            is_blackwell=is_blackwell,
            enable_stochastic_rounding=self._mamba_config.enable_stochastic_rounding,
            cache_philox_rounds=self._mamba_config.stochastic_rounding_philox_rounds,
        )

get_mamba_ssu_backend 

get_mamba_ssu_backend() -> MambaSSUBackend

Get the current Mamba SSU backend. Raises if not initialized.

Source code in vllm/model_executor/layers/mamba/ops/ssu_dispatch.py 
210
211
212
213
214
215
216
217
def get_mamba_ssu_backend() -> MambaSSUBackend:
    """Get the current Mamba SSU backend. Raises if not initialized."""
    if _mamba_ssu_backend is None:
        raise RuntimeError(
            "Mamba SSU backend has not been initialized. "
            "Call initialize_mamba_ssu_backend() first."
        )
    return _mamba_ssu_backend

initialize_mamba_ssu_backend 

initialize_mamba_ssu_backend(
    mamba_config: MambaConfig,
) -> None

Initialize the global Mamba SSU backend.

Parameters:

Name Type Description Default
mamba_config MambaConfig

Mamba configuration.

 required
Source code in vllm/model_executor/layers/mamba/ops/ssu_dispatch.py 
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
def initialize_mamba_ssu_backend(mamba_config: MambaConfig) -> None:
    """Initialize the global Mamba SSU backend.

    Args:
        mamba_config: Mamba configuration.
    """
    global _mamba_ssu_backend

    backend = mamba_config.backend
    if backend not in _BACKEND_REGISTRY:
        raise ValueError(
            f"Unknown Mamba SSU backend: {backend}. "
            f"Valid options: {list(_BACKEND_REGISTRY.keys())}"
        )

    _mamba_ssu_backend = _BACKEND_REGISTRY[backend](mamba_config)
    logger.info("Using %s Mamba SSU backend.", _mamba_ssu_backend.name)

selective_state_update 

selective_state_update(
    state: Tensor,
    x: Tensor,
    dt: Tensor,
    A: Tensor,
    B: Tensor,
    C: Tensor,
    D: Tensor,
    dt_bias: Tensor,
    z: Tensor | None = None,
    dt_softplus: bool = False,
    state_batch_indices: Tensor | None = None,
    dst_state_batch_indices: Tensor | None = None,
    null_block_id: int = NULL_BLOCK_ID,
    out: Tensor | None = None,
    num_accepted_tokens: Tensor | None = None,
    cu_seqlens: Tensor | None = None,
    is_blackwell: bool = False,
) -> None

Unified dispatch for Mamba selective state update.

Delegates to the initialized backend (Triton or FlashInfer).

Source code in vllm/model_executor/layers/mamba/ops/ssu_dispatch.py 
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
def selective_state_update(
    state: torch.Tensor,
    x: torch.Tensor,
    dt: torch.Tensor,
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
    D: torch.Tensor,
    dt_bias: torch.Tensor,
    z: torch.Tensor | None = None,
    dt_softplus: bool = False,
    state_batch_indices: torch.Tensor | None = None,
    dst_state_batch_indices: torch.Tensor | None = None,
    null_block_id: int = NULL_BLOCK_ID,
    out: torch.Tensor | None = None,
    num_accepted_tokens: torch.Tensor | None = None,
    cu_seqlens: torch.Tensor | None = None,
    is_blackwell: bool = False,
) -> None:
    """Unified dispatch for Mamba selective state update.

    Delegates to the initialized backend (Triton or FlashInfer).
    """
    get_mamba_ssu_backend()(
        state,
        x,
        dt,
        A,
        B,
        C,
        D,
        dt_bias,
        z=z,
        dt_softplus=dt_softplus,
        state_batch_indices=state_batch_indices,
        dst_state_batch_indices=dst_state_batch_indices,
        null_block_id=null_block_id,
        out=out,
        num_accepted_tokens=num_accepted_tokens,
        cu_seqlens=cu_seqlens,
        is_blackwell=is_blackwell,
    )