Skip to content

vllm.models.minimax_m3.nvidia.msa_cutlass_sparse_decode

MiniMax CUTLASS sparse decode using per-query-token page indices.

Classes:

Functions:

MSACutlassDecodePlanCache dataclass

Reusable plans whose mutable tensors retain cudagraph-stable addresses.

Source code in vllm/models/minimax_m3/nvidia/msa_cutlass_sparse_decode.py 
 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
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
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
@dataclass
class MSACutlassDecodePlanCache:
    """Reusable plans whose mutable tensors retain cudagraph-stable addresses."""

    plans: dict[tuple[int, ...], Any] = field(init=False, default_factory=dict)

    def _build_plan(
        self,
        *,
        batch: int,
        decode_query_len: int,
        page_table_stride: int,
        initial_seq_lens_cpu: torch.Tensor,
        device: torch.device,
        num_q_heads: int,
        num_kv_heads: int,
        page_size: int,
        topk_blocks: int,
    ) -> Any:
        from vllm.third_party.fmha_sm100.api import fmha_sm100_plan

        qo_lens_cpu = torch.full((batch,), decode_query_len, dtype=torch.int32)
        kv_lens_cpu = initial_seq_lens_cpu
        plan = fmha_sm100_plan(
            qo_lens_cpu,
            kv_lens_cpu,
            num_q_heads,
            num_kv_heads=num_kv_heads,
            qo_offset=kv_lens_cpu - qo_lens_cpu,
            page_size=page_size,
            output_maxscore=False,
            kv_block_num=topk_blocks,
            causal=True,
            sparse_kernel_mode="decode",
            use_fp8_kvcache=True,
            split_prefill_decode=False,
            device=device,
        )

        plan_info = plan[3]
        row_starts = (
            torch.arange(batch, dtype=torch.int32, device=device)
            .mul_(page_table_stride)
            .repeat_interleave(decode_query_len)
        )
        page_indptr = torch.cat(
            (
                row_starts,
                torch.tensor(
                    [batch * page_table_stride],
                    dtype=torch.int32,
                    device=device,
                ),
            )
        )
        plan_info["kv_page_indptr"].copy_(page_indptr)
        return plan

    def prepare(
        self,
        block_table: torch.Tensor,
        seq_lens: torch.Tensor,
        seq_lens_cpu: torch.Tensor,
        decode_query_len: int,
        *,
        num_q_heads: int,
        num_kv_heads: int,
        page_size: int,
        topk_blocks: int,
    ) -> MSACutlassDecodeMetadata:
        batch = int(seq_lens.shape[0])
        if (
            block_table.device.type != "cuda"
            or block_table.dtype != torch.int32
            or not block_table.is_contiguous()
            or block_table.shape[0] != batch
        ):
            raise ValueError(
                "MSA sparse decode requires a contiguous CUDA int32 block "
                "table with one row per request"
            )
        if (
            seq_lens.dtype != torch.int32
            or not seq_lens.is_contiguous()
            or seq_lens.device != block_table.device
        ):
            raise ValueError(
                "MSA sparse decode requires contiguous CUDA int32 sequence "
                "lengths on the block table device"
            )
        if (
            seq_lens_cpu.device.type != "cpu"
            or seq_lens_cpu.dtype != torch.int32
            or not seq_lens_cpu.is_contiguous()
            or seq_lens_cpu.shape != seq_lens.shape
        ):
            raise ValueError(
                "MSA sparse decode requires contiguous CPU int32 sequence "
                "lengths matching the device sequence lengths"
            )

        page_table_stride = int(block_table.stride(0))
        key = (
            batch,
            decode_query_len,
            page_table_stride,
            num_q_heads,
            num_kv_heads,
            page_size,
            topk_blocks,
        )
        plan = self.plans.get(key)
        if plan is None:
            plan = self._build_plan(
                batch=batch,
                decode_query_len=decode_query_len,
                page_table_stride=page_table_stride,
                initial_seq_lens_cpu=seq_lens_cpu,
                device=seq_lens.device,
                num_q_heads=num_q_heads,
                num_kv_heads=num_kv_heads,
                page_size=page_size,
                topk_blocks=topk_blocks,
            )
            self.plans[key] = plan

        plan_info = plan[3]
        num_rows = batch * decode_query_len
        _update_runtime_metadata_kernel[(triton.cdiv(num_rows, 128),)](
            seq_lens,
            plan_info["kv_segment_lens"],
            plan_info["qo_offset"],
            num_rows=num_rows,
            decode_query_len=decode_query_len,
            BLOCK_SIZE=128,
        )
        return MSACutlassDecodeMetadata(
            plan=plan,
            page_table=block_table.view(-1),
        )

msa_cutlass_sparse_decode(query_fp8, kv_cache, topk, output, metadata, *, scale, q_scale_float, k_scale_float, v_scale_float)

Run CUTLASS sparse decode with metadata prepared by the MSA builder.

Source code in vllm/models/minimax_m3/nvidia/msa_cutlass_sparse_decode.py 
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
@torch.no_grad()
def msa_cutlass_sparse_decode(
    query_fp8: torch.Tensor,
    kv_cache: torch.Tensor,
    topk: torch.Tensor,
    output: torch.Tensor,
    metadata: MSACutlassDecodeMetadata,
    *,
    scale: float,
    q_scale_float: float,
    k_scale_float: float,
    v_scale_float: float,
) -> None:
    """Run CUTLASS sparse decode with metadata prepared by the MSA builder."""
    key, value = kv_cache.split(_HEAD_DIM, dim=-1)

    from vllm.third_party.fmha_sm100.api import fmha_sm100

    fmha_sm100(
        query_fp8,
        key,
        value,
        metadata.plan,
        kv_indices=metadata.page_table,
        kv_block_indexes=topk,
        out=output,
        output_maxscore=False,
        output_o=True,
        sm_scale=scale,
        q_scale=q_scale_float,
        k_scale=k_scale_float,
        v_scale=v_scale_float,
        o_scale=1.0,
    )

prepare_decode_metadata(block_table, seq_lens, seq_lens_cpu, decode_query_len, *, num_q_heads, num_kv_heads, page_size, topk_blocks, plan_cache=None)

Prepare graph-stable runtime metadata for one sparse decode step.

Source code in vllm/models/minimax_m3/nvidia/msa_cutlass_sparse_decode.py 
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
@torch.no_grad()
def prepare_decode_metadata(
    block_table: torch.Tensor,
    seq_lens: torch.Tensor,
    seq_lens_cpu: torch.Tensor,
    decode_query_len: int,
    *,
    num_q_heads: int,
    num_kv_heads: int,
    page_size: int,
    topk_blocks: int,
    plan_cache: MSACutlassDecodePlanCache | None = None,
) -> MSACutlassDecodeMetadata:
    """Prepare graph-stable runtime metadata for one sparse decode step."""
    cache = plan_cache or MSACutlassDecodePlanCache()
    return cache.prepare(
        block_table,
        seq_lens,
        seq_lens_cpu,
        decode_query_len,
        num_q_heads=num_q_heads,
        num_kv_heads=num_kv_heads,
        page_size=page_size,
        topk_blocks=topk_blocks,
    )

should_prepare_decode_metadata(batch_size, decode_query_len, *, decode_backend, num_q_heads, num_kv_heads, kv_cache_dtype, page_size, topk_blocks)

Return whether a graph shape can use the CUTLASS decode path.

Source code in vllm/models/minimax_m3/nvidia/msa_cutlass_sparse_decode.py 
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
def should_prepare_decode_metadata(
    batch_size: int,
    decode_query_len: int,
    *,
    decode_backend: MiniMaxM3MSADecodeBackend,
    num_q_heads: int,
    num_kv_heads: int,
    kv_cache_dtype: str,
    page_size: int,
    topk_blocks: int,
) -> bool:
    """Return whether a graph shape can use the CUTLASS decode path."""
    total_q = batch_size * decode_query_len
    return (
        supports_cutlass_sparse_decode(
            decode_backend=decode_backend,
            num_q_heads=num_q_heads,
            num_kv_heads=num_kv_heads,
            kv_cache_dtype=kv_cache_dtype,
            page_size=page_size,
            topk_blocks=topk_blocks,
        )
        and 1 <= decode_query_len <= _MAX_DECODE_QUERY_LEN
        and batch_size >= _MIN_CUTLASS_BATCH_SIZE
        and total_q * num_q_heads <= _MAX_QUERY_HEAD_ROWS
    )

supports_cutlass_sparse_decode(*, decode_backend, num_q_heads, num_kv_heads, kv_cache_dtype, page_size, topk_blocks)

Return whether static model geometry supports CUTLASS sparse decode.

Source code in vllm/models/minimax_m3/nvidia/msa_cutlass_sparse_decode.py 
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
def supports_cutlass_sparse_decode(
    *,
    decode_backend: MiniMaxM3MSADecodeBackend,
    num_q_heads: int,
    num_kv_heads: int,
    kv_cache_dtype: str,
    page_size: int,
    topk_blocks: int,
) -> bool:
    """Return whether static model geometry supports CUTLASS sparse decode."""
    return (
        decode_backend == "cutlass"
        and current_platform.is_cuda()
        and current_platform.is_device_capability_family(100)
        and kv_cache_dtype in ("fp8", "fp8_e4m3")
        and _supported_head_geometry(num_q_heads, num_kv_heads)
        and page_size == _PAGE_SIZE
        and topk_blocks == _TOPK
    )