| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """DeepseekV4 MLA Attention Layer.""" |
| 4 | |
| 5 | from abc import ABC, abstractmethod |
| 6 | from collections.abc import Callable |
| 7 | from typing import TYPE_CHECKING, Any, ClassVar, cast |
| 8 | |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | import torch.nn.functional as F |
| 12 | from transformers import DeepseekV2Config, DeepseekV3Config |
| 13 | |
| 14 | import vllm.envs as envs |
| 15 | from vllm.compilation.breakable_cudagraph import eager_break_during_capture |
| 16 | from vllm.model_executor.layers.linear import ( |
| 17 | ColumnParallelLinear, |
| 18 | MergedColumnParallelLinear, |
| 19 | ReplicatedLinear, |
| 20 | RowParallelLinear, |
| 21 | ) |
| 22 | from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer |
| 23 | from vllm.models.common.ops.fused_qk_rmsnorm import ( |
| 24 | _FUSED_Q_KV_RMSNORM_KERNEL, |
| 25 | fused_q_kv_rmsnorm, |
| 26 | ) |
| 27 | from vllm.models.deepseek_v4.common.ops import ( |
| 28 | fused_indexer_q_rope_quant, |
| 29 | ) |
| 30 | from vllm.models.deepseek_v4.common.ops.fused_indexer_q import MXFP4_BLOCK_SIZE |
| 31 | |
| 32 | if TYPE_CHECKING: |
| 33 | from vllm.v1.attention.backends.mla.sparse_swa import ( |
| 34 | DeepseekSparseSWAMetadata, |
| 35 | ) |
| 36 | |
| 37 | from vllm.config import ( |
| 38 | CacheConfig, |
| 39 | VllmConfig, |
| 40 | get_current_vllm_config, |
| 41 | ) |
| 42 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 43 | from vllm.forward_context import get_forward_context |
| 44 | from vllm.logger import init_logger |
| 45 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 46 | from vllm.model_executor.layers.fusion.quant_activation import QuantizedActivation |
| 47 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 48 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 49 | from vllm.model_executor.models.utils import extract_layer_index |
| 50 | from vllm.models.deepseek_v4.common.rope import build_deepseek_v4_rope |
| 51 | from vllm.models.deepseek_v4.compressor import DeepseekCompressor |
| 52 | from vllm.platforms import current_platform |
| 53 | from vllm.triton_utils import tl, triton |
| 54 | from vllm.utils.multi_stream_utils import ( |
| 55 | execute_in_parallel, |
| 56 | maybe_execute_in_parallel, |
| 57 | ) |
| 58 | from vllm.v1.attention.backend import AttentionBackend, AttentionMetadata |
| 59 | from vllm.v1.attention.backends.mla.indexer import ( |
| 60 | DeepseekV4IndexerBackend, |
| 61 | dsa_indexer_uses_fp4, |
| 62 | get_max_prefill_buffer_size, |
| 63 | ) |
| 64 | from vllm.v1.attention.backends.mla.sparse_swa import DeepseekV4SWACache |
| 65 | from vllm.v1.kv_cache_interface import ( |
| 66 | KVCacheSpec, |
| 67 | MLAAttentionSpec, |
| 68 | get_kv_quant_mode, |
| 69 | ) |
| 70 | |
| 71 | logger = init_logger(__name__) |
| 72 | |
| 73 | |
| 74 | @triton.jit |
| 75 | def _fill_short_context_topk_indices( |
| 76 | output, |
| 77 | positions, |
| 78 | TOP_K: tl.constexpr, |
| 79 | COMPRESS_RATIO: tl.constexpr, |
| 80 | PADDED_TOP_K: tl.constexpr, |
| 81 | ): |
| 82 | # small triton kernel that selects every candidate, -1 otherwise |
| 83 | row = tl.program_id(0) |
| 84 | offsets = tl.arange(0, PADDED_TOP_K) |
| 85 | num_compressed = (tl.load(positions + row) + 1) // COMPRESS_RATIO |
| 86 | tl.store( |
| 87 | output + row * TOP_K + offsets, |
| 88 | tl.where(offsets < num_compressed, offsets, -1), |
| 89 | mask=offsets < TOP_K, |
| 90 | ) |
| 91 | |
| 92 | |
| 93 | def _resolve_dsv4_kv_cache_dtype( |
| 94 | use_fp8_ds_mla_layout: bool, |
| 95 | kv_cache_dtype: str, |
| 96 | cache_config: CacheConfig | None, |
| 97 | ) -> tuple[str, torch.dtype]: |
| 98 | """Map ``(layout, --kv-cache-dtype)`` to ``(cache_dtype_str, torch_dtype)``. |
| 99 | |
| 100 | Both layouts are paged; they differ in the per-token block format. The |
| 101 | ``fp8_ds_mla`` format is UE8M0 block-scaled fp8 packed as ``uint8`` (the |
| 102 | canonical ``fp8_ds_mla`` string is written back onto ``cache_config`` so the |
| 103 | page-size specs pick the 576B per-token slot). Plain-row backends store each |
| 104 | token's KV row in its element dtype: bf16 or per-tensor FP8 E4M3. |
| 105 | """ |
| 106 | if use_fp8_ds_mla_layout: |
| 107 | # fp8_ds_mla block format: UE8M0 block-scaled fp8 packed as uint8. |
| 108 | if kv_cache_dtype == "auto": |
| 109 | kv_cache_dtype = "fp8" |
| 110 | if not kv_cache_dtype.startswith("fp8"): |
| 111 | raise ValueError( |
| 112 | "DeepseekV4 fp8_ds_mla layout only supports fp8 " |
| 113 | f"kv-cache, got {kv_cache_dtype}. Please set " |
| 114 | "`--kv-cache-dtype fp8` or select a backend that supports " |
| 115 | "bfloat16 KV cache." |
| 116 | ) |
| 117 | if kv_cache_dtype != "fp8_ds_mla": |
| 118 | if cache_config is not None: |
| 119 | cache_config.cache_dtype = "fp8_ds_mla" |
| 120 | kv_cache_dtype = "fp8_ds_mla" |
| 121 | logger.info_once("Using DeepSeek's fp8_ds_mla KV cache format.") |
| 122 | return kv_cache_dtype, torch.uint8 |
| 123 | |
| 124 | # Plain bf16 / per-tensor fp8 KV row (FlashInfer). |
| 125 | if kv_cache_dtype.startswith("fp8"): |
| 126 | return kv_cache_dtype, torch.float8_e4m3fn |
| 127 | # auto / bfloat16 -> plain bf16 KV row. |
| 128 | return kv_cache_dtype, torch.bfloat16 |
| 129 | |
| 130 | |
| 131 | class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC): |
| 132 | """DeepseekV4 MLA attention layer. |
| 133 | |
| 134 | The platform-specific sparse-MLA forward (``forward_mqa`` / |
| 135 | ``get_padded_num_q_heads`` / ``_o_proj`` / ``backend_cls``) is provided by a |
| 136 | subclass — ``DeepseekV4FlashMLAAttention`` / |
| 137 | ``DeepseekV4FlashInferSM120Attention`` / |
| 138 | ``DeepseekV4FlashInferMLAAttention`` (CUDA) or |
| 139 | ``DeepseekV4ROCMAiterMLAAttention`` (ROCm) — selected by the platform-specific |
| 140 | deepseek_v4 model module. The base is never instantiated directly. |
| 141 | """ |
| 142 | |
| 143 | # Provided by the platform subclass. |
| 144 | backend_cls: ClassVar[type[AttentionBackend]] |
| 145 | # Backend for the SWA cache layer; None uses the default SWA backend. |
| 146 | swa_backend_cls: ClassVar[type[AttentionBackend] | None] = None |
| 147 | # KV-cache per-token block format (both layouts are paged). True (default) |
| 148 | # = fp8_ds_mla (UE8M0 block-scaled fp8 packed as uint8); False = plain |
| 149 | # bf16 / per-tensor fp8 KV row. Backends can override the instance hook when |
| 150 | # a single attention class dispatches across arch-specific layouts. |
| 151 | use_fp8_ds_mla_layout: ClassVar[bool] = True |
| 152 | # Prefill is processed in fixed-size chunks; this bounds the bf16 kv-gather |
| 153 | # workspace allocated in _forward_prefill and is also read by the dummy-run |
| 154 | # path to pre-reserve that workspace. |
| 155 | PREFILL_CHUNK_SIZE: ClassVar[int] = 4 |
| 156 | |
| 157 | @classmethod |
| 158 | @abstractmethod |
| 159 | def get_padded_num_q_heads(cls, num_heads: int) -> int: |
| 160 | """Q head count the q/output buffers are allocated at. |
| 161 | |
| 162 | The layer allocates the q/output buffers at |
| 163 | ``[N, get_padded_num_q_heads(n_local_heads), head_dim]``. Must satisfy |
| 164 | ``result >= num_heads``. Backends with no padding constraint return |
| 165 | ``num_heads``. |
| 166 | """ |
| 167 | raise NotImplementedError |
| 168 | |
| 169 | @abstractmethod |
| 170 | def forward_mqa( |
| 171 | self, |
| 172 | q: torch.Tensor, |
| 173 | kv: torch.Tensor, |
| 174 | positions: torch.Tensor, |
| 175 | output: "torch.Tensor | QuantizedActivation", |
| 176 | ) -> None: |
| 177 | """Platform-specific sparse MLA forward; writes attention into ``output``.""" |
| 178 | raise NotImplementedError |
| 179 | |
| 180 | @abstractmethod |
| 181 | def _o_proj( |
| 182 | self, o: "torch.Tensor | QuantizedActivation", positions: torch.Tensor |
| 183 | ) -> torch.Tensor: |
| 184 | """Inverse-RoPE + wo_a + wo_b output projection (platform-specific). |
| 185 | |
| 186 | ``o`` is the live heads of the bf16 attention output, or the |
| 187 | QuantizedActivation of a layer whose ``_alloc_attn_out`` returns one. |
| 188 | """ |
| 189 | raise NotImplementedError |
| 190 | |
| 191 | def _uses_fp8_ds_mla_layout(self) -> bool: |
| 192 | """Return whether this instance stores fp8 KV in fp8_ds_mla layout.""" |
| 193 | return self.use_fp8_ds_mla_layout |
| 194 | |
| 195 | def __init__( |
| 196 | self, |
| 197 | vllm_config: VllmConfig, |
| 198 | prefix: str, |
| 199 | topk_indices_buffer: torch.Tensor | None = None, |
| 200 | aux_stream_list: list[torch.cuda.Stream] | None = None, |
| 201 | ) -> None: |
| 202 | super().__init__() |
| 203 | config = vllm_config.model_config.hf_config |
| 204 | quant_config = vllm_config.quant_config |
| 205 | cache_config = vllm_config.cache_config |
| 206 | tp_size = get_tensor_model_parallel_world_size() |
| 207 | layer_id = extract_layer_index(prefix) |
| 208 | |
| 209 | self.prefix = prefix # Alias for compatibility with compressor |
| 210 | self.hidden_size = config.hidden_size |
| 211 | self.n_heads = config.num_attention_heads |
| 212 | assert self.n_heads % tp_size == 0 |
| 213 | self.n_local_heads = self.n_heads // tp_size |
| 214 | self.q_lora_rank = config.q_lora_rank |
| 215 | self.o_lora_rank = config.o_lora_rank |
| 216 | self.head_dim = config.head_dim |
| 217 | self.rope_head_dim = config.qk_rope_head_dim |
| 218 | self.nope_head_dim = self.head_dim - self.rope_head_dim |
| 219 | self.n_groups = config.o_groups |
| 220 | self.n_local_groups = self.n_groups // tp_size |
| 221 | self.window_size = config.sliding_window |
| 222 | # Vision variant: image spans are visible bidirectionally, widening |
| 223 | # prefill SWA index rows by up to max_image_tokens columns. |
| 224 | self.max_image_tokens = ( |
| 225 | getattr(config, "vision_max_n_token", 0) |
| 226 | if getattr(config, "vision_n_layers", 0) > 0 |
| 227 | else 0 |
| 228 | ) |
| 229 | # NOTE(zyongye) Compress ratio can't be 0 |
| 230 | # we do this for because MTP layer is not included |
| 231 | # in the compress ratio list |
| 232 | if layer_id < config.num_hidden_layers: |
| 233 | self.compress_ratio = max(1, config.compress_ratios[layer_id]) |
| 234 | else: |
| 235 | self.compress_ratio = 1 |
| 236 | self.eps = config.rms_norm_eps |
| 237 | self.scale = self.head_dim**-0.5 |
| 238 | |
| 239 | # Padded Q head count is dictated by the platform subclass. |
| 240 | self.padded_heads = self.get_padded_num_q_heads(self.n_local_heads) |
| 241 | # Sink padded to the same head count, initialized to -inf (no sink |
| 242 | # effect). Weight loading fills the first n_local_heads slots. |
| 243 | self.attn_sink = nn.Parameter( |
| 244 | torch.full((self.padded_heads,), -float("inf"), dtype=torch.float32), |
| 245 | requires_grad=False, |
| 246 | ) |
| 247 | |
| 248 | self.fused_wqa_wkv = MergedColumnParallelLinear( |
| 249 | self.hidden_size, |
| 250 | [self.q_lora_rank, self.head_dim], |
| 251 | bias=False, |
| 252 | quant_config=quant_config, |
| 253 | prefix=f"{prefix}.fused_wqa_wkv", |
| 254 | disable_tp=True, # fused ReplicatedLinear |
| 255 | ) |
| 256 | self.q_norm = RMSNorm(self.q_lora_rank, self.eps) |
| 257 | self.wq_b = ColumnParallelLinear( |
| 258 | self.q_lora_rank, |
| 259 | self.n_heads * self.head_dim, |
| 260 | bias=False, |
| 261 | quant_config=quant_config, |
| 262 | return_bias=False, |
| 263 | prefix=f"{prefix}.wq_b", |
| 264 | ) |
| 265 | |
| 266 | self.kv_norm = RMSNorm(self.head_dim, self.eps) |
| 267 | self.wo_a = ColumnParallelLinear( |
| 268 | self.n_heads * self.head_dim // self.n_groups, |
| 269 | self.n_groups * self.o_lora_rank, |
| 270 | bias=False, |
| 271 | quant_config=quant_config, |
| 272 | return_bias=False, |
| 273 | prefix=f"{prefix}.wo_a", |
| 274 | ) |
| 275 | self.wo_a.is_bmm = True |
| 276 | self.wo_a.bmm_batch_size = self.n_local_groups |
| 277 | self.wo_b = RowParallelLinear( |
| 278 | self.n_groups * self.o_lora_rank, |
| 279 | self.hidden_size, |
| 280 | bias=False, |
| 281 | quant_config=quant_config, |
| 282 | return_bias=False, |
| 283 | prefix=f"{prefix}.wo_b", |
| 284 | ) |
| 285 | |
| 286 | # Initialize rotary embedding before the indexer/compressor consume it. |
| 287 | self.rotary_emb = build_deepseek_v4_rope( |
| 288 | config, |
| 289 | head_dim=self.head_dim, |
| 290 | rope_head_dim=self.rope_head_dim, |
| 291 | max_position_embeddings=config.max_position_embeddings, |
| 292 | compress_ratio=self.compress_ratio, |
| 293 | ) |
| 294 | self.indexer_rotary_emb = self.rotary_emb |
| 295 | self.topk_indices_buffer = topk_indices_buffer |
| 296 | |
| 297 | self.indexer = None |
| 298 | if self.compress_ratio == 4: |
| 299 | # Only C4A uses sparse attention and hence has indexer. |
| 300 | # aux_stream_list[2] is free here (outer GEMMs joined) for the inner |
| 301 | # overlap of wq_b+fused_indexer_q_rope_quant vs compressor. None on |
| 302 | # ROCm, where aux_stream_list is None. |
| 303 | indexer_aux_stream = ( |
| 304 | aux_stream_list[2] if aux_stream_list is not None else None |
| 305 | ) |
| 306 | self.indexer = DeepseekV4Indexer( |
| 307 | vllm_config, |
| 308 | config=config, |
| 309 | hidden_size=self.hidden_size, |
| 310 | q_lora_rank=self.q_lora_rank, |
| 311 | quant_config=quant_config, |
| 312 | cache_config=cache_config, |
| 313 | topk_indices_buffer=topk_indices_buffer, |
| 314 | compress_ratio=self.compress_ratio, |
| 315 | prefix=f"{prefix}.indexer", |
| 316 | aux_stream=indexer_aux_stream, |
| 317 | ) |
| 318 | |
| 319 | self._prepare_and_attn_fn = self._prepare_and_attn |
| 320 | if not vllm_config.use_v2_model_runner: |
| 321 | # MRV1's piecewise capture only tolerates the wide eager region: with |
| 322 | # the narrow one the attention input preparation stays in the captured |
| 323 | # graph and MRV1 produces garbage (#51430). |
| 324 | self._prepare_and_attn_fn = self._prepare_and_attn_eager |
| 325 | |
| 326 | self.aux_stream_list = aux_stream_list |
| 327 | # [0]: GEMM start / post-GEMM event0. [1..3]: GEMM done events; |
| 328 | # [1] doubles as post-GEMM event1. Reuse is safe: GEMM fully joins |
| 329 | # before post-GEMM starts. |
| 330 | self.ln_events = [torch.cuda.Event() for _ in range(4)] |
| 331 | |
| 332 | assert cache_config is not None, "DeepseekV4 attention requires cache_config" |
| 333 | # ---- Attention / KV-cache setup ---- |
| 334 | self.max_num_batched_tokens = ( |
| 335 | vllm_config.scheduler_config.max_num_batched_tokens |
| 336 | ) |
| 337 | self.max_model_len = vllm_config.model_config.max_model_len |
| 338 | |
| 339 | # Resolve the kv-cache dtype from this backend's block format. The same |
| 340 | # resolution drives the SWA cache tensor dtype below. |
| 341 | self.kv_cache_dtype, self.kv_cache_torch_dtype = _resolve_dsv4_kv_cache_dtype( |
| 342 | self._uses_fp8_ds_mla_layout(), cache_config.cache_dtype, cache_config |
| 343 | ) |
| 344 | |
| 345 | self.swa_cache_layer = DeepseekV4SWACache( |
| 346 | head_dim=self.head_dim, |
| 347 | window_size=self.window_size, |
| 348 | dtype=self.kv_cache_torch_dtype, |
| 349 | prefix=f"{prefix}.swa_cache", |
| 350 | cache_config=cache_config, |
| 351 | backend_cls=self.swa_backend_cls, |
| 352 | ) |
| 353 | |
| 354 | # Register with compilation context for metadata lookup. |
| 355 | compilation_config = vllm_config.compilation_config |
| 356 | if prefix and prefix in compilation_config.static_forward_context: |
| 357 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 358 | if prefix: |
| 359 | compilation_config.static_forward_context[prefix] = self |
| 360 | self.kv_cache = torch.tensor([]) |
| 361 | |
| 362 | # Create the compressor for layers with compress_ratio > 1; after the |
| 363 | # attention setup above so its KV-cache prefix (self.prefix) is set. |
| 364 | self.compressor = None |
| 365 | if self.compress_ratio > 1: |
| 366 | self.compressor = DeepseekCompressor( |
| 367 | vllm_config=vllm_config, |
| 368 | compress_ratio=self.compress_ratio, |
| 369 | hidden_size=self.hidden_size, |
| 370 | head_dim=self.head_dim, |
| 371 | rotate=True, |
| 372 | prefix=f"{prefix}.compressor", |
| 373 | k_cache_prefix=self.prefix, |
| 374 | ) |
| 375 | |
| 376 | if vllm_config.kernel_config.enable_jit_warmup: |
| 377 | from vllm.v1.attention.backends.mla.sparse_swa import ( |
| 378 | _COMPUTE_PREFILL_METADATA_KERNEL, |
| 379 | _COMPUTE_SWA_INDICES_AND_LENS_KERNEL, |
| 380 | ) |
| 381 | |
| 382 | _COMPUTE_PREFILL_METADATA_KERNEL.register_warmup() |
| 383 | _COMPUTE_SWA_INDICES_AND_LENS_KERNEL.register_warmup( |
| 384 | window_size=self.window_size, |
| 385 | block_size=self.swa_cache_layer.block_size, |
| 386 | max_image_tokens=self.max_image_tokens, |
| 387 | ) |
| 388 | |
| 389 | if self.compress_ratio > 1: |
| 390 | from vllm.v1.attention.backends.mla.compressor_utils import ( |
| 391 | _COMPRESSED_SLOT_MAPPING_KERNEL, |
| 392 | ) |
| 393 | |
| 394 | _COMPRESSED_SLOT_MAPPING_KERNEL.register_warmup() |
| 395 | |
| 396 | if self.indexer is not None: |
| 397 | from vllm.v1.attention.backends.mla.indexer import ( |
| 398 | _BUILD_PREFILL_CHUNK_METADATA_KERNEL, |
| 399 | _PREPARE_UNIFORM_DECODE_KERNEL, |
| 400 | ) |
| 401 | |
| 402 | _PREPARE_UNIFORM_DECODE_KERNEL.register_warmup() |
| 403 | _BUILD_PREFILL_CHUNK_METADATA_KERNEL.register_warmup() |
| 404 | |
| 405 | spec_config = vllm_config.speculative_config |
| 406 | if spec_config is not None and spec_config.use_dspark(): |
| 407 | from vllm.v1.attention.backends.mla.sparse_swa import ( |
| 408 | _COMPUTE_DSPARK_NONCAUSAL_SWA_INDICES_KERNEL, |
| 409 | ) |
| 410 | |
| 411 | _COMPUTE_DSPARK_NONCAUSAL_SWA_INDICES_KERNEL.register_warmup( |
| 412 | window_size=self.window_size, |
| 413 | num_speculative_tokens=spec_config.num_speculative_tokens, |
| 414 | block_size=self.swa_cache_layer.block_size, |
| 415 | ) |
| 416 | |
| 417 | if self.backend_cls.get_name() in ( |
| 418 | "FLASHMLA_SPARSE_DSV4", |
| 419 | "ROCM_FLASHMLA_SPARSE_DSV4", |
| 420 | "XPU_V4_MLA_SPARSE", |
| 421 | ): |
| 422 | from vllm.models.deepseek_v4.common.ops.cache_utils import ( |
| 423 | _COMBINE_TOPK_SWA_INDICES_KERNEL, |
| 424 | ) |
| 425 | |
| 426 | _COMBINE_TOPK_SWA_INDICES_KERNEL.register_warmup() |
| 427 | |
| 428 | from vllm.utils.import_utils import has_cutedsl |
| 429 | |
| 430 | _FUSED_Q_KV_RMSNORM_KERNEL.register_warmup() |
| 431 | |
| 432 | backend_name = self.backend_cls.get_name() |
| 433 | if current_platform.is_cuda(): |
| 434 | from vllm.models.deepseek_v4.common.ops.fused_inv_rope_fp8_quant import ( # noqa: E501 |
| 435 | _FUSED_INV_ROPE_FP8_QUANT_KERNEL, |
| 436 | ) |
| 437 | |
| 438 | _FUSED_INV_ROPE_FP8_QUANT_KERNEL.register_warmup() |
| 439 | |
| 440 | if self.compress_ratio == 128: |
| 441 | from vllm.models.deepseek_v4.sparse_mla import ( |
| 442 | _BUILD_C128A_TOPK_METADATA_KERNEL, |
| 443 | ) |
| 444 | |
| 445 | _BUILD_C128A_TOPK_METADATA_KERNEL.register_warmup() |
| 446 | |
| 447 | if backend_name == "FLASHMLA_SPARSE_DSV4": |
| 448 | from vllm.models.deepseek_v4.common.ops.cache_utils import ( |
| 449 | _COMPUTE_GLOBAL_TOPK_INDICES_AND_LENS_KERNEL, |
| 450 | _DEQUANTIZE_AND_GATHER_K_CACHE_KERNEL, |
| 451 | ) |
| 452 | |
| 453 | _COMPUTE_GLOBAL_TOPK_INDICES_AND_LENS_KERNEL.register_warmup() |
| 454 | if has_cutedsl(): |
| 455 | from vllm.models.deepseek_v4.nvidia.ops.dequant_gather_k_cutedsl import ( # noqa: E501 |
| 456 | _DEQUANT_GATHER_K_CACHE_CUTEDSL_KERNEL, |
| 457 | ) |
| 458 | |
| 459 | _DEQUANT_GATHER_K_CACHE_CUTEDSL_KERNEL.register_warmup() |
| 460 | else: |
| 461 | _DEQUANTIZE_AND_GATHER_K_CACHE_KERNEL.register_warmup() |
| 462 | elif backend_name == "FLASHINFER_MLA_SPARSE_DSV4": |
| 463 | from vllm.models.deepseek_v4.common.ops.cache_utils import ( |
| 464 | _BUILD_FLASHINFER_MIXED_SPARSE_INDICES_KERNEL, |
| 465 | _COMPUTE_GLOBAL_TOPK_INDICES_AND_LENS_KERNEL, |
| 466 | ) |
| 467 | |
| 468 | _COMPUTE_GLOBAL_TOPK_INDICES_AND_LENS_KERNEL.register_warmup() |
| 469 | _BUILD_FLASHINFER_MIXED_SPARSE_INDICES_KERNEL.register_warmup() |
| 470 | |
| 471 | def forward( |
| 472 | self, |
| 473 | positions: torch.Tensor, |
| 474 | hidden_states: torch.Tensor, |
| 475 | llama_4_scaling: torch.Tensor | None = None, |
| 476 | ) -> torch.Tensor: |
| 477 | # The eager attention region writes into a caller-owned buffer |
| 478 | # (breakable_cudagraph needs in-place outputs). |
| 479 | attn_out = self._alloc_attn_out(hidden_states.shape[0], hidden_states) |
| 480 | |
| 481 | # Keep the attention input preparation in the captured graph. Only the |
| 482 | # sparse indexer and MLA attention run in the eager break below. |
| 483 | qr_kv, kv_score, indexer_kv_score, indexer_weights = ( |
| 484 | self._run_parallel_input_projections(hidden_states) |
| 485 | ) |
| 486 | qr, qr_scale, kv = self._split_qkv_and_norm(qr_kv) |
| 487 | |
| 488 | self._prepare_and_attn_fn( |
| 489 | hidden_states, |
| 490 | qr, |
| 491 | kv, |
| 492 | qr_scale, |
| 493 | kv_score, |
| 494 | indexer_kv_score, |
| 495 | indexer_weights, |
| 496 | positions, |
| 497 | attn_out, |
| 498 | ) |
| 499 | if isinstance(attn_out, torch.Tensor): |
| 500 | attn_out = attn_out[:, : self.n_local_heads, :] |
| 501 | return self._o_proj(attn_out, positions) |
| 502 | |
| 503 | def _alloc_attn_out( |
| 504 | self, num_tokens: int, hidden_states: torch.Tensor |
| 505 | ) -> "torch.Tensor | QuantizedActivation": |
| 506 | """The buffer ``forward_mqa`` fills. |
| 507 | |
| 508 | A bf16 ``[num_tokens, padded_heads, head_dim]`` buffer by default, whose |
| 509 | padding heads are sliced off before ``_o_proj``. A layer whose kernel |
| 510 | also does the inverse RoPE and the FP8 cast returns a |
| 511 | QuantizedActivation instead and gets it back in ``_o_proj`` whole. |
| 512 | """ |
| 513 | return torch.empty( |
| 514 | (num_tokens, self.padded_heads, self.head_dim), |
| 515 | dtype=hidden_states.dtype, |
| 516 | device=hidden_states.device, |
| 517 | ) |
| 518 | |
| 519 | def _split_qkv_and_norm( |
| 520 | self, qr_kv: torch.Tensor |
| 521 | ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor]: |
| 522 | """Split the fused q-lora / kv projection and RMSNorm both halves. |
| 523 | |
| 524 | ``qr_scale`` is None on the shared path. The ROCm subclass returns |
| 525 | a pre-quantized fp8 ``qr`` together with its per-1x128 fp32 scales, |
| 526 | which the downstream wq_b projections consume directly. |
| 527 | """ |
| 528 | qr, kv = qr_kv.split([self.q_lora_rank, self.head_dim], dim=-1) |
| 529 | qr, kv = fused_q_kv_rmsnorm( |
| 530 | qr, |
| 531 | kv, |
| 532 | self.q_norm.weight.data, |
| 533 | self.kv_norm.weight.data, |
| 534 | self.eps, |
| 535 | ) |
| 536 | return qr, None, kv |
| 537 | |
| 538 | @eager_break_during_capture |
| 539 | def _prepare_and_attn_eager( |
| 540 | self, |
| 541 | hidden_states: torch.Tensor, |
| 542 | qr: torch.Tensor, |
| 543 | kv: torch.Tensor, |
| 544 | qr_scale: torch.Tensor | None, |
| 545 | kv_score: torch.Tensor, |
| 546 | indexer_kv_score: torch.Tensor, |
| 547 | indexer_weights: torch.Tensor, |
| 548 | positions: torch.Tensor, |
| 549 | o_padded: "torch.Tensor | QuantizedActivation", |
| 550 | ) -> None: |
| 551 | """Wide eager region: the whole of ``_prepare_and_attn`` runs eagerly. |
| 552 | |
| 553 | The nested ``_sparse_indexer_and_attn`` break runs inline, since |
| 554 | ``add_eager`` clears ``_capturing`` before invoking this. |
| 555 | """ |
| 556 | self._prepare_and_attn( |
| 557 | hidden_states, |
| 558 | qr, |
| 559 | kv, |
| 560 | qr_scale, |
| 561 | kv_score, |
| 562 | indexer_kv_score, |
| 563 | indexer_weights, |
| 564 | positions, |
| 565 | o_padded, |
| 566 | ) |
| 567 | |
| 568 | def _prepare_and_attn( |
| 569 | self, |
| 570 | hidden_states: torch.Tensor, |
| 571 | qr: torch.Tensor, |
| 572 | kv: torch.Tensor, |
| 573 | qr_scale: torch.Tensor | None, |
| 574 | kv_score: torch.Tensor, |
| 575 | indexer_kv_score: torch.Tensor, |
| 576 | indexer_weights: torch.Tensor, |
| 577 | positions: torch.Tensor, |
| 578 | o_padded: "torch.Tensor | QuantizedActivation", |
| 579 | ) -> None: |
| 580 | """Attention input preparation followed by the sparse indexer and MLA. |
| 581 | |
| 582 | Only the latter runs in the eager break. |
| 583 | """ |
| 584 | attn_metadata = get_forward_context().attn_metadata |
| 585 | indexer = self.indexer |
| 586 | compressor = self.compressor |
| 587 | aux_streams = self.aux_stream_list |
| 588 | |
| 589 | def project_query_and_cache_kv() -> torch.Tensor: |
| 590 | q = self._wq_b_proj(qr, qr_scale).view( |
| 591 | -1, self.n_local_heads, self.head_dim |
| 592 | ) |
| 593 | return self._fused_qnorm_rope_kv_insert(q, kv, positions, attn_metadata) |
| 594 | |
| 595 | index_q: torch.Tensor | None = None |
| 596 | index_q_scale: torch.Tensor | None = None |
| 597 | index_weights_out: torch.Tensor | None = None |
| 598 | |
| 599 | # Keep Q projection and KV insertion on the default stream. The indexer |
| 600 | # and MLA compressor use aux streams 0 and 1; aux 2 is internal to the |
| 601 | # indexer. |
| 602 | if indexer is not None: |
| 603 | assert compressor is not None |
| 604 | q, (indexer_inputs, _) = execute_in_parallel( |
| 605 | project_query_and_cache_kv, |
| 606 | [ |
| 607 | lambda: indexer( |
| 608 | hidden_states, |
| 609 | qr, |
| 610 | indexer_kv_score, |
| 611 | indexer_weights, |
| 612 | positions, |
| 613 | self.indexer_rotary_emb, |
| 614 | qr_scale, |
| 615 | ), |
| 616 | lambda: compressor(kv_score, positions, self.rotary_emb), |
| 617 | ], |
| 618 | self.ln_events[0], |
| 619 | [self.ln_events[1], self.ln_events[2]], |
| 620 | [aux_streams[0], aux_streams[1]] if aux_streams is not None else None, |
| 621 | enable=aux_streams is not None, |
| 622 | ) |
| 623 | index_q, index_q_scale, index_weights_out = indexer_inputs |
| 624 | elif compressor is not None: |
| 625 | aux_stream = aux_streams[0] if aux_streams is not None else None |
| 626 | q, _ = maybe_execute_in_parallel( |
| 627 | project_query_and_cache_kv, |
| 628 | lambda: compressor(kv_score, positions, self.rotary_emb), |
| 629 | self.ln_events[0], |
| 630 | self.ln_events[1], |
| 631 | aux_stream, |
| 632 | ) |
| 633 | else: |
| 634 | q = project_query_and_cache_kv() |
| 635 | |
| 636 | self._sparse_indexer_and_attn( |
| 637 | hidden_states, |
| 638 | index_q, |
| 639 | index_q_scale, |
| 640 | index_weights_out, |
| 641 | q, |
| 642 | kv, |
| 643 | positions, |
| 644 | o_padded, |
| 645 | ) |
| 646 | |
| 647 | def _fused_wqa_wkv_gemm(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 648 | # Override point: the ROCm layer preshuffles this weight in place, so |
| 649 | # it cannot go through fused_wqa_wkv directly. |
| 650 | # MergedColumnParallelLinear returns (output, bias); bias is None. |
| 651 | qr_kv, _ = self.fused_wqa_wkv(hidden_states) |
| 652 | return qr_kv |
| 653 | |
| 654 | def _wq_b_proj( |
| 655 | self, qr: torch.Tensor, qr_scale: torch.Tensor | None = None |
| 656 | ) -> torch.Tensor: |
| 657 | """Project the q-lora input to the query heads. |
| 658 | |
| 659 | ``qr_scale`` is only set by the ROCm fused path, where ``qr`` is |
| 660 | already fp8-quantized with per-1x128 scales and must bypass the |
| 661 | linear's internal re-quantization (apply_weights would quantize |
| 662 | the fp8 input again). |
| 663 | """ |
| 664 | if qr_scale is None: |
| 665 | return self.wq_b(qr) |
| 666 | from vllm.models.deepseek_v4.amd.rocm import ( |
| 667 | apply_pre_quantized_block_scaled_mm, |
| 668 | ) |
| 669 | |
| 670 | return apply_pre_quantized_block_scaled_mm(self.wq_b, qr, qr_scale) |
| 671 | |
| 672 | def _run_parallel_input_projections( |
| 673 | self, hidden_states: torch.Tensor |
| 674 | ) -> tuple[ |
| 675 | torch.Tensor, |
| 676 | torch.Tensor | None, |
| 677 | torch.Tensor | None, |
| 678 | torch.Tensor | None, |
| 679 | ]: |
| 680 | aux_streams = self.aux_stream_list |
| 681 | if aux_streams is not None: |
| 682 | assert len(aux_streams) >= 3 |
| 683 | aux_streams = aux_streams[:3] |
| 684 | |
| 685 | # fused_wqa_wkv (heaviest) on default; the three lighter input GEMMs |
| 686 | # on aux streams 0..2 when their owning module exists. ln_events[0] |
| 687 | # is the fan-out start event; ln_events[1..3] are per-aux done events. |
| 688 | aux_fns: list[Callable[[], Any] | None] = [None, None, None] |
| 689 | |
| 690 | if self.compressor is not None: |
| 691 | # Local ref so the closure keeps a non-None type for mypy. |
| 692 | compressor = self.compressor |
| 693 | |
| 694 | def compressor_kv_score() -> torch.Tensor: |
| 695 | return torch.mm( |
| 696 | hidden_states, |
| 697 | compressor.fused_wkv_wgate.weight.T, |
| 698 | out_dtype=torch.float32, |
| 699 | ) |
| 700 | |
| 701 | aux_fns[0] = compressor_kv_score |
| 702 | |
| 703 | if self.indexer is not None: |
| 704 | indexer = self.indexer |
| 705 | |
| 706 | def indexer_weights_proj() -> torch.Tensor: |
| 707 | # ReplicatedLinear returns (output, bias); bias is None. |
| 708 | weights, _ = indexer.weights_proj(hidden_states) |
| 709 | return weights |
| 710 | |
| 711 | def indexer_compressor_kv_score() -> torch.Tensor: |
| 712 | return torch.mm( |
| 713 | hidden_states, |
| 714 | indexer.compressor.fused_wkv_wgate.weight.T, |
| 715 | out_dtype=torch.float32, |
| 716 | ) |
| 717 | |
| 718 | aux_fns[1] = indexer_weights_proj |
| 719 | aux_fns[2] = indexer_compressor_kv_score |
| 720 | |
| 721 | qr_kv, (kv_score, indexer_weights, indexer_kv_score) = execute_in_parallel( |
| 722 | lambda: self._fused_wqa_wkv_gemm(hidden_states), |
| 723 | aux_fns, |
| 724 | self.ln_events[0], |
| 725 | self.ln_events[1:4], |
| 726 | aux_streams, |
| 727 | enable=hidden_states.shape[0] |
| 728 | <= envs.VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD, |
| 729 | ) |
| 730 | |
| 731 | return qr_kv, kv_score, indexer_kv_score, indexer_weights |
| 732 | |
| 733 | @eager_break_during_capture |
| 734 | def _sparse_indexer_and_attn( |
| 735 | self, |
| 736 | hidden_states: torch.Tensor, |
| 737 | index_q: torch.Tensor | None, |
| 738 | index_q_scale: torch.Tensor | None, |
| 739 | index_weights: torch.Tensor | None, |
| 740 | q: torch.Tensor, |
| 741 | kv: torch.Tensor, |
| 742 | positions: torch.Tensor, |
| 743 | out: "torch.Tensor | QuantizedActivation", |
| 744 | ) -> None: |
| 745 | if self.indexer is not None and index_q is not None: |
| 746 | assert index_weights is not None |
| 747 | q_quant = (index_q, index_q_scale) if index_q_scale is not None else index_q |
| 748 | self.indexer.indexer_op( |
| 749 | hidden_states, |
| 750 | q_quant, |
| 751 | None, |
| 752 | index_weights, |
| 753 | ) |
| 754 | |
| 755 | # MLA attention writes into the pre-allocated `out` buffer |
| 756 | # (see _alloc_attn_out). |
| 757 | self.forward_mqa(q, kv, positions, out) |
| 758 | |
| 759 | def _fused_qnorm_rope_kv_insert( |
| 760 | self, |
| 761 | q: torch.Tensor, |
| 762 | kv: torch.Tensor, |
| 763 | positions: torch.Tensor, |
| 764 | attn_metadata: ( |
| 765 | dict[str, AttentionMetadata] | list[dict[str, AttentionMetadata]] | None |
| 766 | ), |
| 767 | ) -> torch.Tensor: |
| 768 | if not isinstance(attn_metadata, dict): |
| 769 | # Profile run: kernel doesn't fire; produce a padded tensor so |
| 770 | # downstream FlashMLA gets the right shape. |
| 771 | if self.n_local_heads < self.padded_heads: |
| 772 | return F.pad( |
| 773 | q, |
| 774 | (0, 0, 0, self.padded_heads - self.n_local_heads), |
| 775 | value=0.0, |
| 776 | ) |
| 777 | return q |
| 778 | |
| 779 | swa_metadata = cast( |
| 780 | "DeepseekSparseSWAMetadata | None", |
| 781 | attn_metadata.get(self.swa_cache_layer.prefix), |
| 782 | ) |
| 783 | assert swa_metadata is not None |
| 784 | |
| 785 | swa_kv_cache = self.swa_cache_layer.kv_cache |
| 786 | # The fused insert ops require int64 position_ids; the runner's positions |
| 787 | # buffer is already int64, so no cast is needed. |
| 788 | assert positions.dtype == torch.int64 |
| 789 | cos_sin_cache = self.rotary_emb.cos_sin_cache |
| 790 | cache_dtype = swa_kv_cache.dtype |
| 791 | |
| 792 | # kv is unchanged; attention reads kv solely via swa_kv_cache. |
| 793 | if cache_dtype == torch.uint8: |
| 794 | # fp8_ds_mla UE8M0 paged path. Horizontally fused: |
| 795 | # Q side: per-head RMSNorm (no weight) + GPT-J RoPE, zero-filling |
| 796 | # the padding head slots; the kernel allocates and returns |
| 797 | # the padded q tensor. |
| 798 | # KV side: GPT-J RoPE + UE8M0 FP8 quant + paged cache insert. |
| 799 | swa_kv_cache_2d = swa_kv_cache.view(swa_kv_cache.shape[0], -1) |
| 800 | return torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert( |
| 801 | q, |
| 802 | kv, |
| 803 | swa_kv_cache_2d, |
| 804 | swa_metadata.slot_mapping, |
| 805 | positions, |
| 806 | cos_sin_cache, |
| 807 | self.padded_heads, |
| 808 | self.eps, |
| 809 | swa_metadata.block_size, |
| 810 | ) |
| 811 | |
| 812 | # Plain-row path: the [num_blocks, block_size, 512] cache stores the KV |
| 813 | # row in its element dtype (no Q padding). bf16 rewrites q in place; |
| 814 | # per-tensor fp8 writes a separately-allocated fp8 q and quantizes the |
| 815 | # KV row. |
| 816 | block_size = swa_metadata.block_size |
| 817 | assert swa_kv_cache.shape[1:] == (block_size, self.head_dim) |
| 818 | swa_kv_cache_3d = swa_kv_cache |
| 819 | if cache_dtype == torch.bfloat16: |
| 820 | torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert( |
| 821 | q, |
| 822 | kv, |
| 823 | swa_kv_cache_3d, |
| 824 | swa_metadata.slot_mapping, |
| 825 | positions, |
| 826 | cos_sin_cache, |
| 827 | self.eps, |
| 828 | block_size, |
| 829 | ) |
| 830 | return q |
| 831 | |
| 832 | # per-tensor fp8 (torch.float8_e4m3fn) |
| 833 | q_fp8 = torch.empty_like(q, dtype=torch.float8_e4m3fn) |
| 834 | torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert( |
| 835 | q, |
| 836 | kv, |
| 837 | q_fp8, |
| 838 | swa_kv_cache_3d, |
| 839 | swa_metadata.slot_mapping, |
| 840 | positions, |
| 841 | cos_sin_cache, |
| 842 | self._flashinfer_fp8_kv_scale, |
| 843 | self._flashinfer_fp8_q_scale_inv, |
| 844 | self.eps, |
| 845 | block_size, |
| 846 | ) |
| 847 | return q_fp8 |
| 848 | |
| 849 | def bind_kv_cache(self, kv_cache: torch.Tensor) -> None: |
| 850 | # [B, H=1, N, C] -> [B, N, C] |
| 851 | self.kv_cache = kv_cache.squeeze(1) |
| 852 | |
| 853 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 854 | return self.backend_cls |
| 855 | |
| 856 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None: |
| 857 | if ( |
| 858 | self.compress_ratio <= 1 |
| 859 | ): # SWA part. Allocated separately as DeepseekV4SWACache. |
| 860 | return None |
| 861 | # fp8_ds_mla is a UE8M0 block-scaled uint8 layout and needs 576B |
| 862 | # alignment; plain bf16 / per-tensor fp8 rows use natural element-size |
| 863 | # pages. |
| 864 | uses_fp8_ds_mla_layout = self.kv_cache_dtype == "fp8_ds_mla" |
| 865 | return MLAAttentionSpec( |
| 866 | block_size=vllm_config.cache_config.block_size, |
| 867 | num_kv_heads=1, |
| 868 | max_tp_shards=1, |
| 869 | head_size=self.head_dim, |
| 870 | dtype=torch.uint8 if uses_fp8_ds_mla_layout else self.kv_cache_torch_dtype, |
| 871 | tokens_per_state=self.compress_ratio, |
| 872 | cache_dtype_str=self.kv_cache_dtype, |
| 873 | alignment=576 if uses_fp8_ds_mla_layout else 512, |
| 874 | model_version="deepseek_v4", |
| 875 | kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype), |
| 876 | # DeepseekV4: 448B NoPE + 128B RoPE + 8B fp8 scale = 584B per token; |
| 877 | # head_size stays semantic (512). |
| 878 | state_content_bytes=584 if uses_fp8_ds_mla_layout else None, |
| 879 | ) |
| 880 | |
| 881 | |
| 882 | class DeepseekV4IndexerCache(torch.nn.Module, AttentionLayerBase): |
| 883 | def __init__( |
| 884 | self, |
| 885 | head_dim: int, |
| 886 | dtype: torch.dtype, |
| 887 | prefix: str, |
| 888 | cache_config: CacheConfig, |
| 889 | compress_ratio: int = 1, |
| 890 | ): |
| 891 | super().__init__() |
| 892 | self.kv_cache = torch.tensor([]) |
| 893 | self.head_dim = head_dim |
| 894 | self.prefix = prefix |
| 895 | self.cache_config = cache_config |
| 896 | self.dtype = dtype |
| 897 | self.compress_ratio = compress_ratio |
| 898 | compilation_config = get_current_vllm_config().compilation_config |
| 899 | if prefix in compilation_config.static_forward_context: |
| 900 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 901 | compilation_config.static_forward_context[prefix] = self |
| 902 | |
| 903 | def bind_kv_cache(self, kv_cache: torch.Tensor) -> None: |
| 904 | # [B, H=1, N, C] -> [B, N, C] |
| 905 | self.kv_cache = kv_cache.squeeze(1) |
| 906 | |
| 907 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec: |
| 908 | # head_dim already carries the fp8 scale padding |
| 909 | # tokens_per_state=1 for V3.2, >1 for DeepseekV4; same cache layout. |
| 910 | uses_fp8_ds_mla_layout = vllm_config.cache_config.cache_dtype == "fp8_ds_mla" |
| 911 | return MLAAttentionSpec( |
| 912 | block_size=self.cache_config.block_size, |
| 913 | num_kv_heads=1, |
| 914 | max_tp_shards=1, |
| 915 | head_size=self.head_dim, |
| 916 | dtype=self.dtype, |
| 917 | tokens_per_state=self.compress_ratio, |
| 918 | # 576B for FlashMLA packing; 512B for FlashInfer sparse (#44577). |
| 919 | alignment=576 if uses_fp8_ds_mla_layout else 512, |
| 920 | ) |
| 921 | |
| 922 | def forward(self): ... |
| 923 | |
| 924 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 925 | return DeepseekV4IndexerBackend |
| 926 | |
| 927 | |
| 928 | class DeepseekV4Indexer(nn.Module): |
| 929 | def __init__( |
| 930 | self, |
| 931 | vllm_config: VllmConfig, |
| 932 | config: DeepseekV2Config | DeepseekV3Config, |
| 933 | hidden_size: int, |
| 934 | q_lora_rank: int, |
| 935 | quant_config: QuantizationConfig | None, |
| 936 | cache_config: CacheConfig | None, |
| 937 | topk_indices_buffer: torch.Tensor | None, |
| 938 | compress_ratio: int = 1, |
| 939 | prefix: str = "", |
| 940 | aux_stream: torch.cuda.Stream | None = None, |
| 941 | ): |
| 942 | super().__init__() |
| 943 | self.vllm_config = vllm_config |
| 944 | self.config = config |
| 945 | self.quant_config = quant_config |
| 946 | # self.indexer_cfg = config.attn_module_list_cfg[0]["attn_index"] |
| 947 | self.topk_tokens = config.index_topk |
| 948 | self.n_head = config.index_n_heads # 64 |
| 949 | self.head_dim = config.index_head_dim # 128 |
| 950 | self.rope_dim = config.qk_rope_head_dim # 64 |
| 951 | self.q_lora_rank = q_lora_rank # 1536 |
| 952 | self.compress_ratio = compress_ratio |
| 953 | self.use_fp4_kv = dsa_indexer_uses_fp4(vllm_config) |
| 954 | logger.info_once( |
| 955 | "Using %s indexer cache for Lightning Indexer.", |
| 956 | "MXFP4" if self.use_fp4_kv else "FP8", |
| 957 | ) |
| 958 | if vllm_config.kernel_config.enable_jit_warmup: |
| 959 | from vllm.utils.import_utils import has_cutedsl |
| 960 | |
| 961 | if current_platform.is_cuda() and has_cutedsl(): |
| 962 | from vllm.models.deepseek_v4.nvidia.ops.fused_indexer_q_cutedsl import ( # noqa: E501 |
| 963 | _INDEXER_Q_FP8_KERNEL, |
| 964 | _INDEXER_Q_MXFP4_KERNEL, |
| 965 | ) |
| 966 | |
| 967 | indexer_q_kernel = ( |
| 968 | _INDEXER_Q_MXFP4_KERNEL |
| 969 | if self.use_fp4_kv |
| 970 | else _INDEXER_Q_FP8_KERNEL |
| 971 | ) |
| 972 | indexer_q_kernel.register_warmup() |
| 973 | |
| 974 | # no tensor parallel, just replicated |
| 975 | self.wq_b = ReplicatedLinear( |
| 976 | self.q_lora_rank, |
| 977 | self.head_dim * self.n_head, |
| 978 | bias=False, |
| 979 | quant_config=quant_config, |
| 980 | prefix=f"{prefix}.wq_b", |
| 981 | ) |
| 982 | self.weights_proj = ReplicatedLinear( |
| 983 | hidden_size, |
| 984 | self.n_head, |
| 985 | bias=False, |
| 986 | quant_config=None, |
| 987 | prefix=f"{prefix}.weights_proj", |
| 988 | ) |
| 989 | self.softmax_scale = self.head_dim**-0.5 |
| 990 | |
| 991 | self.scale_fmt = "ue8m0" |
| 992 | self.quant_block_size = 128 # TODO: get from config |
| 993 | self.topk_indices_buffer = topk_indices_buffer |
| 994 | |
| 995 | self.max_model_len = ( |
| 996 | vllm_config.model_config.max_model_len // self.compress_ratio |
| 997 | ) |
| 998 | self.prefix = prefix |
| 999 | |
| 1000 | self.max_total_seq_len = ( |
| 1001 | get_max_prefill_buffer_size(vllm_config) // self.compress_ratio |
| 1002 | ) |
| 1003 | |
| 1004 | assert cache_config is not None, "Deepseek V4 indexer requires cache_config" |
| 1005 | if self.use_fp4_kv: |
| 1006 | # MXFP4 stores two values per byte plus one UE8M0 byte per 32 values. |
| 1007 | # head_dim bytes = 64 packed values + 4 UE8M0 scales = 68. |
| 1008 | k_cache_head_dim = self.head_dim // 2 + self.head_dim // MXFP4_BLOCK_SIZE |
| 1009 | else: |
| 1010 | # NOTE(yifan): FP8 indexer cache uses the same layout as V3.2: |
| 1011 | # head_dim bytes = 128 fp8 + 4 fp32 scale = 132. |
| 1012 | k_cache_head_dim = ( |
| 1013 | self.head_dim + self.head_dim // self.quant_block_size * 4 |
| 1014 | ) |
| 1015 | self.k_cache = DeepseekV4IndexerCache( |
| 1016 | head_dim=k_cache_head_dim, |
| 1017 | dtype=torch.uint8, |
| 1018 | prefix=f"{prefix}.k_cache", |
| 1019 | cache_config=cache_config, |
| 1020 | compress_ratio=self.compress_ratio, |
| 1021 | ) |
| 1022 | self.compressor = DeepseekCompressor( |
| 1023 | vllm_config=vllm_config, |
| 1024 | compress_ratio=self.compress_ratio, |
| 1025 | hidden_size=hidden_size, |
| 1026 | head_dim=self.head_dim, |
| 1027 | rotate=True, |
| 1028 | prefix=f"{prefix}.compressor", |
| 1029 | k_cache_prefix=self.k_cache.prefix, |
| 1030 | use_fp4_cache=self.use_fp4_kv, |
| 1031 | ) |
| 1032 | |
| 1033 | self.indexer_op = SparseAttnIndexer( |
| 1034 | self.k_cache, |
| 1035 | self.quant_block_size, |
| 1036 | self.scale_fmt, |
| 1037 | self.topk_tokens, |
| 1038 | self.head_dim, |
| 1039 | self.max_model_len, |
| 1040 | self.max_total_seq_len, |
| 1041 | self.topk_indices_buffer, |
| 1042 | skip_k_cache_insert=True, |
| 1043 | use_fp4_cache=self.use_fp4_kv, |
| 1044 | compress_ratio=self.compress_ratio, |
| 1045 | ) |
| 1046 | |
| 1047 | # None on ROCm — maybe_execute_in_parallel falls back to sequential. |
| 1048 | self.aux_stream = aux_stream |
| 1049 | self.ln_events: list[torch.cuda.Event] = [ |
| 1050 | torch.cuda.Event(), |
| 1051 | torch.cuda.Event(), |
| 1052 | ] |
| 1053 | |
| 1054 | if vllm_config.kernel_config.enable_jit_warmup: |
| 1055 | from vllm.utils.import_utils import has_cutedsl |
| 1056 | |
| 1057 | if not has_cutedsl() and not current_platform.is_xpu(): |
| 1058 | from vllm.models.deepseek_v4.common.ops.fused_indexer_q import ( |
| 1059 | _FUSED_INDEXER_Q_ROPE_MXFP4_TRITON_KERNEL, |
| 1060 | _FUSED_INDEXER_Q_ROPE_QUANT_TRITON_KERNEL, |
| 1061 | ) |
| 1062 | |
| 1063 | ( |
| 1064 | _FUSED_INDEXER_Q_ROPE_MXFP4_TRITON_KERNEL |
| 1065 | if self.use_fp4_kv |
| 1066 | else _FUSED_INDEXER_Q_ROPE_QUANT_TRITON_KERNEL |
| 1067 | ).register_warmup() |
| 1068 | |
| 1069 | def forward( |
| 1070 | self, |
| 1071 | hidden_states: torch.Tensor, |
| 1072 | qr: torch.Tensor, |
| 1073 | compressed_kv_score: torch.Tensor | None, |
| 1074 | indexer_weights: torch.Tensor, |
| 1075 | positions: torch.Tensor, |
| 1076 | rotary_emb: nn.Module, |
| 1077 | qr_scale: torch.Tensor | None = None, |
| 1078 | skip_compressor: bool = False, |
| 1079 | ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor | None]: |
| 1080 | compressor = self.compressor |
| 1081 | |
| 1082 | attn_metadata = get_forward_context().attn_metadata |
| 1083 | if isinstance(attn_metadata, dict): |
| 1084 | indexer_metadata = cast(Any, attn_metadata[self.k_cache.prefix]) |
| 1085 | if ( |
| 1086 | indexer_metadata.max_seq_len // self.compress_ratio <= self.topk_tokens |
| 1087 | and not torch.cuda.is_current_stream_capturing() |
| 1088 | ): |
| 1089 | # candidates num smaller than topk, every candidate is selected |
| 1090 | # but we still need to build k cache |
| 1091 | if not skip_compressor: |
| 1092 | compressor(compressed_kv_score, positions, rotary_emb) |
| 1093 | assert self.topk_indices_buffer is not None |
| 1094 | num_tokens = ( |
| 1095 | indexer_metadata.num_decode_tokens |
| 1096 | + indexer_metadata.num_prefill_tokens |
| 1097 | ) |
| 1098 | if num_tokens > 0: |
| 1099 | _fill_short_context_topk_indices[(num_tokens,)]( |
| 1100 | self.topk_indices_buffer, |
| 1101 | positions, |
| 1102 | TOP_K=self.topk_tokens, |
| 1103 | COMPRESS_RATIO=self.compress_ratio, |
| 1104 | PADDED_TOP_K=triton.next_power_of_2(self.topk_tokens), |
| 1105 | num_warps=8, |
| 1106 | ) |
| 1107 | return None, None, None |
| 1108 | |
| 1109 | def wq_b_and_q_quant(): |
| 1110 | q = self._wq_b_proj(qr, qr_scale) |
| 1111 | q = q.view(-1, self.n_head, self.head_dim) |
| 1112 | return fused_indexer_q_rope_quant( |
| 1113 | positions, |
| 1114 | q, |
| 1115 | rotary_emb.cos_sin_cache, |
| 1116 | indexer_weights, |
| 1117 | self.softmax_scale, |
| 1118 | self.n_head**-0.5, |
| 1119 | use_fp4=self.use_fp4_kv, |
| 1120 | ) |
| 1121 | |
| 1122 | if not skip_compressor: |
| 1123 | # compressor returns None and writes K to the indexer KV cache; the |
| 1124 | # join orders that write before indexer_op (skip_k_cache_insert=True). |
| 1125 | (q_quant, weights), _ = maybe_execute_in_parallel( |
| 1126 | wq_b_and_q_quant, |
| 1127 | lambda: compressor(compressed_kv_score, positions, rotary_emb), |
| 1128 | self.ln_events[0], |
| 1129 | self.ln_events[1], |
| 1130 | self.aux_stream, |
| 1131 | ) |
| 1132 | else: |
| 1133 | q_quant, weights = wq_b_and_q_quant() |
| 1134 | if isinstance(q_quant, tuple): |
| 1135 | q, q_scale = q_quant |
| 1136 | else: |
| 1137 | q, q_scale = q_quant, None |
| 1138 | return q, q_scale, weights |
| 1139 | |
| 1140 | def _wq_b_proj( |
| 1141 | self, qr: torch.Tensor, qr_scale: torch.Tensor | None = None |
| 1142 | ) -> torch.Tensor: |
| 1143 | """Project the q-lora input to the indexer query heads. |
| 1144 | |
| 1145 | Same bypass as ``DeepseekV4Attention._wq_b_proj``: on ROCm the fused |
| 1146 | norm path hands over pre-quantized fp8 ``qr`` with per-1x128 scales, |
| 1147 | so the linear's internal re-quantization must be skipped. |
| 1148 | """ |
| 1149 | if qr_scale is None: |
| 1150 | # ReplicatedLinear returns (output, bias); bias is None. |
| 1151 | q, _ = self.wq_b(qr) |
| 1152 | return q |
| 1153 | from vllm.models.deepseek_v4.amd.rocm import ( |
| 1154 | apply_pre_quantized_block_scaled_mm, |
| 1155 | ) |
| 1156 | |
| 1157 | return apply_pre_quantized_block_scaled_mm(self.wq_b, qr, qr_scale) |