| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """DeepSeek V4 FlashInfer sparse MLA backend.""" |
| 4 | |
| 5 | from typing import TYPE_CHECKING, ClassVar, cast |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | from vllm.config import VllmConfig |
| 10 | from vllm.config.cache import CacheDType |
| 11 | from vllm.forward_context import get_forward_context |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.model_executor.layers.fusion.quant_activation import QuantizedActivation |
| 14 | from vllm.models.deepseek_v4.attention import DeepseekV4Attention |
| 15 | from vllm.models.deepseek_v4.common.ops import ( |
| 16 | build_flashinfer_mixed_sparse_indices, |
| 17 | compute_global_topk_indices_and_lens, |
| 18 | ) |
| 19 | from vllm.models.deepseek_v4.nvidia.ops.o_proj import ( |
| 20 | compute_fp8_einsum_recipe, |
| 21 | deep_gemm_fp8_o_proj, |
| 22 | rope_quant_attn_out, |
| 23 | rope_quant_unsupported_reason, |
| 24 | ) |
| 25 | from vllm.models.deepseek_v4.sparse_mla import ( |
| 26 | DeepseekV4FlashMLAMetadata, |
| 27 | DeepseekV4SparseMLABackend, |
| 28 | DeepseekV4SparseMLAMetadataBuilder, |
| 29 | ) |
| 30 | from vllm.platforms.interface import DeviceCapability |
| 31 | from vllm.utils.flashinfer import flashinfer_trtllm_batch_decode_sparse_mla_dsv4 |
| 32 | from vllm.v1.attention.backend import AttentionCGSupport, MultipleOf |
| 33 | from vllm.v1.attention.backends.mla.compressor_utils import ( |
| 34 | get_dspark_swa_index_width, |
| 35 | ) |
| 36 | from vllm.v1.attention.backends.mla.sparse_swa import ( |
| 37 | DeepseekSparseSWABackend, |
| 38 | DeepseekSparseSWAMetadataBuilder, |
| 39 | ) |
| 40 | |
| 41 | if TYPE_CHECKING: |
| 42 | from vllm.v1.attention.backends.mla.sparse_swa import DeepseekSparseSWAMetadata |
| 43 | |
| 44 | logger = init_logger(__name__) |
| 45 | |
| 46 | _FLASHINFER_DSV4_WORKSPACE_BUFFER_SIZE = 128 * 1024 * 1024 |
| 47 | _flashinfer_dsv4_workspace_by_device: dict[torch.device, torch.Tensor] = {} |
| 48 | |
| 49 | |
| 50 | def _get_flashinfer_dsv4_workspace(device: torch.device) -> torch.Tensor: |
| 51 | workspace = _flashinfer_dsv4_workspace_by_device.get(device) |
| 52 | if workspace is None: |
| 53 | workspace = torch.zeros( |
| 54 | _FLASHINFER_DSV4_WORKSPACE_BUFFER_SIZE, |
| 55 | dtype=torch.uint8, |
| 56 | device=device, |
| 57 | ) |
| 58 | _flashinfer_dsv4_workspace_by_device[device] = workspace |
| 59 | return workspace |
| 60 | |
| 61 | |
| 62 | def _packed_block_span(pool: torch.Tensor) -> int: |
| 63 | """Per-block stride of ``pool`` in tokens (``stride(0)//stride(-2)``): == |
| 64 | block_size for unpacked KV, larger when packed (#44577). Raises if not |
| 65 | token-aligned.""" |
| 66 | block_stride = pool.stride(0) |
| 67 | token_stride = pool.stride(-2) |
| 68 | if block_stride % token_stride != 0: |
| 69 | raise NotImplementedError( |
| 70 | "FLASHINFER_MLA_SPARSE_DSV4 packed KV requires the per-block stride " |
| 71 | f"({block_stride}) to be a multiple of the per-token stride " |
| 72 | f"({token_stride}); this layout is not supported yet." |
| 73 | ) |
| 74 | return block_stride // token_stride |
| 75 | |
| 76 | |
| 77 | # Sparse MLA h_q counts accepted natively (flashinfer>=0.6.14, #3545). |
| 78 | _SPARSE_MLA_SUPPORTED_Q_HEADS = (8, 16, 32, 64, 128) |
| 79 | |
| 80 | |
| 81 | def _pad_to_supported_q_heads(num_heads: int) -> int: |
| 82 | for supported in _SPARSE_MLA_SUPPORTED_Q_HEADS: |
| 83 | if num_heads <= supported: |
| 84 | return supported |
| 85 | raise ValueError( |
| 86 | f"DeepseekV4 FlashInfer MLA Sparse does not support {num_heads} heads " |
| 87 | "(sparse MLA kernel requires h_q in {8, 16, 32, 64, 128})." |
| 88 | ) |
| 89 | |
| 90 | |
| 91 | def _required_sm120_sparse_topk(vllm_config: VllmConfig, window_size: int) -> int: |
| 92 | """Return the SM120 DSV4 SWA specialization needed by this model.""" |
| 93 | if not vllm_config.attention_config.use_non_causal: |
| 94 | return window_size |
| 95 | speculative_config = vllm_config.speculative_config |
| 96 | if speculative_config is None: |
| 97 | return window_size |
| 98 | return get_dspark_swa_index_width( |
| 99 | window_size, |
| 100 | speculative_config.num_speculative_tokens, |
| 101 | ) |
| 102 | |
| 103 | |
| 104 | class DeepseekV4FlashInferMLASparseBackend(DeepseekV4SparseMLABackend): |
| 105 | """FlashInfer backend using the DSv4 sparse metadata/cache layout. |
| 106 | |
| 107 | Inherits the base and backend reuses its``DeepseekV4SparseMLAMetadataBuilder`` |
| 108 | """ |
| 109 | |
| 110 | supported_dtypes: ClassVar[list[torch.dtype]] = [torch.bfloat16] |
| 111 | supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [ |
| 112 | "auto", |
| 113 | "bfloat16", |
| 114 | "fp8", |
| 115 | "fp8_e4m3", |
| 116 | "fp8_ds_mla", |
| 117 | ] |
| 118 | |
| 119 | @staticmethod |
| 120 | def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]: |
| 121 | return [256] |
| 122 | |
| 123 | @staticmethod |
| 124 | def get_name() -> str: |
| 125 | return "FLASHINFER_MLA_SPARSE_DSV4" |
| 126 | |
| 127 | @classmethod |
| 128 | def get_supported_head_sizes(cls) -> list[int]: |
| 129 | return [512] |
| 130 | |
| 131 | @classmethod |
| 132 | def supports_sink(cls) -> bool: |
| 133 | return True |
| 134 | |
| 135 | @classmethod |
| 136 | def is_sparse(cls) -> bool: |
| 137 | return True |
| 138 | |
| 139 | @classmethod |
| 140 | def supports_compute_capability(cls, capability: DeviceCapability) -> bool: |
| 141 | return capability.major in [10, 12] |
| 142 | |
| 143 | @classmethod |
| 144 | def supports_combination( |
| 145 | cls, |
| 146 | head_size: int, |
| 147 | dtype: torch.dtype, |
| 148 | kv_cache_dtype: CacheDType | None, |
| 149 | block_size: int | None, |
| 150 | use_mla: bool, |
| 151 | has_sink: bool, |
| 152 | use_sparse: bool, |
| 153 | use_mm_prefix: bool, |
| 154 | device_capability: DeviceCapability, |
| 155 | ) -> str | None: |
| 156 | if device_capability.major == 10: |
| 157 | if kv_cache_dtype == "fp8_ds_mla": |
| 158 | return ( |
| 159 | "FLASHINFER_MLA_SPARSE_DSV4 SM10x uses the plain " |
| 160 | "per-tensor FP8 KV layout, not fp8_ds_mla" |
| 161 | ) |
| 162 | if kv_cache_dtype not in (None, "auto", "bfloat16", "fp8", "fp8_e4m3"): |
| 163 | return "kv_cache_dtype not supported" |
| 164 | return None |
| 165 | if device_capability.major == 12: |
| 166 | if kv_cache_dtype not in ("fp8", "fp8_e4m3", "fp8_ds_mla"): |
| 167 | return "kv_cache_dtype not supported" |
| 168 | from vllm.utils.flashinfer import has_flashinfer_sparse_mla_sm120 |
| 169 | |
| 170 | if not has_flashinfer_sparse_mla_sm120(): |
| 171 | return ( |
| 172 | "FLASHINFER_MLA_SPARSE_DSV4 SM120 requires FlashInfer's " |
| 173 | "sparse MLA decode API" |
| 174 | ) |
| 175 | return None |
| 176 | return "FLASHINFER_MLA_SPARSE_DSV4 requires SM10x or SM12x" |
| 177 | |
| 178 | @staticmethod |
| 179 | def get_builder_cls() -> type["DeepseekV4FlashInferSparseMLAMetadataBuilder"]: |
| 180 | return DeepseekV4FlashInferSparseMLAMetadataBuilder |
| 181 | |
| 182 | |
| 183 | class DeepseekV4FlashInferSparseMLAMetadataBuilder(DeepseekV4SparseMLAMetadataBuilder): |
| 184 | """Varlen-capable metadata builder for the FlashInfer sparse MLA backend.""" |
| 185 | |
| 186 | _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.ALWAYS |
| 187 | |
| 188 | |
| 189 | class DeepseekSparseSWAFlashInferMetadataBuilder(DeepseekSparseSWAMetadataBuilder): |
| 190 | """SWA metadata for the FlashInfer sparse decode path (varlen decode).""" |
| 191 | |
| 192 | _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.ALWAYS |
| 193 | |
| 194 | |
| 195 | class DeepseekSparseSWAFlashInferBackend(DeepseekSparseSWABackend): |
| 196 | @staticmethod |
| 197 | def get_builder_cls() -> type[DeepseekSparseSWAFlashInferMetadataBuilder]: |
| 198 | return DeepseekSparseSWAFlashInferMetadataBuilder |
| 199 | |
| 200 | |
| 201 | class DeepseekV4FlashInferMLAAttention(DeepseekV4Attention): |
| 202 | """FlashInfer TRTLLM-gen sparse MLA attention layer for SM100 DeepSeek V4.""" |
| 203 | |
| 204 | backend_cls = DeepseekV4FlashInferMLASparseBackend |
| 205 | swa_backend_cls = DeepseekSparseSWAFlashInferBackend |
| 206 | use_fp8_ds_mla_layout: ClassVar[bool] = False |
| 207 | |
| 208 | @classmethod |
| 209 | def get_padded_num_q_heads(cls, num_heads: int) -> int: |
| 210 | return _pad_to_supported_q_heads(num_heads) |
| 211 | |
| 212 | def _o_proj( |
| 213 | self, o: torch.Tensor | QuantizedActivation, positions: torch.Tensor |
| 214 | ) -> torch.Tensor: |
| 215 | return deep_gemm_fp8_o_proj( |
| 216 | o, |
| 217 | positions, |
| 218 | self.rotary_emb.cos_sin_cache, |
| 219 | self.wo_a, |
| 220 | self.wo_b, |
| 221 | n_groups=self.n_local_groups, |
| 222 | heads_per_group=self.n_local_heads // self.n_local_groups, |
| 223 | nope_dim=self.nope_head_dim, |
| 224 | rope_dim=self.rope_head_dim, |
| 225 | o_lora_rank=self.o_lora_rank, |
| 226 | einsum_recipe=self._einsum_recipe, |
| 227 | tma_aligned_scales=self._tma_aligned_scales, |
| 228 | ) |
| 229 | |
| 230 | def _alloc_attn_out( |
| 231 | self, num_tokens: int, hidden_states: torch.Tensor |
| 232 | ) -> torch.Tensor | QuantizedActivation: |
| 233 | if self._fuse_rope_quant: |
| 234 | return rope_quant_attn_out(num_tokens, hidden_states.device) |
| 235 | return super()._alloc_attn_out(num_tokens, hidden_states) |
| 236 | |
| 237 | def __init__(self, vllm_config: VllmConfig, *args, **kwargs) -> None: |
| 238 | super().__init__(vllm_config, *args, **kwargs) |
| 239 | self._einsum_recipe, self._tma_aligned_scales = compute_fp8_einsum_recipe() |
| 240 | # RopeQuant derives each query's RoPE position as seq_len - q_len + i, |
| 241 | # which matches `positions` for DSpark's non-causal draft block too, |
| 242 | # since every call passes the real per-request seq_lens. The one |
| 243 | # exception is a draft block running past max_model_len, whose clamped |
| 244 | # positions it does not reproduce; that only costs those drafts. |
| 245 | reason = rope_quant_unsupported_reason(self) |
| 246 | self._fuse_rope_quant = reason is None |
| 247 | if reason is None: |
| 248 | logger.info_once( |
| 249 | "FLASHINFER_MLA_SPARSE_DSV4 fuses the inverse RoPE and FP8 " |
| 250 | "quant of the attention output." |
| 251 | ) |
| 252 | else: |
| 253 | logger.debug_once("FLASHINFER_MLA_SPARSE_DSV4 RopeQuant off: %s", reason) |
| 254 | # Per-tensor FP8 scale buffers + precomputed scalar BMM scales. Only the |
| 255 | # per-tensor FP8 cache path consumes these; bf16 reads ``self.scale``. |
| 256 | if self.kv_cache_torch_dtype != torch.float8_e4m3fn: |
| 257 | return |
| 258 | fp8_q_scale = 1.0 |
| 259 | fp8_kv_scale = 1.0 |
| 260 | self.register_buffer( |
| 261 | "_flashinfer_fp8_q_scale", |
| 262 | torch.tensor([fp8_q_scale], dtype=torch.float32), |
| 263 | persistent=False, |
| 264 | ) |
| 265 | self.register_buffer( |
| 266 | "_flashinfer_fp8_q_scale_inv", |
| 267 | torch.tensor([1.0 / fp8_q_scale], dtype=torch.float32), |
| 268 | persistent=False, |
| 269 | ) |
| 270 | self.register_buffer( |
| 271 | "_flashinfer_fp8_kv_scale", |
| 272 | torch.tensor([fp8_kv_scale], dtype=torch.float32), |
| 273 | persistent=False, |
| 274 | ) |
| 275 | # TRTLLM-gen takes scalar scale args on a distinct C++ path vs |
| 276 | # one-element tensors, so these are Python floats. |
| 277 | self._flashinfer_fp8_bmm1_scale = self.scale * fp8_q_scale * fp8_kv_scale |
| 278 | self._flashinfer_fp8_bmm2_scale = fp8_kv_scale |
| 279 | |
| 280 | def forward_mqa( |
| 281 | self, |
| 282 | q: torch.Tensor, |
| 283 | kv: torch.Tensor, |
| 284 | positions: torch.Tensor, |
| 285 | output: torch.Tensor | QuantizedActivation, |
| 286 | ) -> None: |
| 287 | if isinstance(output, QuantizedActivation): |
| 288 | assert output.data.shape[0] == q.shape[0] |
| 289 | else: |
| 290 | # The TRTLLM-gen kernel requires h_q in {64, 128}, so the output |
| 291 | # buffer is allocated at the padded head count while q arrives at |
| 292 | # the local head count; _forward pads q to match before the launcher. |
| 293 | assert output.shape[0] == q.shape[0] and output.shape[-1] == q.shape[-1], ( |
| 294 | f"output buffer shape {output.shape} incompatible with q {q.shape}" |
| 295 | ) |
| 296 | assert output.shape[1] >= q.shape[1], ( |
| 297 | f"output heads {output.shape[1]} must be >= q heads {q.shape[1]}" |
| 298 | ) |
| 299 | # Per-tensor FP8 q produces a bf16 attention output. |
| 300 | expected_output_dtype = ( |
| 301 | torch.bfloat16 if q.dtype == torch.float8_e4m3fn else q.dtype |
| 302 | ) |
| 303 | assert output.dtype == expected_output_dtype, ( |
| 304 | f"output dtype {output.dtype} must match expected " |
| 305 | f"{expected_output_dtype} for q dtype {q.dtype}" |
| 306 | ) |
| 307 | |
| 308 | forward_context = get_forward_context() |
| 309 | attn_metadata = forward_context.attn_metadata |
| 310 | if attn_metadata is None: |
| 311 | # Warmup dummy run: FlashInfer reads the cache directly and lazily |
| 312 | # allocates its workspace, so nothing to reserve here. |
| 313 | if isinstance(output, QuantizedActivation): |
| 314 | output.data.zero_() |
| 315 | output.scale.zero_() |
| 316 | else: |
| 317 | output.zero_() |
| 318 | return |
| 319 | |
| 320 | assert isinstance(attn_metadata, dict) |
| 321 | flashmla_metadata = cast( |
| 322 | DeepseekV4FlashMLAMetadata | None, attn_metadata.get(self.prefix) |
| 323 | ) |
| 324 | swa_metadata = cast( |
| 325 | "DeepseekSparseSWAMetadata | None", |
| 326 | attn_metadata.get(self.swa_cache_layer.prefix), |
| 327 | ) |
| 328 | assert swa_metadata is not None |
| 329 | |
| 330 | swa_only = self.compress_ratio <= 1 |
| 331 | # SWA-only layers don't allocate their own compressed KV cache. |
| 332 | self_kv_cache = self.kv_cache if not swa_only else None |
| 333 | swa_kv_cache = self.swa_cache_layer.kv_cache |
| 334 | |
| 335 | self._forward( |
| 336 | q=q, |
| 337 | kv_cache=self_kv_cache, |
| 338 | swa_k_cache=swa_kv_cache, |
| 339 | swa_metadata=swa_metadata, |
| 340 | attn_metadata=flashmla_metadata, |
| 341 | swa_only=swa_only, |
| 342 | output=output, |
| 343 | ) |
| 344 | |
| 345 | def _build_sparse_index_metadata( |
| 346 | self, |
| 347 | kv_cache: torch.Tensor | None, |
| 348 | swa_k_cache: torch.Tensor, |
| 349 | swa_metadata: "DeepseekSparseSWAMetadata", |
| 350 | attn_metadata: DeepseekV4FlashMLAMetadata | None, |
| 351 | swa_only: bool, |
| 352 | num_rows: int | None = None, |
| 353 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 354 | """Build the combined sparse-index tensors for the mixed batch. |
| 355 | |
| 356 | Returns ``(compressed_kv_cache, seq_lens, sparse_indices, |
| 357 | sparse_topk_lens)``, sized ``num_rows`` (default: the real tokens). |
| 358 | """ |
| 359 | num_decodes = swa_metadata.num_decodes |
| 360 | num_prefills = swa_metadata.num_prefills |
| 361 | num_decode_tokens = swa_metadata.num_decode_tokens |
| 362 | num_prefill_tokens = swa_metadata.num_prefill_tokens |
| 363 | num_reqs = num_decodes + num_prefills |
| 364 | num_tokens = num_decode_tokens + num_prefill_tokens |
| 365 | |
| 366 | assert swa_metadata.seq_lens is not None |
| 367 | assert swa_metadata.query_start_loc is not None |
| 368 | assert swa_metadata.token_to_req_indices is not None |
| 369 | assert swa_metadata.decode_swa_indices is not None |
| 370 | assert swa_metadata.block_table is not None |
| 371 | |
| 372 | decode_swa_indices = swa_metadata.decode_swa_indices.reshape( |
| 373 | num_decode_tokens, swa_metadata.decode_swa_width |
| 374 | ) |
| 375 | decode_compressed_topk_lens = None |
| 376 | decode_compressed_indices_are_local = False |
| 377 | decode_is_valid_token = None |
| 378 | |
| 379 | if swa_only: |
| 380 | assert self.topk_indices_buffer is not None |
| 381 | compressed_kv_cache = swa_k_cache |
| 382 | decode_compressed_indices = None |
| 383 | prefill_topk_indices = self.topk_indices_buffer[ |
| 384 | num_decode_tokens:num_tokens, :0 |
| 385 | ] |
| 386 | compressed_block_table = None |
| 387 | compressed_block_size = swa_metadata.block_size |
| 388 | top_k = 0 |
| 389 | else: |
| 390 | assert kv_cache is not None |
| 391 | assert attn_metadata is not None |
| 392 | compressed_kv_cache = kv_cache |
| 393 | compressed_block_table = attn_metadata.block_table[:num_reqs] |
| 394 | compressed_block_size = attn_metadata.block_size // self.compress_ratio |
| 395 | |
| 396 | if self.compress_ratio == 4: |
| 397 | assert self.topk_indices_buffer is not None |
| 398 | if num_prefill_tokens > 0: |
| 399 | prefill_topk_indices = self.topk_indices_buffer[ |
| 400 | num_decode_tokens:num_tokens |
| 401 | ] |
| 402 | top_k = prefill_topk_indices.shape[-1] |
| 403 | else: |
| 404 | prefill_topk_indices = self.topk_indices_buffer[:0, :0] |
| 405 | top_k = 0 |
| 406 | |
| 407 | decode_compressed_indices_are_local = True |
| 408 | assert swa_metadata.is_valid_token is not None |
| 409 | decode_is_valid_token = swa_metadata.is_valid_token[:num_decode_tokens] |
| 410 | if num_decode_tokens > 0: |
| 411 | decode_compressed_indices = self.topk_indices_buffer[ |
| 412 | :num_decode_tokens |
| 413 | ] |
| 414 | else: |
| 415 | # Keep the logical width aligned with the mixed-batch case so |
| 416 | # pure-prefill steps reuse the same Triton specialization. |
| 417 | decode_compressed_indices = prefill_topk_indices[:0] |
| 418 | else: |
| 419 | if num_prefill_tokens > 0: |
| 420 | assert attn_metadata.c128a_prefill_topk_indices is not None |
| 421 | prefill_topk_indices = attn_metadata.c128a_prefill_topk_indices |
| 422 | top_k = prefill_topk_indices.shape[-1] |
| 423 | else: |
| 424 | prefill_topk_indices = decode_swa_indices[:0, :0] |
| 425 | top_k = 0 |
| 426 | |
| 427 | if num_decode_tokens > 0: |
| 428 | assert attn_metadata.c128a_global_decode_topk_indices is not None |
| 429 | assert attn_metadata.c128a_decode_topk_lens is not None |
| 430 | decode_compressed_indices = ( |
| 431 | attn_metadata.c128a_global_decode_topk_indices.view( |
| 432 | num_decode_tokens, -1 |
| 433 | ) |
| 434 | ) |
| 435 | decode_compressed_topk_lens = attn_metadata.c128a_decode_topk_lens |
| 436 | if num_prefill_tokens == 0: |
| 437 | prefill_topk_indices = decode_compressed_indices[:0, :0] |
| 438 | else: |
| 439 | decode_compressed_indices = prefill_topk_indices[:0] |
| 440 | decode_compressed_topk_lens = swa_metadata.seq_lens[:0] |
| 441 | |
| 442 | query_start_loc = swa_metadata.query_start_loc[: num_reqs + 1] |
| 443 | seq_lens = swa_metadata.seq_lens[:num_reqs] |
| 444 | assert seq_lens.dtype == torch.int32 |
| 445 | # cache for SWA-only and C128A that build the same mixed sparse indices |
| 446 | # C4A stays uncached. |
| 447 | cache_key = ( |
| 448 | "swa_only" |
| 449 | if swa_only |
| 450 | else ("c128a" if self.compress_ratio == 128 else "c4a") |
| 451 | ) |
| 452 | if num_rows is not None: |
| 453 | cache_key = f"{cache_key}:{num_rows}" |
| 454 | cached_sparse = swa_metadata.flashinfer_sparse_index_cache.get(cache_key, None) |
| 455 | if cached_sparse is None: |
| 456 | swa_block_span = _packed_block_span(swa_k_cache) |
| 457 | compressed_block_span = _packed_block_span(compressed_kv_cache) |
| 458 | sparse_indices, sparse_topk_lens = build_flashinfer_mixed_sparse_indices( |
| 459 | decode_swa_indices, |
| 460 | decode_compressed_indices, |
| 461 | decode_compressed_topk_lens, |
| 462 | prefill_topk_indices[:num_prefill_tokens], |
| 463 | query_start_loc, |
| 464 | seq_lens, |
| 465 | swa_metadata.token_to_req_indices[:num_tokens], |
| 466 | swa_metadata.block_table[:num_reqs], |
| 467 | swa_metadata.block_size, |
| 468 | compressed_block_table, |
| 469 | compressed_block_size, |
| 470 | self.window_size, |
| 471 | self.compress_ratio, |
| 472 | top_k, |
| 473 | decode_compressed_indices_are_local=decode_compressed_indices_are_local, |
| 474 | decode_is_valid_token=decode_is_valid_token, |
| 475 | swa_block_span=swa_block_span, |
| 476 | compressed_block_span=compressed_block_span, |
| 477 | prefill_left_visible=swa_metadata.prefill_left_visible, |
| 478 | prefill_right_visible=swa_metadata.prefill_right_visible, |
| 479 | # getattr for tests that bypass __init__ via object.__new__. |
| 480 | max_image_tokens=getattr(self, "max_image_tokens", 0), |
| 481 | num_rows=num_rows, |
| 482 | ) |
| 483 | if not cache_key.startswith("c4a"): |
| 484 | swa_metadata.flashinfer_sparse_index_cache[cache_key] = ( |
| 485 | sparse_indices, |
| 486 | sparse_topk_lens, |
| 487 | ) |
| 488 | else: |
| 489 | sparse_indices, sparse_topk_lens = cached_sparse |
| 490 | return compressed_kv_cache, seq_lens, sparse_indices, sparse_topk_lens |
| 491 | |
| 492 | def _forward( |
| 493 | self, |
| 494 | q: torch.Tensor, |
| 495 | kv_cache: torch.Tensor | None, |
| 496 | swa_k_cache: torch.Tensor, |
| 497 | swa_metadata: "DeepseekSparseSWAMetadata", |
| 498 | attn_metadata: DeepseekV4FlashMLAMetadata | None, |
| 499 | swa_only: bool, |
| 500 | output: torch.Tensor | QuantizedActivation, |
| 501 | ) -> None: |
| 502 | assert self.kv_cache_torch_dtype in (torch.bfloat16, torch.float8_e4m3fn) |
| 503 | num_decodes = swa_metadata.num_decodes |
| 504 | num_prefills = swa_metadata.num_prefills |
| 505 | num_decode_tokens = swa_metadata.num_decode_tokens |
| 506 | num_prefill_tokens = swa_metadata.num_prefill_tokens |
| 507 | num_reqs = num_decodes + num_prefills |
| 508 | num_tokens = num_decode_tokens + num_prefill_tokens |
| 509 | if num_tokens == 0: |
| 510 | return |
| 511 | # RopeQuant's output is group-major with a group stride of exactly |
| 512 | # sum_q tokens, so both calls span every row of the (CUDA-graph |
| 513 | # padded) buffers and select their tokens through cum_seq_lens_q |
| 514 | # alone; the kernel never touches rows outside it. |
| 515 | rope_quant = isinstance(output, QuantizedActivation) |
| 516 | |
| 517 | ( |
| 518 | compressed_kv_cache, |
| 519 | seq_lens, |
| 520 | sparse_indices, |
| 521 | sparse_topk_lens, |
| 522 | ) = self._build_sparse_index_metadata( |
| 523 | kv_cache=kv_cache, |
| 524 | swa_k_cache=swa_k_cache, |
| 525 | swa_metadata=swa_metadata, |
| 526 | attn_metadata=attn_metadata, |
| 527 | swa_only=swa_only, |
| 528 | num_rows=q.shape[0] if rope_quant else None, |
| 529 | ) |
| 530 | |
| 531 | # CUDA graph execution can pad q/output past the scheduled token count; |
| 532 | # restrict to the real tokens (the launcher validates sparse indices). |
| 533 | query = q if rope_quant else q[:num_tokens] |
| 534 | bmm1_scale: float | torch.Tensor = self.scale |
| 535 | bmm2_scale: float | torch.Tensor = 1.0 |
| 536 | if self.kv_cache_torch_dtype == torch.float8_e4m3fn: |
| 537 | assert query.dtype == torch.float8_e4m3fn |
| 538 | bmm1_scale = self._flashinfer_fp8_bmm1_scale |
| 539 | bmm2_scale = self._flashinfer_fp8_bmm2_scale |
| 540 | else: |
| 541 | assert query.dtype == torch.bfloat16 |
| 542 | query = query.contiguous() |
| 543 | |
| 544 | # The TRTLLM-gen sparse-MLA kernel requires h_q in {64, 128}; zero-pad |
| 545 | # the query heads to the allocated output head count. Padded heads attend |
| 546 | # to the shared KV and are sliced off downstream (output is padded too). |
| 547 | padded_heads = ( |
| 548 | output.shape[1] if isinstance(output, torch.Tensor) else self.padded_heads |
| 549 | ) |
| 550 | if query.shape[1] < padded_heads: |
| 551 | padded_query = query.new_zeros( |
| 552 | (query.shape[0], padded_heads, query.shape[2]) |
| 553 | ) |
| 554 | padded_query[:, : query.shape[1], :] = query |
| 555 | query = padded_query |
| 556 | |
| 557 | workspace = _get_flashinfer_dsv4_workspace(q.device) |
| 558 | query_start_loc = swa_metadata.query_start_loc |
| 559 | query_start_loc_cpu = swa_metadata.query_start_loc_cpu |
| 560 | assert query_start_loc is not None and query_start_loc_cpu is not None |
| 561 | |
| 562 | def run( |
| 563 | rows: slice, reqs: slice, cum_seq_lens_q: torch.Tensor, max_q_len: int |
| 564 | ) -> None: |
| 565 | if rope_quant: |
| 566 | assert isinstance(output, QuantizedActivation) |
| 567 | rows = slice(None) |
| 568 | out_kwargs = dict( |
| 569 | out=output.data, |
| 570 | dsv4_output_scale=output.scale, |
| 571 | dsv4_inv_rope_cos_sin_cache=self.rotary_emb.cos_sin_cache, |
| 572 | ) |
| 573 | else: |
| 574 | assert isinstance(output, torch.Tensor) |
| 575 | out_kwargs = dict(out=output[rows]) |
| 576 | flashinfer_trtllm_batch_decode_sparse_mla_dsv4( |
| 577 | query=query[rows], |
| 578 | swa_kv_cache=swa_k_cache, |
| 579 | workspace_buffer=workspace, |
| 580 | sparse_indices=sparse_indices[rows], |
| 581 | sparse_indices_are_storage_offsets=True, |
| 582 | compressed_kv_cache=compressed_kv_cache, |
| 583 | sparse_topk_lens=sparse_topk_lens[rows], |
| 584 | seq_lens=seq_lens[reqs], |
| 585 | bmm1_scale=bmm1_scale, |
| 586 | bmm2_scale=bmm2_scale, |
| 587 | sinks=self.attn_sink, |
| 588 | cum_seq_lens_q=cum_seq_lens_q, |
| 589 | max_q_len=max_q_len, |
| 590 | **out_kwargs, |
| 591 | ) |
| 592 | |
| 593 | # Keep the TRTLLM-gen decode/prefill split: the launcher is tuned for |
| 594 | # uniform-q batches, and this avoids flattening mixed batches into one call. |
| 595 | if num_decode_tokens > 0: |
| 596 | run( |
| 597 | slice(0, num_decode_tokens), |
| 598 | slice(0, num_decodes), |
| 599 | query_start_loc[: num_decodes + 1], |
| 600 | swa_metadata.max_decode_query_len, |
| 601 | ) |
| 602 | |
| 603 | if num_prefill_tokens > 0: |
| 604 | prefill_cu = query_start_loc[num_decodes : num_reqs + 1] |
| 605 | if not rope_quant: |
| 606 | # The prefill query view re-anchors at offset 0, so rebase the |
| 607 | # cumulative query offsets to start at 0. |
| 608 | prefill_cu = prefill_cu - query_start_loc[num_decodes] |
| 609 | prefill_cu_cpu = query_start_loc_cpu[num_decodes : num_reqs + 1] |
| 610 | prefill_lens_cpu = prefill_cu_cpu[1:] - prefill_cu_cpu[:-1] |
| 611 | run( |
| 612 | slice(num_decode_tokens, num_tokens), |
| 613 | slice(num_decodes, num_reqs), |
| 614 | prefill_cu, |
| 615 | int(prefill_lens_cpu.max().item()), |
| 616 | ) |
| 617 | |
| 618 | |
| 619 | class DeepseekV4FlashInferSM120Attention(DeepseekV4Attention): |
| 620 | """DeepSeek V4 sparse MLA attention through FlashInfer's SM120 kernels.""" |
| 621 | |
| 622 | backend_cls = DeepseekV4FlashInferMLASparseBackend |
| 623 | swa_backend_cls = DeepseekSparseSWAFlashInferBackend |
| 624 | use_fp8_ds_mla_layout: ClassVar[bool] = True |
| 625 | |
| 626 | @staticmethod |
| 627 | def _get_workspace(device: torch.device) -> torch.Tensor: |
| 628 | return _get_flashinfer_dsv4_workspace(device) |
| 629 | |
| 630 | @staticmethod |
| 631 | def _as_sparse_cache(kv_cache: torch.Tensor) -> torch.Tensor: |
| 632 | if kv_cache.dtype == torch.float8_e4m3fn: |
| 633 | kv_cache = kv_cache.view(torch.uint8) |
| 634 | if kv_cache.dim() == 4: |
| 635 | return kv_cache |
| 636 | return kv_cache.unsqueeze(-2) |
| 637 | |
| 638 | @classmethod |
| 639 | def get_padded_num_q_heads(cls, num_heads: int) -> int: |
| 640 | return _pad_to_supported_q_heads(num_heads) |
| 641 | |
| 642 | def _o_proj(self, o: torch.Tensor, positions: torch.Tensor) -> torch.Tensor: |
| 643 | return deep_gemm_fp8_o_proj( |
| 644 | o, |
| 645 | positions, |
| 646 | self.rotary_emb.cos_sin_cache, |
| 647 | self.wo_a, |
| 648 | self.wo_b, |
| 649 | n_groups=self.n_local_groups, |
| 650 | heads_per_group=self.n_local_heads // self.n_local_groups, |
| 651 | nope_dim=self.nope_head_dim, |
| 652 | rope_dim=self.rope_head_dim, |
| 653 | o_lora_rank=self.o_lora_rank, |
| 654 | einsum_recipe=self._einsum_recipe, |
| 655 | tma_aligned_scales=self._tma_aligned_scales, |
| 656 | ) |
| 657 | |
| 658 | def __init__(self, vllm_config: VllmConfig, *args, **kwargs) -> None: |
| 659 | super().__init__(vllm_config, *args, **kwargs) |
| 660 | from vllm.utils.flashinfer import has_flashinfer_sparse_mla_sm120_config |
| 661 | |
| 662 | required_topk = _required_sm120_sparse_topk(vllm_config, self.window_size) |
| 663 | if not has_flashinfer_sparse_mla_sm120_config(self.padded_heads, required_topk): |
| 664 | raise RuntimeError( |
| 665 | "FLASHINFER_MLA_SPARSE_DSV4 on SM120 requires a FlashInfer " |
| 666 | "DSV4 sparse MLA decode specialization for " |
| 667 | f"(num_q_heads={self.padded_heads}, top_k={required_topk}). " |
| 668 | "Install a FlashInfer build containing " |
| 669 | "flashinfer-ai/flashinfer#4380." |
| 670 | ) |
| 671 | self._einsum_recipe, self._tma_aligned_scales = compute_fp8_einsum_recipe() |
| 672 | # Per-tensor FP8 cache path scales. |
| 673 | if self.kv_cache_torch_dtype != torch.float8_e4m3fn: |
| 674 | return |
| 675 | fp8_q_scale = 1.0 |
| 676 | fp8_kv_scale = 1.0 |
| 677 | self.register_buffer( |
| 678 | "_flashinfer_fp8_q_scale", |
| 679 | torch.tensor([fp8_q_scale], dtype=torch.float32), |
| 680 | persistent=False, |
| 681 | ) |
| 682 | self.register_buffer( |
| 683 | "_flashinfer_fp8_q_scale_inv", |
| 684 | torch.tensor([1.0 / fp8_q_scale], dtype=torch.float32), |
| 685 | persistent=False, |
| 686 | ) |
| 687 | self.register_buffer( |
| 688 | "_flashinfer_fp8_kv_scale", |
| 689 | torch.tensor([fp8_kv_scale], dtype=torch.float32), |
| 690 | persistent=False, |
| 691 | ) |
| 692 | # FlashInfer expects scalar scale arguments for this path. |
| 693 | self._flashinfer_fp8_bmm1_scale = self.scale * fp8_q_scale * fp8_kv_scale |
| 694 | self._flashinfer_fp8_bmm2_scale = fp8_kv_scale |
| 695 | |
| 696 | def _reserve_empty_forward_workspace(self) -> None: |
| 697 | self._get_workspace( |
| 698 | torch.device("cuda", torch.accelerator.current_device_index()) |
| 699 | ) |
| 700 | |
| 701 | def _forward_sparse_impl( |
| 702 | self, |
| 703 | q: torch.Tensor, |
| 704 | output: torch.Tensor, |
| 705 | flashmla_metadata: DeepseekV4FlashMLAMetadata | None, |
| 706 | swa_metadata: "DeepseekSparseSWAMetadata", |
| 707 | self_kv_cache: torch.Tensor | None, |
| 708 | swa_kv_cache: torch.Tensor, |
| 709 | swa_only: bool, |
| 710 | ) -> None: |
| 711 | num_decode_tokens = swa_metadata.num_decode_tokens |
| 712 | if swa_metadata.num_prefills > 0: |
| 713 | self._forward_prefill( |
| 714 | q=q[num_decode_tokens:], |
| 715 | compressed_k_cache=self_kv_cache, |
| 716 | swa_k_cache=swa_kv_cache, |
| 717 | output=output[num_decode_tokens:], |
| 718 | attn_metadata=flashmla_metadata, |
| 719 | swa_metadata=swa_metadata, |
| 720 | ) |
| 721 | if swa_metadata.num_decodes > 0: |
| 722 | self._forward_decode( |
| 723 | q=q[:num_decode_tokens], |
| 724 | kv_cache=self_kv_cache, |
| 725 | swa_metadata=swa_metadata, |
| 726 | attn_metadata=flashmla_metadata, |
| 727 | swa_only=swa_only, |
| 728 | output=output[:num_decode_tokens], |
| 729 | ) |
| 730 | |
| 731 | def forward_mqa( |
| 732 | self, |
| 733 | q: torch.Tensor, |
| 734 | kv: torch.Tensor, |
| 735 | positions: torch.Tensor, |
| 736 | output: torch.Tensor, |
| 737 | ) -> None: |
| 738 | # Output may be padded to backend-supported head counts. |
| 739 | assert output.shape[0] == q.shape[0] and output.shape[-1] == q.shape[-1], ( |
| 740 | f"output buffer shape {output.shape} incompatible with q shape {q.shape}" |
| 741 | ) |
| 742 | assert output.shape[1] >= q.shape[1], ( |
| 743 | f"output heads {output.shape[1]} must be >= q heads {q.shape[1]}" |
| 744 | ) |
| 745 | # Per-tensor FP8 q produces a bf16 attention output. |
| 746 | expected_output_dtype = ( |
| 747 | torch.bfloat16 if q.dtype == torch.float8_e4m3fn else q.dtype |
| 748 | ) |
| 749 | assert output.dtype == expected_output_dtype, ( |
| 750 | f"output dtype {output.dtype} must match expected {expected_output_dtype} " |
| 751 | f"for q dtype {q.dtype}" |
| 752 | ) |
| 753 | |
| 754 | forward_context = get_forward_context() |
| 755 | attn_metadata = forward_context.attn_metadata |
| 756 | if attn_metadata is None: |
| 757 | self._reserve_empty_forward_workspace() |
| 758 | output.zero_() |
| 759 | return |
| 760 | |
| 761 | assert isinstance(attn_metadata, dict) |
| 762 | flashmla_metadata = cast( |
| 763 | DeepseekV4FlashMLAMetadata | None, attn_metadata.get(self.prefix) |
| 764 | ) |
| 765 | swa_metadata = cast( |
| 766 | "DeepseekSparseSWAMetadata | None", |
| 767 | attn_metadata.get(self.swa_cache_layer.prefix), |
| 768 | ) |
| 769 | assert swa_metadata is not None |
| 770 | |
| 771 | swa_only = self.compress_ratio <= 1 |
| 772 | # SWA-only layers don't allocate their own compressed KV cache. |
| 773 | self_kv_cache = self.kv_cache if not swa_only else None |
| 774 | swa_kv_cache = self.swa_cache_layer.kv_cache |
| 775 | |
| 776 | self._forward_sparse_impl( |
| 777 | q=q, |
| 778 | output=output, |
| 779 | flashmla_metadata=flashmla_metadata, |
| 780 | swa_metadata=swa_metadata, |
| 781 | self_kv_cache=self_kv_cache, |
| 782 | swa_kv_cache=swa_kv_cache, |
| 783 | swa_only=swa_only, |
| 784 | ) |
| 785 | |
| 786 | def _prepare_query(self, q: torch.Tensor, output: torch.Tensor) -> torch.Tensor: |
| 787 | if self.kv_cache_torch_dtype == torch.float8_e4m3fn: |
| 788 | assert q.dtype == torch.float8_e4m3fn |
| 789 | q = q.to(torch.bfloat16) |
| 790 | else: |
| 791 | assert q.dtype == torch.bfloat16 |
| 792 | padded_heads = output.shape[1] |
| 793 | if q.shape[1] < padded_heads: |
| 794 | padded_query = q.new_zeros((q.shape[0], padded_heads, q.shape[2])) |
| 795 | padded_query[:, : q.shape[1], :] = q |
| 796 | q = padded_query |
| 797 | return q.contiguous() |
| 798 | |
| 799 | def _forward_decode( |
| 800 | self, |
| 801 | q: torch.Tensor, |
| 802 | kv_cache: torch.Tensor | None, |
| 803 | swa_metadata: "DeepseekSparseSWAMetadata", |
| 804 | attn_metadata: DeepseekV4FlashMLAMetadata | None, |
| 805 | swa_only: bool, |
| 806 | output: torch.Tensor, |
| 807 | ) -> None: |
| 808 | num_decodes = swa_metadata.num_decodes |
| 809 | num_decode_tokens = swa_metadata.num_decode_tokens |
| 810 | |
| 811 | extra_sparse_indices = None |
| 812 | extra_sparse_lengths = None |
| 813 | if not swa_only: |
| 814 | if attn_metadata is None: |
| 815 | raise RuntimeError( |
| 816 | "Sparse MLA metadata is required for compressed layers." |
| 817 | ) |
| 818 | if swa_metadata.is_valid_token is None: |
| 819 | raise RuntimeError( |
| 820 | "SWA validity metadata is required for compressed layers." |
| 821 | ) |
| 822 | is_valid = swa_metadata.is_valid_token[:num_decode_tokens] |
| 823 | if self.compress_ratio == 4: |
| 824 | if self.topk_indices_buffer is None: |
| 825 | raise RuntimeError( |
| 826 | "C4A decode requires top-k indices from the indexer." |
| 827 | ) |
| 828 | block_size = attn_metadata.block_size // self.compress_ratio |
| 829 | global_indices, extra_sparse_lengths = ( |
| 830 | compute_global_topk_indices_and_lens( |
| 831 | self.topk_indices_buffer[:num_decode_tokens], |
| 832 | swa_metadata.token_to_req_indices, |
| 833 | attn_metadata.block_table[:num_decodes], |
| 834 | block_size, |
| 835 | is_valid, |
| 836 | ) |
| 837 | ) |
| 838 | extra_sparse_indices = global_indices.view(num_decode_tokens, 1, -1) |
| 839 | else: |
| 840 | extra_sparse_indices = attn_metadata.c128a_global_decode_topk_indices |
| 841 | extra_sparse_lengths = attn_metadata.c128a_decode_topk_lens |
| 842 | |
| 843 | swa_indices = swa_metadata.decode_swa_indices |
| 844 | swa_lens = swa_metadata.decode_swa_lens |
| 845 | assert swa_indices is not None |
| 846 | assert swa_lens is not None |
| 847 | q = self._prepare_query(q, output) |
| 848 | swa_cache = self._as_sparse_cache(self.swa_cache_layer.kv_cache) |
| 849 | extra_cache = self._as_sparse_cache(kv_cache) if kv_cache is not None else None |
| 850 | if extra_cache is not None and extra_sparse_indices is None: |
| 851 | raise RuntimeError( |
| 852 | "Compressed sparse MLA decode requires compressed sparse indices." |
| 853 | ) |
| 854 | flashinfer_trtllm_batch_decode_sparse_mla_dsv4( |
| 855 | query=q, |
| 856 | swa_kv_cache=swa_cache, |
| 857 | workspace_buffer=self._get_workspace(q.device), |
| 858 | sparse_indices=swa_indices, |
| 859 | compressed_kv_cache=extra_cache, |
| 860 | out=output, |
| 861 | bmm1_scale=self.scale, |
| 862 | sinks=self.attn_sink, |
| 863 | kv_layout="NHD", |
| 864 | swa_topk_lens=swa_lens, |
| 865 | extra_sparse_indices=extra_sparse_indices, |
| 866 | extra_sparse_topk_lens=extra_sparse_lengths, |
| 867 | ) |
| 868 | |
| 869 | def _forward_prefill( |
| 870 | self, |
| 871 | q: torch.Tensor, |
| 872 | compressed_k_cache: torch.Tensor | None, |
| 873 | swa_k_cache: torch.Tensor, |
| 874 | output: torch.Tensor, |
| 875 | attn_metadata: DeepseekV4FlashMLAMetadata | None, |
| 876 | swa_metadata: "DeepseekSparseSWAMetadata", |
| 877 | ) -> None: |
| 878 | swa_only = self.compress_ratio <= 1 |
| 879 | |
| 880 | num_prefills = swa_metadata.num_prefills |
| 881 | num_decodes = swa_metadata.num_decodes |
| 882 | num_decode_tokens = swa_metadata.num_decode_tokens |
| 883 | num_prefill_tokens = swa_metadata.num_prefill_tokens |
| 884 | |
| 885 | query_start_loc_cpu = swa_metadata.query_start_loc_cpu |
| 886 | assert query_start_loc_cpu is not None |
| 887 | prefill_token_base = query_start_loc_cpu[num_decodes] |
| 888 | |
| 889 | local_topk_indices: torch.Tensor | None |
| 890 | if swa_only: |
| 891 | local_topk_indices = None |
| 892 | elif self.compress_ratio == 4: |
| 893 | if self.topk_indices_buffer is None: |
| 894 | raise RuntimeError( |
| 895 | "C4A prefill requires top-k indices from the indexer." |
| 896 | ) |
| 897 | local_topk_indices = self.topk_indices_buffer[ |
| 898 | num_decode_tokens : num_decode_tokens + num_prefill_tokens |
| 899 | ] |
| 900 | else: |
| 901 | if attn_metadata is None: |
| 902 | raise RuntimeError("C128A prefill metadata is missing.") |
| 903 | local_topk_indices = attn_metadata.c128a_prefill_topk_indices |
| 904 | |
| 905 | extra_sparse_indices: torch.Tensor | None = None |
| 906 | extra_sparse_lengths: torch.Tensor | None = None |
| 907 | if local_topk_indices is not None: |
| 908 | if attn_metadata is None: |
| 909 | raise RuntimeError("C4A prefill metadata is missing.") |
| 910 | if swa_metadata.token_to_req_indices is None: |
| 911 | raise RuntimeError("C4A prefill request mapping is missing.") |
| 912 | if swa_metadata.is_valid_token is None: |
| 913 | raise RuntimeError("C4A prefill validity metadata is missing.") |
| 914 | prefill_token_slice = slice( |
| 915 | num_decode_tokens, num_decode_tokens + num_prefill_tokens |
| 916 | ) |
| 917 | block_size = attn_metadata.block_size // self.compress_ratio |
| 918 | extra_sparse_indices, extra_sparse_lengths = ( |
| 919 | compute_global_topk_indices_and_lens( |
| 920 | local_topk_indices, |
| 921 | swa_metadata.token_to_req_indices[prefill_token_slice], |
| 922 | attn_metadata.block_table, |
| 923 | block_size, |
| 924 | swa_metadata.is_valid_token[prefill_token_slice], |
| 925 | ) |
| 926 | ) |
| 927 | |
| 928 | assert swa_metadata.prefill_swa_indices is not None |
| 929 | assert swa_metadata.prefill_swa_lens is not None |
| 930 | |
| 931 | q = self._prepare_query(q, output) |
| 932 | swa_kv_paged = self._as_sparse_cache(swa_k_cache) |
| 933 | if swa_only: |
| 934 | extra_kv_paged = None |
| 935 | else: |
| 936 | if compressed_k_cache is None: |
| 937 | raise RuntimeError( |
| 938 | "Compressed sparse MLA layers require their compressed KV cache." |
| 939 | ) |
| 940 | extra_kv_paged = self._as_sparse_cache(compressed_k_cache) |
| 941 | |
| 942 | num_chunks = ( |
| 943 | num_prefills + self.PREFILL_CHUNK_SIZE - 1 |
| 944 | ) // self.PREFILL_CHUNK_SIZE |
| 945 | for chunk_idx in range(num_chunks): |
| 946 | chunk_start = chunk_idx * self.PREFILL_CHUNK_SIZE |
| 947 | chunk_end = min(chunk_start + self.PREFILL_CHUNK_SIZE, num_prefills) |
| 948 | query_start = ( |
| 949 | query_start_loc_cpu[num_decodes + chunk_start] - prefill_token_base |
| 950 | ) |
| 951 | query_end = ( |
| 952 | query_start_loc_cpu[num_decodes + chunk_end] - prefill_token_base |
| 953 | ) |
| 954 | |
| 955 | extra_sparse_indices_chunk = ( |
| 956 | extra_sparse_indices[query_start:query_end] |
| 957 | if extra_sparse_indices is not None |
| 958 | else None |
| 959 | ) |
| 960 | extra_sparse_lengths_chunk = ( |
| 961 | extra_sparse_lengths[query_start:query_end] |
| 962 | if extra_sparse_lengths is not None |
| 963 | else None |
| 964 | ) |
| 965 | |
| 966 | q_chunk = q[query_start:query_end] |
| 967 | swa_indices_chunk = swa_metadata.prefill_swa_indices[query_start:query_end] |
| 968 | swa_lens_chunk = swa_metadata.prefill_swa_lens[query_start:query_end] |
| 969 | if extra_kv_paged is not None and extra_sparse_indices_chunk is None: |
| 970 | raise RuntimeError( |
| 971 | "Compressed sparse MLA prefill requires compressed sparse indices." |
| 972 | ) |
| 973 | flashinfer_trtllm_batch_decode_sparse_mla_dsv4( |
| 974 | query=q_chunk, |
| 975 | swa_kv_cache=swa_kv_paged, |
| 976 | workspace_buffer=self._get_workspace(q.device), |
| 977 | sparse_indices=swa_indices_chunk, |
| 978 | compressed_kv_cache=extra_kv_paged, |
| 979 | out=output[query_start:query_end], |
| 980 | bmm1_scale=self.scale, |
| 981 | sinks=self.attn_sink, |
| 982 | kv_layout="NHD", |
| 983 | swa_topk_lens=swa_lens_chunk, |
| 984 | extra_sparse_indices=extra_sparse_indices_chunk, |
| 985 | extra_sparse_topk_lens=extra_sparse_lengths_chunk, |
| 986 | ) |