| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """DeepseekV4 MLA Attention Layer""" |
| 4 | |
| 5 | import math |
| 6 | from abc import ABC, abstractmethod |
| 7 | from collections.abc import Callable |
| 8 | from functools import cached_property |
| 9 | from typing import TYPE_CHECKING, Any, ClassVar, cast |
| 10 | |
| 11 | import regex as re |
| 12 | import torch |
| 13 | import torch.nn as nn |
| 14 | import torch.nn.functional as F |
| 15 | from transformers import DeepseekV2Config, DeepseekV3Config |
| 16 | |
| 17 | import vllm.envs as envs |
| 18 | from vllm.compilation.breakable_cudagraph import eager_break_during_capture |
| 19 | from vllm.model_executor.layers.fusion.quant_activation import QuantizedActivation |
| 20 | from vllm.model_executor.layers.linear import ( |
| 21 | ColumnParallelLinear, |
| 22 | MergedColumnParallelLinear, |
| 23 | ReplicatedLinear, |
| 24 | RowParallelLinear, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer |
| 27 | from vllm.model_executor.layers.sparse_mqa_indexer import SparseMQAIndexer |
| 28 | from vllm.models.common.ops import fused_q_kv_rmsnorm |
| 29 | from vllm.models.common.ops.sequence_parallel import sp_reduce_scatter |
| 30 | from vllm.models.deepseek_v41.common.ops import ( |
| 31 | MXFP4_BLOCK_SIZE, |
| 32 | fused_indexer_q_rope_quant, |
| 33 | indexer_k_norm_rope_store, |
| 34 | ) |
| 35 | |
| 36 | if TYPE_CHECKING: |
| 37 | from vllm.model_executor.kernels.linear.cute_dsl.gemm_rs_ar import GemmRsAr |
| 38 | from vllm.model_executor.layers.rocm_paged_mxfp4_indexer import ( |
| 39 | RocmSparseMQAIndexer, |
| 40 | ) |
| 41 | from vllm.v1.attention.backends.mla.sparse_swa import ( |
| 42 | DeepseekSparseSWAMetadata, |
| 43 | ) |
| 44 | |
| 45 | from vllm.config import ( |
| 46 | CacheConfig, |
| 47 | VllmConfig, |
| 48 | get_current_vllm_config, |
| 49 | ) |
| 50 | from vllm.config.cache import CacheDType |
| 51 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 52 | from vllm.forward_context import get_forward_context |
| 53 | from vllm.logger import init_logger |
| 54 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 55 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 56 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 57 | from vllm.model_executor.models.utils import extract_layer_index |
| 58 | from vllm.models.deepseek_v41.common.rope import build_deepseek_v4_rope |
| 59 | from vllm.models.deepseek_v41.compressor import DeepseekCompressor |
| 60 | from vllm.platforms import current_platform |
| 61 | from vllm.triton_utils import tl, triton |
| 62 | from vllm.utils.multi_stream_utils import ( |
| 63 | execute_in_parallel, |
| 64 | maybe_execute_in_parallel, |
| 65 | ) |
| 66 | from vllm.v1.attention.backend import AttentionBackend, AttentionMetadata |
| 67 | from vllm.v1.attention.backends.mla.indexer import ( |
| 68 | DeepseekV41IndexerBackend, |
| 69 | dsa_indexer_uses_fp4, |
| 70 | get_max_prefill_buffer_size, |
| 71 | ) |
| 72 | from vllm.v1.attention.backends.mla.sparse_indexer import ( |
| 73 | DeepseekV41SparseIndexerBackend, |
| 74 | ) |
| 75 | from vllm.v1.attention.backends.mla.sparse_swa import DeepseekV4SWACache |
| 76 | from vllm.v1.kv_cache_interface import ( |
| 77 | KVCacheSpec, |
| 78 | MLAAttentionSpec, |
| 79 | get_kv_quant_mode, |
| 80 | ) |
| 81 | |
| 82 | logger = init_logger(__name__) |
| 83 | |
| 84 | |
| 85 | def _replace_layer_index(prefix: str, layer_id: int) -> str: |
| 86 | """Swap the layer index inside a ``...layers.<idx>...`` prefix.""" |
| 87 | new_prefix, n = re.subn(r"\.layers\.\d+\.", f".layers.{layer_id}.", prefix) |
| 88 | assert n == 1, f"Cannot locate layer index in prefix {prefix}" |
| 89 | return new_prefix |
| 90 | |
| 91 | |
| 92 | def _indexer_k_cache_head_dim(index_head_dim: int, use_fp4_kv: bool) -> int: |
| 93 | """Per-token byte width of the paged indexer K cache row.""" |
| 94 | if use_fp4_kv: |
| 95 | # MXFP4 stores two values per byte plus one UE8M0 byte per 32 values. |
| 96 | # head_dim bytes = 64 packed values + 4 UE8M0 scales = 68. |
| 97 | return index_head_dim // 2 + index_head_dim // MXFP4_BLOCK_SIZE |
| 98 | # NOTE(yifan): FP8 indexer cache uses the same layout as V3.2: |
| 99 | # head_dim bytes = 128 fp8 + 4 fp32 scale = 132. |
| 100 | return index_head_dim + index_head_dim // 128 * 4 |
| 101 | |
| 102 | |
| 103 | @triton.jit |
| 104 | def _fill_short_context_topk_indices( |
| 105 | output, |
| 106 | positions, |
| 107 | TOP_K: tl.constexpr, |
| 108 | COMPRESS_RATIO: tl.constexpr, |
| 109 | PADDED_TOP_K: tl.constexpr, |
| 110 | ): |
| 111 | # small triton kernel that selects every candidate, -1 otherwise |
| 112 | row = tl.program_id(0) |
| 113 | offsets = tl.arange(0, PADDED_TOP_K) |
| 114 | num_compressed = (tl.load(positions + row) + 1) // COMPRESS_RATIO |
| 115 | tl.store( |
| 116 | output + row * TOP_K + offsets, |
| 117 | tl.where(offsets < num_compressed, offsets, -1), |
| 118 | mask=offsets < TOP_K, |
| 119 | ) |
| 120 | |
| 121 | |
| 122 | # Which packed fp8_ds_mla record a V4.1 layer writes. FlashMLA decodes |
| 123 | # DeepSeek's V4.1 record -- all 512 dims (RoPE included) as fp8 e4m3 with one |
| 124 | # UE8M0 scale per 32 dims, 512 data bytes + 16 scale bytes per token, pages |
| 125 | # rounded to the kernel's 512 B TMA stride -- only in its SM100 sparse-decode |
| 126 | # kernels. Every other arch keeps the V4 record: 448 fp8 NoPE + 64 bf16 RoPE |
| 127 | # plus 7 UE8M0 scales of 64 dims and a pad byte (584 B, 576 B pages). |
| 128 | def _use_v41_mxfp8_kv_record() -> bool: |
| 129 | return current_platform.is_device_capability_family(100) |
| 130 | |
| 131 | |
| 132 | def _resolve_dsv4_kv_cache_dtype( |
| 133 | use_fp8_ds_mla_layout: bool, |
| 134 | kv_cache_dtype: CacheDType, |
| 135 | cache_config: CacheConfig | None, |
| 136 | packed_kv_cache_dtype: CacheDType = "fp8_ds_mla", |
| 137 | ) -> tuple[CacheDType, torch.dtype]: |
| 138 | """Map ``(layout, --kv-cache-dtype)`` to ``(cache_dtype_str, torch_dtype)``. |
| 139 | |
| 140 | Both layouts are paged; they differ in the per-token block format. The |
| 141 | packed formats are ``uint8``-backed: ``fp8_ds_mla`` is UE8M0 block-scaled |
| 142 | fp8 throughout, ``nvfp4_ds_mla`` keeps that sliding-window record and |
| 143 | stores the compressed cache as NVFP4. An unspecific ``--kv-cache-dtype`` |
| 144 | (``auto`` / ``fp8``) resolves to ``packed_kv_cache_dtype``, the record this |
| 145 | layer's kernel prefers, and the canonical string is written back onto |
| 146 | ``cache_config`` so the page-size specs pick the right per-token slot. |
| 147 | Plain-row backends store each token's KV row in its element dtype: bf16 or |
| 148 | per-tensor FP8 E4M3. |
| 149 | """ |
| 150 | if use_fp8_ds_mla_layout: |
| 151 | if kv_cache_dtype in ("auto", "fp8"): |
| 152 | kv_cache_dtype = packed_kv_cache_dtype |
| 153 | elif not kv_cache_dtype.endswith("_ds_mla"): |
| 154 | raise ValueError( |
| 155 | "DeepseekV4 packed KV layouts only support fp8 kv-cache, got " |
| 156 | f"{kv_cache_dtype}. Please set `--kv-cache-dtype fp8` or " |
| 157 | "select a backend that supports bfloat16 KV cache." |
| 158 | ) |
| 159 | if cache_config is not None and cache_config.cache_dtype != kv_cache_dtype: |
| 160 | cache_config.cache_dtype = kv_cache_dtype |
| 161 | logger.info_once("Using DeepSeek's %s KV cache format.", kv_cache_dtype) |
| 162 | return kv_cache_dtype, torch.uint8 |
| 163 | if kv_cache_dtype.endswith("_ds_mla"): |
| 164 | raise ValueError( |
| 165 | f"{kv_cache_dtype} is a packed FlashMLA DeepSeek V4.1 KV cache " |
| 166 | "format; the selected backend stores plain KV rows." |
| 167 | ) |
| 168 | |
| 169 | # Plain bf16 / per-tensor fp8 KV row (FlashInfer). |
| 170 | if kv_cache_dtype.startswith("fp8"): |
| 171 | return kv_cache_dtype, torch.float8_e4m3fn |
| 172 | # auto / bfloat16 -> plain bf16 KV row. |
| 173 | return kv_cache_dtype, torch.bfloat16 |
| 174 | |
| 175 | |
| 176 | class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC): |
| 177 | """DeepseekV4 MLA attention layer. |
| 178 | |
| 179 | The platform-specific sparse-MLA forward (``forward_mqa`` / |
| 180 | ``get_padded_num_q_heads`` / ``_o_proj`` / ``backend_cls``) is provided by a |
| 181 | subclass — ``DeepseekV4FlashMLAAttention`` / |
| 182 | ``DeepseekV4FlashInferSM120Attention`` / |
| 183 | ``DeepseekV4FlashInferMLAAttention`` (CUDA) or |
| 184 | ``DeepseekV41ROCMAiterMLAAttention`` (ROCm) — selected by the platform-specific |
| 185 | deepseek_v41 model module. The base is never instantiated directly. |
| 186 | """ |
| 187 | |
| 188 | # Provided by the platform subclass. |
| 189 | backend_cls: ClassVar[type[AttentionBackend]] |
| 190 | # Backend for the SWA cache layer; None uses the default SWA backend. |
| 191 | swa_backend_cls: ClassVar[type[AttentionBackend] | None] = None |
| 192 | # KV-cache per-token block format (both layouts are paged). True (default) |
| 193 | # = fp8_ds_mla (UE8M0 block-scaled fp8 packed as uint8); False = plain |
| 194 | # bf16 / per-tensor fp8 KV row. Backends can override the instance hook when |
| 195 | # a single attention class dispatches across arch-specific layouts. |
| 196 | use_fp8_ds_mla_layout: ClassVar[bool] = True |
| 197 | # Prefill is processed in fixed-size chunks; this bounds the bf16 kv-gather |
| 198 | # workspace allocated in _forward_prefill and is also read by the dummy-run |
| 199 | # path to pre-reserve that workspace. |
| 200 | PREFILL_CHUNK_SIZE: ClassVar[int] = 4 |
| 201 | |
| 202 | # ---- attention-interface contract, declared by the platform subclass ---- |
| 203 | |
| 204 | @property |
| 205 | def accepts_unnormed_unroped_query(self) -> bool: |
| 206 | """Whether ``forward_mqa``'s ``q`` is the raw ``wq_b`` output. |
| 207 | |
| 208 | True when the attention kernel applies the Q norm and RoPE itself, and |
| 209 | reads Q in its own chunk-interleaved layout, so the layer only |
| 210 | zero-pads Q to ``padded_heads`` and inserts KV before calling it. |
| 211 | """ |
| 212 | return False |
| 213 | |
| 214 | @property |
| 215 | def packed_kv_cache_dtype(self) -> CacheDType: |
| 216 | """The packed KV record this layer's kernel prefers. |
| 217 | |
| 218 | What an unspecific ``--kv-cache-dtype`` (``auto`` / ``fp8``) resolves |
| 219 | to. Mega attention overrides it: its kernel is the one that can read |
| 220 | an NVFP4 compressed cache. |
| 221 | """ |
| 222 | return "fp8_ds_mla" |
| 223 | |
| 224 | @classmethod |
| 225 | @abstractmethod |
| 226 | def get_padded_num_q_heads(cls, num_heads: int) -> int: |
| 227 | """Q head count the q/output buffers are allocated at. |
| 228 | |
| 229 | The layer allocates the q/output buffers at |
| 230 | ``[N, get_padded_num_q_heads(n_local_heads), head_dim]``. Must satisfy |
| 231 | ``result >= num_heads``. Backends with no padding constraint return |
| 232 | ``num_heads``. |
| 233 | """ |
| 234 | raise NotImplementedError |
| 235 | |
| 236 | @abstractmethod |
| 237 | def forward_mqa( |
| 238 | self, |
| 239 | q: torch.Tensor, |
| 240 | kv: torch.Tensor, |
| 241 | positions: torch.Tensor, |
| 242 | output: torch.Tensor, |
| 243 | ) -> None: |
| 244 | """Platform-specific sparse MLA forward; writes attention into ``output``.""" |
| 245 | raise NotImplementedError |
| 246 | |
| 247 | @abstractmethod |
| 248 | def _o_proj( |
| 249 | self, |
| 250 | attn_out: "torch.Tensor | QuantizedActivation", |
| 251 | positions: torch.Tensor, |
| 252 | ) -> torch.Tensor: |
| 253 | """Project whatever ``_alloc_attn_out`` produced through wo_a and wo_b. |
| 254 | |
| 255 | Takes the buffer whole, so each layer owns the shape it allocated: the |
| 256 | bf16 layers slice off their padding heads and apply the inverse RoPE, |
| 257 | while a layer whose attention kernel already did the inverse RoPE and |
| 258 | the FP8 cast gets a QuantizedActivation and has only wo_a and wo_b |
| 259 | left. |
| 260 | """ |
| 261 | raise NotImplementedError |
| 262 | |
| 263 | def _uses_fp8_ds_mla_layout(self) -> bool: |
| 264 | """Return whether this instance stores fp8 KV in fp8_ds_mla layout.""" |
| 265 | return self.use_fp8_ds_mla_layout |
| 266 | |
| 267 | def _indexer_cls( |
| 268 | self, k_cache: "DeepseekV4IndexerCache | None" |
| 269 | ) -> type["DeepseekV4Indexer"]: |
| 270 | """The indexer for this layer's index K cache. A platform subclass |
| 271 | returns one that writes and scores its own cache layout.""" |
| 272 | return DeepseekV4Indexer |
| 273 | |
| 274 | def __init__( |
| 275 | self, |
| 276 | vllm_config: VllmConfig, |
| 277 | prefix: str, |
| 278 | topk_indices_buffer: torch.Tensor | None = None, |
| 279 | aux_stream_list: list[torch.cuda.Stream] | None = None, |
| 280 | candidate_block_buffer: torch.Tensor | None = None, |
| 281 | ) -> None: |
| 282 | super().__init__() |
| 283 | config = vllm_config.model_config.hf_config |
| 284 | quant_config = vllm_config.quant_config |
| 285 | cache_config = vllm_config.cache_config |
| 286 | tp_size = get_tensor_model_parallel_world_size() |
| 287 | layer_id = extract_layer_index(prefix) |
| 288 | self.layer_id = layer_id |
| 289 | |
| 290 | self.prefix = prefix # Alias for compatibility with compressor |
| 291 | self.hidden_size = config.hidden_size |
| 292 | self.n_heads = config.num_attention_heads |
| 293 | assert self.n_heads % tp_size == 0 |
| 294 | self.n_local_heads = self.n_heads // tp_size |
| 295 | self.q_lora_rank = config.q_lora_rank |
| 296 | self.o_lora_rank = config.o_lora_rank |
| 297 | self.head_dim = config.head_dim |
| 298 | self.rope_head_dim = config.qk_rope_head_dim |
| 299 | self.nope_head_dim = self.head_dim - self.rope_head_dim |
| 300 | self.n_groups = config.o_groups |
| 301 | self.n_local_groups = self.n_groups // tp_size |
| 302 | self.window_size = config.sliding_window |
| 303 | # ---- v4.1 sparse-attention topology ---- |
| 304 | # compress_ratios has one entry per layer (MTP layers included): |
| 305 | # 0 = pure sliding window, 1 = full-length compressed cache, |
| 306 | # 2 = ratio-2 compressed. Compressors and compressed-KV caches live |
| 307 | # only on ``kv_source_layer_ids``; indexers only on |
| 308 | # ``index_source_layer_ids``. Consumers reuse the most recently |
| 309 | # published source below them. |
| 310 | compress_ratios = getattr(config, "compress_ratios", None) |
| 311 | if compress_ratios is not None and layer_id < len(compress_ratios): |
| 312 | self.compress_ratio = int(compress_ratios[layer_id]) |
| 313 | else: |
| 314 | # MTP layers past the configured list are pure sliding-window. |
| 315 | self.compress_ratio = 0 |
| 316 | if self.compress_ratio not in (0, 1, 2): |
| 317 | raise ValueError( |
| 318 | f"DeepSeek V4.1 layer {layer_id} has compress_ratio=" |
| 319 | f"{self.compress_ratio}; only 0 (sliding window), 1 and 2 are " |
| 320 | "supported." |
| 321 | ) |
| 322 | self.kv_source_layers = tuple( |
| 323 | getattr(config, "kv_source_layer_ids", None) or () |
| 324 | ) |
| 325 | self.index_source_layers = tuple( |
| 326 | getattr(config, "index_source_layer_ids", None) or () |
| 327 | ) |
| 328 | self.candidate_source_layer = getattr(config, "candidate_source_layer_id", -1) |
| 329 | self.candidate_topk_blocks = getattr(config, "candidate_topk_blocks", 0) |
| 330 | self.candidate_block_size = getattr(config, "candidate_block_size", 0) |
| 331 | |
| 332 | is_backbone = layer_id < config.num_hidden_layers |
| 333 | self.is_kv_source = is_backbone and layer_id in self.kv_source_layers |
| 334 | self.is_index_source = is_backbone and layer_id in self.index_source_layers |
| 335 | if self.compress_ratio > 0: |
| 336 | if not self.kv_source_layers or not self.index_source_layers: |
| 337 | raise ValueError( |
| 338 | "DeepSeek V4.1 requires kv_source_layer_ids / " |
| 339 | "index_source_layer_ids in the config for compressed " |
| 340 | f"layers (layer {layer_id} has " |
| 341 | f"compress_ratio={self.compress_ratio})." |
| 342 | ) |
| 343 | self.kv_source_layer_id = max( |
| 344 | s for s in self.kv_source_layers if s <= layer_id |
| 345 | ) |
| 346 | self.index_source_layer_id = max( |
| 347 | s for s in self.index_source_layers if s <= layer_id |
| 348 | ) |
| 349 | else: |
| 350 | self.kv_source_layer_id = None |
| 351 | self.index_source_layer_id = None |
| 352 | self.eps = config.rms_norm_eps |
| 353 | self.scale = self.head_dim**-0.5 |
| 354 | |
| 355 | # Padded Q head count is dictated by the platform subclass. |
| 356 | self.padded_heads = self.get_padded_num_q_heads(self.n_local_heads) |
| 357 | # Sink padded to the same head count, initialized to -inf (no sink |
| 358 | # effect). Weight loading fills the first n_local_heads slots. |
| 359 | self.attn_sink = nn.Parameter( |
| 360 | torch.full((self.padded_heads,), -float("inf"), dtype=torch.float32), |
| 361 | requires_grad=False, |
| 362 | ) |
| 363 | |
| 364 | self.fused_wqa_wkv = MergedColumnParallelLinear( |
| 365 | self.hidden_size, |
| 366 | [self.q_lora_rank, self.head_dim], |
| 367 | bias=False, |
| 368 | quant_config=quant_config, |
| 369 | prefix=f"{prefix}.fused_wqa_wkv", |
| 370 | disable_tp=True, # fused ReplicatedLinear |
| 371 | ) |
| 372 | self.q_norm = RMSNorm(self.q_lora_rank, self.eps) |
| 373 | self.wq_b = ColumnParallelLinear( |
| 374 | self.q_lora_rank, |
| 375 | self.n_heads * self.head_dim, |
| 376 | bias=False, |
| 377 | quant_config=quant_config, |
| 378 | return_bias=False, |
| 379 | prefix=f"{prefix}.wq_b", |
| 380 | ) |
| 381 | |
| 382 | self.kv_norm = RMSNorm(self.head_dim, self.eps) |
| 383 | self.wo_a = ColumnParallelLinear( |
| 384 | self.n_heads * self.head_dim // self.n_groups, |
| 385 | self.n_groups * self.o_lora_rank, |
| 386 | bias=False, |
| 387 | quant_config=quant_config, |
| 388 | return_bias=False, |
| 389 | prefix=f"{prefix}.wo_a", |
| 390 | ) |
| 391 | self.wo_a.is_bmm = True |
| 392 | self.wo_a.bmm_batch_size = self.n_local_groups |
| 393 | self._o_proj_block_size = ( |
| 394 | 32 if getattr(self.wo_a, "weight_block_size", None) == [1, 32] else 128 |
| 395 | ) |
| 396 | self.wo_b = RowParallelLinear( |
| 397 | self.n_groups * self.o_lora_rank, |
| 398 | self.hidden_size, |
| 399 | bias=False, |
| 400 | quant_config=quant_config, |
| 401 | return_bias=False, |
| 402 | prefix=f"{prefix}.wo_b", |
| 403 | ) |
| 404 | # Set by ``bind_gemm_rs`` when the decoder layer runs sequence |
| 405 | # parallel and the fused GEMM + reduce-scatter kernel accepts wo_b. |
| 406 | self.gemm_rs: GemmRsAr | None = None |
| 407 | |
| 408 | # Initialize rotary embedding before the indexer/compressor consume it. |
| 409 | self.rotary_emb = build_deepseek_v4_rope( |
| 410 | config, |
| 411 | head_dim=self.head_dim, |
| 412 | rope_head_dim=self.rope_head_dim, |
| 413 | max_position_embeddings=config.max_position_embeddings, |
| 414 | compress_ratio=self.compress_ratio, |
| 415 | ) |
| 416 | self.indexer_rotary_emb = self.rotary_emb |
| 417 | self.topk_indices_buffer = topk_indices_buffer |
| 418 | self.candidate_block_buffer = candidate_block_buffer |
| 419 | |
| 420 | # Register with compilation context for metadata lookup. Done before |
| 421 | # indexer/compressor creation so consumers can resolve their source |
| 422 | # layers through it. |
| 423 | compilation_config = vllm_config.compilation_config |
| 424 | if prefix and prefix in compilation_config.static_forward_context: |
| 425 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 426 | if prefix: |
| 427 | compilation_config.static_forward_context[prefix] = self |
| 428 | self.kv_cache = torch.tensor([]) |
| 429 | self._static_forward_context = compilation_config.static_forward_context |
| 430 | |
| 431 | self.indexer = None |
| 432 | if self.is_index_source: |
| 433 | index_k_cache: DeepseekV4IndexerCache | None |
| 434 | # Index K cache: owned only by kv-source layers (their indexer has |
| 435 | # wk/k_norm); non-owning index sources share the K cache of the |
| 436 | # latest kv source below them (its indexer produces the keys). |
| 437 | if self.is_kv_source: |
| 438 | index_k_cache = DeepseekV4IndexerCache( |
| 439 | head_dim=_indexer_k_cache_head_dim( |
| 440 | config.index_head_dim, dsa_indexer_uses_fp4(vllm_config) |
| 441 | ), |
| 442 | dtype=torch.uint8, |
| 443 | prefix=f"{prefix}.indexer.k_cache", |
| 444 | cache_config=cache_config, |
| 445 | compress_ratio=self.compress_ratio, |
| 446 | ) |
| 447 | else: |
| 448 | assert self.kv_source_layer_id is not None |
| 449 | k_cache_prefix = ( |
| 450 | f"{_replace_layer_index(prefix, self.kv_source_layer_id)}" |
| 451 | ".indexer.k_cache" |
| 452 | ) |
| 453 | index_k_cache = self._static_forward_context.get(k_cache_prefix) |
| 454 | if index_k_cache is None: |
| 455 | raise NotImplementedError( |
| 456 | f"Indexer K cache source {k_cache_prefix} not found on " |
| 457 | "this rank; PP splits inside a v4.1 kv-sharing group " |
| 458 | "are not supported." |
| 459 | ) |
| 460 | is_candidate_source = layer_id == self.candidate_source_layer |
| 461 | uses_candidates = 0 <= self.candidate_source_layer < layer_id |
| 462 | self.indexer = self._indexer_cls(index_k_cache)( |
| 463 | vllm_config, |
| 464 | config=config, |
| 465 | hidden_size=self.hidden_size, |
| 466 | q_lora_rank=self.q_lora_rank, |
| 467 | quant_config=quant_config, |
| 468 | cache_config=cache_config, |
| 469 | topk_indices_buffer=topk_indices_buffer, |
| 470 | compress_ratio=self.compress_ratio, |
| 471 | prefix=f"{prefix}.indexer", |
| 472 | owns_k=self.is_kv_source, |
| 473 | k_cache=index_k_cache, |
| 474 | main_head_dim=self.head_dim, |
| 475 | candidate_block_buffer=( |
| 476 | candidate_block_buffer |
| 477 | if (is_candidate_source or uses_candidates) |
| 478 | else None |
| 479 | ), |
| 480 | candidate_block_size=self.candidate_block_size, |
| 481 | candidate_write=is_candidate_source, |
| 482 | ) |
| 483 | |
| 484 | self._prepare_and_attn_fn = self._prepare_and_attn |
| 485 | if not vllm_config.use_v2_model_runner: |
| 486 | # MRV1's piecewise capture only tolerates the wide eager region: with |
| 487 | # the narrow one the attention input preparation stays in the captured |
| 488 | # graph and MRV1 produces garbage (#51430). |
| 489 | self._prepare_and_attn_fn = self._prepare_and_attn_eager |
| 490 | |
| 491 | self.aux_stream_list = aux_stream_list |
| 492 | # [0]: GEMM start / post-GEMM event0. [1..3]: GEMM done events; |
| 493 | # [1] doubles as post-GEMM event1. Reuse is safe: GEMM fully joins |
| 494 | # before post-GEMM starts. |
| 495 | self.ln_events = [torch.cuda.Event() for _ in range(4)] |
| 496 | |
| 497 | assert cache_config is not None, "DeepseekV4 attention requires cache_config" |
| 498 | # ---- Attention / KV-cache setup ---- |
| 499 | self.max_num_batched_tokens = ( |
| 500 | vllm_config.scheduler_config.max_num_batched_tokens |
| 501 | ) |
| 502 | self.max_model_len = vllm_config.model_config.max_model_len |
| 503 | |
| 504 | # Resolve the kv-cache dtype from this backend's block format. The same |
| 505 | # resolution drives the SWA cache tensor dtype below. |
| 506 | self.kv_cache_dtype, self.kv_cache_torch_dtype = _resolve_dsv4_kv_cache_dtype( |
| 507 | self._uses_fp8_ds_mla_layout(), |
| 508 | cache_config.cache_dtype, |
| 509 | cache_config, |
| 510 | self.packed_kv_cache_dtype, |
| 511 | ) |
| 512 | self.kv_mxfp8 = _use_v41_mxfp8_kv_record() |
| 513 | self.swa_bytes_per_token = 528 if self.kv_mxfp8 else 584 |
| 514 | # nvfp4_ds_mla keeps the MXFP8 sliding-window record and stores the |
| 515 | # compressed cache as NVFP4 (256 B of e2m1 pairs + 32 e4m3 scales). |
| 516 | self.compressed_bytes_per_token = ( |
| 517 | 288 if self.kv_cache_dtype == "nvfp4_ds_mla" else self.swa_bytes_per_token |
| 518 | ) |
| 519 | # One alignment for every page in the block: the block stride is their |
| 520 | # sum, and 512 satisfies both TMA strides in play (512 for the V4.1 |
| 521 | # fp8 record, 256 for NVFP4). |
| 522 | self.kv_page_alignment = 512 if self.kv_mxfp8 else 576 |
| 523 | if self.kv_cache_dtype == "nvfp4_ds_mla" and not self.kv_mxfp8: |
| 524 | raise ValueError( |
| 525 | "nvfp4_ds_mla needs the V4.1 KV records, which FlashMLA " |
| 526 | "decodes only on SM100." |
| 527 | ) |
| 528 | |
| 529 | swa_bounded_replay = cache_config.swa_bounded_replay |
| 530 | if swa_bounded_replay and not vllm_config.use_v2_model_runner: |
| 531 | logger.warning_once( |
| 532 | "SWA bounded replay needs model runner V2 (only it skips the " |
| 533 | "paged-KV writes of replayed tokens); the sliding-window cache " |
| 534 | "takes part in prefix caching instead." |
| 535 | ) |
| 536 | swa_bounded_replay = False |
| 537 | if ( |
| 538 | swa_bounded_replay |
| 539 | and vllm_config.parallel_config.prefill_context_parallel_size > 1 |
| 540 | ): |
| 541 | logger.warning_once( |
| 542 | "SWA bounded replay is off under prefill context parallelism " |
| 543 | "(the replayed tokens' slot padding knows the rank-local batch " |
| 544 | "only); the sliding-window cache takes part in prefix caching " |
| 545 | "instead." |
| 546 | ) |
| 547 | swa_bounded_replay = False |
| 548 | if swa_bounded_replay and current_platform.is_rocm(): |
| 549 | logger.warning_once( |
| 550 | "SWA bounded replay is off on ROCm (the sparse SWA metadata " |
| 551 | "builders forward replay_start, but the window clamp it relies " |
| 552 | "on lives in the FlashInfer and FlashMLA prefill kernels, so " |
| 553 | "the padded slots fault); the sliding-window cache takes part " |
| 554 | "in prefix caching instead." |
| 555 | ) |
| 556 | swa_bounded_replay = False |
| 557 | self.swa_cache_layer = DeepseekV4SWACache( |
| 558 | head_dim=self.head_dim, |
| 559 | window_size=self.window_size, |
| 560 | dtype=self.kv_cache_torch_dtype, |
| 561 | prefix=f"{prefix}.swa_cache", |
| 562 | cache_config=cache_config, |
| 563 | backend_cls=self.swa_backend_cls, |
| 564 | block_size=32, |
| 565 | packed_bytes_per_token=self.swa_bytes_per_token, |
| 566 | packed_page_alignment=self.kv_page_alignment, |
| 567 | bounded_replay=swa_bounded_replay, |
| 568 | ) |
| 569 | |
| 570 | # The attention layer itself was already registered with the |
| 571 | # compilation context above (before indexer/compressor creation). |
| 572 | |
| 573 | # Compressors live only on kv-source layers; consumers read the |
| 574 | # source's compressed cache through the forward context. |
| 575 | self.compressor = None |
| 576 | if self.is_kv_source: |
| 577 | self.compressor = DeepseekCompressor( |
| 578 | vllm_config=vllm_config, |
| 579 | compress_ratio=self.compress_ratio, |
| 580 | hidden_size=self.hidden_size, |
| 581 | head_dim=self.head_dim, |
| 582 | rotate=True, |
| 583 | prefix=f"{prefix}.compressor", |
| 584 | k_cache_prefix=self.prefix, |
| 585 | ) |
| 586 | # Prefix of the attention layer owning this layer's compressed KV |
| 587 | # cache (self for kv sources). |
| 588 | if self.compress_ratio > 0: |
| 589 | assert self.kv_source_layer_id is not None |
| 590 | self.compressed_cache_prefix: str | None = _replace_layer_index( |
| 591 | prefix, self.kv_source_layer_id |
| 592 | ) |
| 593 | if ( |
| 594 | not self.is_kv_source |
| 595 | and self.compressed_cache_prefix not in self._static_forward_context |
| 596 | ): |
| 597 | raise NotImplementedError( |
| 598 | f"Compressed-KV source {self.compressed_cache_prefix} not " |
| 599 | "found on this rank; PP splits inside a v4.1 kv-sharing " |
| 600 | "group are not supported." |
| 601 | ) |
| 602 | else: |
| 603 | self.compressed_cache_prefix = None |
| 604 | |
| 605 | if vllm_config.kernel_config.enable_jit_warmup: |
| 606 | from vllm.v1.attention.backends.mla.sparse_swa import ( |
| 607 | _COMPUTE_PREFILL_METADATA_KERNEL, |
| 608 | _COMPUTE_SWA_INDICES_AND_LENS_KERNEL, |
| 609 | ) |
| 610 | |
| 611 | _COMPUTE_PREFILL_METADATA_KERNEL.register_warmup() |
| 612 | _COMPUTE_SWA_INDICES_AND_LENS_KERNEL.register_warmup( |
| 613 | window_size=self.window_size, |
| 614 | block_size=self.swa_cache_layer.block_size, |
| 615 | ) |
| 616 | |
| 617 | if self.compress_ratio > 1: |
| 618 | from vllm.v1.attention.backends.mla.compressor_utils import ( |
| 619 | _COMPRESSED_SLOT_MAPPING_KERNEL, |
| 620 | ) |
| 621 | |
| 622 | _COMPRESSED_SLOT_MAPPING_KERNEL.register_warmup() |
| 623 | |
| 624 | if self.indexer is not None: |
| 625 | from vllm.v1.attention.backends.mla.indexer import ( |
| 626 | _BUILD_PREFILL_CHUNK_METADATA_KERNEL, |
| 627 | _PREPARE_UNIFORM_DECODE_KERNEL, |
| 628 | ) |
| 629 | |
| 630 | _PREPARE_UNIFORM_DECODE_KERNEL.register_warmup() |
| 631 | _BUILD_PREFILL_CHUNK_METADATA_KERNEL.register_warmup() |
| 632 | |
| 633 | spec_config = vllm_config.speculative_config |
| 634 | if spec_config is not None and spec_config.use_dspark(): |
| 635 | from vllm.v1.attention.backends.mla.sparse_swa import ( |
| 636 | _COMPUTE_DSPARK_NONCAUSAL_SWA_INDICES_KERNEL, |
| 637 | ) |
| 638 | |
| 639 | _COMPUTE_DSPARK_NONCAUSAL_SWA_INDICES_KERNEL.register_warmup( |
| 640 | window_size=self.window_size, |
| 641 | num_speculative_tokens=spec_config.num_speculative_tokens, |
| 642 | block_size=self.swa_cache_layer.block_size, |
| 643 | ) |
| 644 | |
| 645 | # Every backend that gathers a chunk's KV through |
| 646 | # combine_topk_swa_indices needs its Triton kernel warmed, mega |
| 647 | # attention included -- it calls it from _forward_prefill_mega. |
| 648 | if self.backend_cls.get_name() in ( |
| 649 | "FLASHMLA_SPARSE_DSV41", |
| 650 | "FLASHMLA_MEGA_ATTN_DSV41", |
| 651 | "ROCM_FLASHMLA_SPARSE_DSV4", |
| 652 | ): |
| 653 | from vllm.models.deepseek_v41.common.ops.cache_utils import ( |
| 654 | _COMBINE_TOPK_SWA_INDICES_KERNEL, |
| 655 | ) |
| 656 | |
| 657 | _COMBINE_TOPK_SWA_INDICES_KERNEL.register_warmup() |
| 658 | |
| 659 | def forward( |
| 660 | self, |
| 661 | positions: torch.Tensor, |
| 662 | hidden_states: torch.Tensor, |
| 663 | llama_4_scaling: torch.Tensor | None = None, |
| 664 | ) -> torch.Tensor: |
| 665 | # The eager attention region writes into a caller-owned buffer |
| 666 | # (breakable_cudagraph needs in-place outputs); its shape and how it is |
| 667 | # projected afterwards follow the interface contract above. |
| 668 | attn_out = self._alloc_attn_out(hidden_states.shape[0], hidden_states) |
| 669 | |
| 670 | # Keep the attention input preparation in the captured graph. Only the |
| 671 | # sparse indexer and MLA attention run in the eager break below. |
| 672 | qr_kv, kv_score, indexer_weights = self._run_parallel_input_projections( |
| 673 | hidden_states |
| 674 | ) |
| 675 | qr, qr_scale, kv = self._split_qkv_and_norm(qr_kv) |
| 676 | |
| 677 | self._prepare_and_attn_fn( |
| 678 | hidden_states, |
| 679 | qr, |
| 680 | kv, |
| 681 | qr_scale, |
| 682 | kv_score, |
| 683 | indexer_weights, |
| 684 | positions, |
| 685 | attn_out, |
| 686 | ) |
| 687 | return self._o_proj(attn_out, positions) |
| 688 | |
| 689 | def bind_gemm_rs(self) -> None: |
| 690 | """Fuse ``wo_b`` with the sequence-parallel TP reduce-scatter. |
| 691 | |
| 692 | The decoder layer calls this after the model initialized the |
| 693 | process-wide GEMM-RS workspace and before weights load: the |
| 694 | eligibility check inspects the projection's linear kernel and weight |
| 695 | shape, which online quantization may later re-layout. |
| 696 | """ |
| 697 | from vllm.model_executor.kernels.linear.cute_dsl.gemm_rs_ar import ( |
| 698 | get_gemm_rs_ar, |
| 699 | ) |
| 700 | |
| 701 | # The layer owns the reduction under sequence parallel. |
| 702 | assert not self.wo_b.reduce_results |
| 703 | gemm_rs = get_gemm_rs_ar() |
| 704 | if gemm_rs.can_run(self.wo_b): |
| 705 | self.gemm_rs = gemm_rs |
| 706 | else: |
| 707 | gemm_rs.warn_incompatible_projection() |
| 708 | |
| 709 | def _wo_b_proj(self, z: torch.Tensor | QuantizedActivation) -> torch.Tensor: |
| 710 | """Apply ``wo_b``; with GEMM-RS bound, also reduce-scatter the result. |
| 711 | |
| 712 | Every ``_o_proj`` implementation projects through this so the decoder |
| 713 | layer sees one contract: when ``gemm_rs`` is bound the output is |
| 714 | already the local sequence-parallel shard, otherwise it is the |
| 715 | unreduced TP partial. |
| 716 | """ |
| 717 | if self.gemm_rs is None: |
| 718 | return self.wo_b(z) |
| 719 | if isinstance(z, torch.Tensor) and self.gemm_rs.should_run(z): |
| 720 | return self.gemm_rs.apply(z, self.wo_b) |
| 721 | # Small batches stay on the unfused path, which is faster there. |
| 722 | return sp_reduce_scatter(self.wo_b(z)) |
| 723 | |
| 724 | def _alloc_attn_out( |
| 725 | self, num_tokens: int, hidden_states: torch.Tensor |
| 726 | ) -> "torch.Tensor | QuantizedActivation": |
| 727 | """The buffer ``forward_mqa`` fills, per the interface contract.""" |
| 728 | return torch.empty( |
| 729 | (num_tokens, self.padded_heads, self.head_dim), |
| 730 | dtype=hidden_states.dtype, |
| 731 | device=hidden_states.device, |
| 732 | ) |
| 733 | |
| 734 | @cached_property |
| 735 | def _can_fuse_query_quant(self) -> bool: |
| 736 | from vllm.models.deepseek_v41.common.ops.query_quant import ( |
| 737 | can_fuse_query_quant, |
| 738 | ) |
| 739 | |
| 740 | linears = [self.wq_b] |
| 741 | if self.indexer is not None: |
| 742 | linears.append(self.indexer.wq_b) |
| 743 | return can_fuse_query_quant(linears) |
| 744 | |
| 745 | def _split_qkv_and_norm( |
| 746 | self, qr_kv: torch.Tensor |
| 747 | ) -> tuple[torch.Tensor | QuantizedActivation, torch.Tensor | None, torch.Tensor]: |
| 748 | """Split the fused q-lora / kv projection and RMSNorm both halves. |
| 749 | |
| 750 | Compatible MXFP8 projections share the quantized Q and scales; |
| 751 | other projection backends consume the normalized Q directly. |
| 752 | """ |
| 753 | qr, kv = qr_kv.split([self.q_lora_rank, self.head_dim], dim=-1) |
| 754 | if self.q_lora_rank % 32 == 0 and self._can_fuse_query_quant: |
| 755 | from vllm.models.deepseek_v41.common.ops.query_quant import ( |
| 756 | fused_q_kv_rmsnorm_quant, |
| 757 | ) |
| 758 | |
| 759 | qr_quant, kv = fused_q_kv_rmsnorm_quant( |
| 760 | qr, |
| 761 | kv, |
| 762 | self.q_norm.weight.data, |
| 763 | self.kv_norm.weight.data, |
| 764 | self.eps, |
| 765 | ) |
| 766 | return qr_quant, None, kv |
| 767 | qr, kv = fused_q_kv_rmsnorm( |
| 768 | qr, |
| 769 | kv, |
| 770 | self.q_norm.weight.data, |
| 771 | self.kv_norm.weight.data, |
| 772 | self.eps, |
| 773 | ) |
| 774 | return qr, None, kv |
| 775 | |
| 776 | @eager_break_during_capture |
| 777 | def _prepare_and_attn_eager( |
| 778 | self, |
| 779 | hidden_states: torch.Tensor, |
| 780 | qr: torch.Tensor | QuantizedActivation, |
| 781 | kv: torch.Tensor, |
| 782 | qr_scale: torch.Tensor | None, |
| 783 | kv_score: torch.Tensor, |
| 784 | indexer_weights: torch.Tensor, |
| 785 | positions: torch.Tensor, |
| 786 | attn_out: "torch.Tensor | QuantizedActivation", |
| 787 | ) -> None: |
| 788 | """Wide eager region: the whole of ``_prepare_and_attn`` runs eagerly. |
| 789 | |
| 790 | The nested ``_sparse_indexer_and_attn`` break runs inline, since |
| 791 | ``add_eager`` clears ``_capturing`` before invoking this. |
| 792 | """ |
| 793 | self._prepare_and_attn( |
| 794 | hidden_states, |
| 795 | qr, |
| 796 | kv, |
| 797 | qr_scale, |
| 798 | kv_score, |
| 799 | indexer_weights, |
| 800 | positions, |
| 801 | attn_out, |
| 802 | ) |
| 803 | |
| 804 | def _prepare_and_attn( |
| 805 | self, |
| 806 | hidden_states: torch.Tensor, |
| 807 | qr: torch.Tensor | QuantizedActivation, |
| 808 | kv: torch.Tensor, |
| 809 | qr_scale: torch.Tensor | None, |
| 810 | kv_score: torch.Tensor, |
| 811 | indexer_weights: torch.Tensor, |
| 812 | positions: torch.Tensor, |
| 813 | attn_out: "torch.Tensor | QuantizedActivation", |
| 814 | ) -> None: |
| 815 | """Attention input preparation followed by the sparse indexer and MLA. |
| 816 | |
| 817 | Only the latter runs in the eager break. |
| 818 | |
| 819 | Q/SWA preparation overlaps state saving and compression. Once the |
| 820 | latent is ready, main-cache insertion overlaps indexer preparation; |
| 821 | both cache writes finish before sparse attention reads either cache. |
| 822 | """ |
| 823 | attn_metadata = get_forward_context().attn_metadata |
| 824 | indexer = self.indexer |
| 825 | compressor = self.compressor |
| 826 | aux_streams = self.aux_stream_list |
| 827 | |
| 828 | def project_query_and_cache_kv() -> torch.Tensor: |
| 829 | q = self._wq_b_proj(qr, qr_scale).view( |
| 830 | -1, self.n_local_heads, self.head_dim |
| 831 | ) |
| 832 | return self._fused_qnorm_rope_kv_insert(q, kv, positions, attn_metadata) |
| 833 | |
| 834 | index_q: torch.Tensor | None = None |
| 835 | index_q_scale: torch.Tensor | None = None |
| 836 | index_weights_out: torch.Tensor | None = None |
| 837 | latent: torch.Tensor | None = None |
| 838 | aux_stream = aux_streams[0] if aux_streams is not None else None |
| 839 | |
| 840 | if compressor is not None: |
| 841 | # Q projection / KV insertion on the default stream overlaps the |
| 842 | # compressor on aux stream 0 (sequential on ROCm). |
| 843 | q, latent = maybe_execute_in_parallel( |
| 844 | project_query_and_cache_kv, |
| 845 | lambda: compressor(kv_score, positions), |
| 846 | self.ln_events[0], |
| 847 | self.ln_events[1], |
| 848 | aux_stream, |
| 849 | ) |
| 850 | else: |
| 851 | q = project_query_and_cache_kv() |
| 852 | |
| 853 | def prepare_indexer(): |
| 854 | if indexer is None: |
| 855 | return None, None, None |
| 856 | return indexer( |
| 857 | qr, |
| 858 | latent, |
| 859 | indexer_weights, |
| 860 | positions, |
| 861 | self.indexer_rotary_emb, |
| 862 | qr_scale, |
| 863 | ) |
| 864 | |
| 865 | if compressor is not None: |
| 866 | indexer_result, _ = maybe_execute_in_parallel( |
| 867 | prepare_indexer, |
| 868 | lambda: compressor.insert_cache(latent, positions, self.rotary_emb), |
| 869 | self.ln_events[0], |
| 870 | self.ln_events[1], |
| 871 | aux_stream, |
| 872 | ) |
| 873 | else: |
| 874 | indexer_result = prepare_indexer() |
| 875 | index_q, index_q_scale, index_weights_out = indexer_result |
| 876 | |
| 877 | self._sparse_indexer_and_attn( |
| 878 | hidden_states, |
| 879 | index_q, |
| 880 | index_q_scale, |
| 881 | index_weights_out, |
| 882 | q, |
| 883 | kv, |
| 884 | positions, |
| 885 | attn_out, |
| 886 | ) |
| 887 | |
| 888 | def _fused_wqa_wkv_gemm(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 889 | # Override point: the ROCm layer preshuffles this weight in place, so |
| 890 | # it cannot go through fused_wqa_wkv directly. |
| 891 | # MergedColumnParallelLinear returns (output, bias); bias is None. |
| 892 | qr_kv, _ = self.fused_wqa_wkv(hidden_states) |
| 893 | return qr_kv |
| 894 | |
| 895 | def _wq_b_proj( |
| 896 | self, |
| 897 | qr: torch.Tensor | QuantizedActivation, |
| 898 | qr_scale: torch.Tensor | None = None, |
| 899 | ) -> torch.Tensor: |
| 900 | """Project normalized Q, bypassing quantization when already fused.""" |
| 901 | assert qr_scale is None, "ROCm-only path" |
| 902 | return self.wq_b(qr) |
| 903 | |
| 904 | def _run_parallel_input_projections( |
| 905 | self, hidden_states: torch.Tensor |
| 906 | ) -> tuple[ |
| 907 | torch.Tensor, |
| 908 | torch.Tensor | None, |
| 909 | torch.Tensor | None, |
| 910 | ]: |
| 911 | aux_streams = self.aux_stream_list |
| 912 | if aux_streams is not None: |
| 913 | aux_streams = aux_streams[:2] |
| 914 | |
| 915 | # fused_wqa_wkv (heaviest) on default; the two lighter input GEMMs on |
| 916 | # aux streams 0/1 when their owning module exists. ln_events[0] is the |
| 917 | # fan-out start event; ln_events[1..2] are per-aux done events. The |
| 918 | # v4.1 indexer derives K from the kv-source compressor's latent, so |
| 919 | # unlike v4.0 there is no indexer K GEMM over hidden_states here. |
| 920 | aux_fns: list[Callable[[], Any] | None] = [None, None] |
| 921 | |
| 922 | if self.compressor is not None: |
| 923 | # Local ref so the closure keeps a non-None type for mypy. |
| 924 | compressor = self.compressor |
| 925 | |
| 926 | def compressor_kv_score() -> torch.Tensor: |
| 927 | return torch.mm( |
| 928 | hidden_states, |
| 929 | compressor.fused_wkv_wgate.weight.T, |
| 930 | out_dtype=torch.float32, |
| 931 | ) |
| 932 | |
| 933 | aux_fns[0] = compressor_kv_score |
| 934 | |
| 935 | if self.indexer is not None: |
| 936 | indexer = self.indexer |
| 937 | |
| 938 | def indexer_weights_proj() -> torch.Tensor: |
| 939 | # ReplicatedLinear returns (output, bias); bias is None. |
| 940 | weights, _ = indexer.weights_proj(hidden_states) |
| 941 | return weights |
| 942 | |
| 943 | aux_fns[1] = indexer_weights_proj |
| 944 | |
| 945 | qr_kv, (kv_score, indexer_weights) = execute_in_parallel( |
| 946 | lambda: self._fused_wqa_wkv_gemm(hidden_states), |
| 947 | aux_fns, |
| 948 | self.ln_events[0], |
| 949 | self.ln_events[1:3], |
| 950 | aux_streams, |
| 951 | enable=hidden_states.shape[0] |
| 952 | <= envs.VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD, |
| 953 | ) |
| 954 | |
| 955 | return qr_kv, kv_score, indexer_weights |
| 956 | |
| 957 | @eager_break_during_capture |
| 958 | def _sparse_indexer_and_attn( |
| 959 | self, |
| 960 | hidden_states: torch.Tensor, |
| 961 | index_q: torch.Tensor | None, |
| 962 | index_q_scale: torch.Tensor | None, |
| 963 | index_weights: torch.Tensor | None, |
| 964 | q: torch.Tensor, |
| 965 | kv: torch.Tensor, |
| 966 | positions: torch.Tensor, |
| 967 | out: torch.Tensor, |
| 968 | ) -> None: |
| 969 | if self.indexer is not None and index_q is not None: |
| 970 | assert index_weights is not None |
| 971 | q_quant = (index_q, index_q_scale) if index_q_scale is not None else index_q |
| 972 | self.indexer.indexer_op( |
| 973 | hidden_states, |
| 974 | q_quant, |
| 975 | None, |
| 976 | index_weights, |
| 977 | ) |
| 978 | |
| 979 | # MLA attention writes into the pre-allocated `out` buffer |
| 980 | # ([num_tokens, padded_heads, head_dim]). |
| 981 | self.forward_mqa(q, kv, positions, out) |
| 982 | |
| 983 | def _fused_qnorm_rope_kv_insert( |
| 984 | self, |
| 985 | q: torch.Tensor, |
| 986 | kv: torch.Tensor, |
| 987 | positions: torch.Tensor, |
| 988 | attn_metadata: ( |
| 989 | dict[str, AttentionMetadata] | list[dict[str, AttentionMetadata]] | None |
| 990 | ), |
| 991 | ) -> torch.Tensor: |
| 992 | """Ready ``q`` for the attention kernel and publish this step's KV. |
| 993 | |
| 994 | One launch does both. With ``accepts_unnormed_unroped_query`` the |
| 995 | attention kernel norms and rotates Q itself and reads it in its own |
| 996 | chunk-interleaved layout, so the Q half of the launch is a zero-pad to |
| 997 | ``padded_heads`` -- and nothing at all once the shard is that wide. |
| 998 | """ |
| 999 | if not isinstance(attn_metadata, dict): |
| 1000 | # Profile run: kernel doesn't fire; produce a padded tensor so |
| 1001 | # downstream FlashMLA gets the right shape. |
| 1002 | if self.n_local_heads >= self.padded_heads: |
| 1003 | return q |
| 1004 | if self.accepts_unnormed_unroped_query: |
| 1005 | # Padding heads sit at the tail of every head-dim chunk in |
| 1006 | # that layout, so no head-major pad of `q` reproduces it -- |
| 1007 | # and nothing reads it on a profile run. |
| 1008 | return q.new_zeros((q.shape[0], self.padded_heads, q.shape[2])) |
| 1009 | return F.pad( |
| 1010 | q, |
| 1011 | (0, 0, 0, self.padded_heads - self.n_local_heads), |
| 1012 | value=0.0, |
| 1013 | ) |
| 1014 | |
| 1015 | swa_metadata = cast( |
| 1016 | "DeepseekSparseSWAMetadata | None", |
| 1017 | attn_metadata.get(self.swa_cache_layer.prefix), |
| 1018 | ) |
| 1019 | assert swa_metadata is not None |
| 1020 | |
| 1021 | swa_kv_cache = self.swa_cache_layer.kv_cache |
| 1022 | # The fused insert ops require int64 position_ids; the runner's positions |
| 1023 | # buffer is already int64, so no cast is needed. |
| 1024 | assert positions.dtype == torch.int64 |
| 1025 | cos_sin_cache = self.rotary_emb.cos_sin_cache |
| 1026 | cache_dtype = swa_kv_cache.dtype |
| 1027 | |
| 1028 | # kv is unchanged; attention reads kv solely via swa_kv_cache. |
| 1029 | if cache_dtype == torch.uint8: |
| 1030 | # fp8_ds_mla UE8M0 paged path. Horizontally fused: |
| 1031 | # Q side: GPT-J RoPE, zero-filling the padding head slots; the |
| 1032 | # kernel allocates and returns the padded q tensor. An |
| 1033 | # interleaved Q skips the RoPE its attention kernel owns |
| 1034 | # and keeps only the pad, which q_head_padded=0 drops |
| 1035 | # too once the shard is already padded_heads wide. |
| 1036 | # KV side: GPT-J RoPE + UE8M0 FP8 quant + paged cache insert. |
| 1037 | swa_kv_cache_2d = swa_kv_cache.view(swa_kv_cache.shape[0], -1) |
| 1038 | pad_to = ( |
| 1039 | 0 |
| 1040 | if self.accepts_unnormed_unroped_query |
| 1041 | and self.n_local_heads == self.padded_heads |
| 1042 | else self.padded_heads |
| 1043 | ) |
| 1044 | q_padded = torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert( |
| 1045 | q, |
| 1046 | kv, |
| 1047 | swa_kv_cache_2d, |
| 1048 | swa_metadata.slot_mapping, |
| 1049 | positions, |
| 1050 | cos_sin_cache, |
| 1051 | pad_to, |
| 1052 | self.eps, |
| 1053 | swa_metadata.block_size, |
| 1054 | False, # apply_q_norm: qr is normed before wq_b |
| 1055 | self.kv_mxfp8, |
| 1056 | not self.accepts_unnormed_unroped_query, # apply_q_rope |
| 1057 | self.accepts_unnormed_unroped_query, # is_q_interleaved |
| 1058 | ) |
| 1059 | return q if pad_to == 0 else q_padded |
| 1060 | |
| 1061 | assert not self.accepts_unnormed_unroped_query, ( |
| 1062 | "the chunk-interleaved Q layout only pairs with a packed KV record" |
| 1063 | ) |
| 1064 | # Plain-row path: the [num_blocks, block_size, 512] cache stores the KV |
| 1065 | # row in its element dtype (no Q padding). bf16 rewrites q in place; |
| 1066 | # per-tensor fp8 writes a separately-allocated fp8 q and quantizes the |
| 1067 | # KV row. |
| 1068 | block_size = swa_metadata.block_size |
| 1069 | assert swa_kv_cache.shape[1:] == (block_size, self.head_dim) |
| 1070 | swa_kv_cache_3d = swa_kv_cache |
| 1071 | if cache_dtype == torch.bfloat16: |
| 1072 | torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert( |
| 1073 | q, |
| 1074 | kv, |
| 1075 | swa_kv_cache_3d, |
| 1076 | swa_metadata.slot_mapping, |
| 1077 | positions, |
| 1078 | cos_sin_cache, |
| 1079 | self.eps, |
| 1080 | block_size, |
| 1081 | False, |
| 1082 | ) |
| 1083 | return q |
| 1084 | |
| 1085 | # per-tensor fp8 (torch.float8_e4m3fn) |
| 1086 | q_fp8 = torch.empty_like(q, dtype=torch.float8_e4m3fn) |
| 1087 | torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert( |
| 1088 | q, |
| 1089 | kv, |
| 1090 | q_fp8, |
| 1091 | swa_kv_cache_3d, |
| 1092 | swa_metadata.slot_mapping, |
| 1093 | positions, |
| 1094 | cos_sin_cache, |
| 1095 | self._flashinfer_fp8_kv_scale, |
| 1096 | self._flashinfer_fp8_q_scale_inv, |
| 1097 | self.eps, |
| 1098 | block_size, |
| 1099 | False, |
| 1100 | ) |
| 1101 | return q_fp8 |
| 1102 | |
| 1103 | def bind_kv_cache(self, kv_cache: torch.Tensor) -> None: |
| 1104 | # [B, H=1, N, C] -> [B, N, C] |
| 1105 | self.kv_cache = kv_cache.squeeze(1) |
| 1106 | |
| 1107 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 1108 | return self.backend_cls |
| 1109 | |
| 1110 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None: |
| 1111 | # Only kv-source layers own a compressed-KV cache; consumers read the |
| 1112 | # source's cache through the forward context, and cr==0 layers are |
| 1113 | # pure SWA. The SWA cache is allocated separately as |
| 1114 | # DeepseekV4SWACache. |
| 1115 | if not self.is_kv_source: |
| 1116 | return None |
| 1117 | # fp8_ds_mla is a UE8M0 block-scaled uint8 layout whose page rounds up |
| 1118 | # to the decode kernel's TMA stride; plain bf16 / per-tensor fp8 rows |
| 1119 | # use natural element-size pages. |
| 1120 | uses_fp8_ds_mla_layout = self.kv_cache_dtype in ("fp8_ds_mla", "nvfp4_ds_mla") |
| 1121 | return MLAAttentionSpec( |
| 1122 | block_size=vllm_config.cache_config.block_size, |
| 1123 | num_kv_heads=1, |
| 1124 | head_size=self.head_dim, |
| 1125 | dtype=torch.uint8 if uses_fp8_ds_mla_layout else self.kv_cache_torch_dtype, |
| 1126 | tokens_per_state=self.compress_ratio, |
| 1127 | cache_dtype_str=self.kv_cache_dtype, |
| 1128 | alignment=self.kv_page_alignment if uses_fp8_ds_mla_layout else 512, |
| 1129 | model_version="deepseek_v4", |
| 1130 | kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype), |
| 1131 | # Packed record width; head_size stays semantic (512). |
| 1132 | state_content_bytes=( |
| 1133 | self.compressed_bytes_per_token if uses_fp8_ds_mla_layout else None |
| 1134 | ), |
| 1135 | ) |
| 1136 | |
| 1137 | def _compressed_kv_cache(self) -> torch.Tensor: |
| 1138 | """The compressed-KV cache tensor of this layer's kv source (own |
| 1139 | cache for kv-source layers).""" |
| 1140 | if self.is_kv_source: |
| 1141 | return self.kv_cache |
| 1142 | assert self.compressed_cache_prefix is not None |
| 1143 | source = self._static_forward_context[self.compressed_cache_prefix] |
| 1144 | return source.kv_cache |
| 1145 | |
| 1146 | |
| 1147 | class DeepseekV4IndexerCache(torch.nn.Module, AttentionLayerBase): |
| 1148 | def __init__( |
| 1149 | self, |
| 1150 | head_dim: int, |
| 1151 | dtype: torch.dtype, |
| 1152 | prefix: str, |
| 1153 | cache_config: CacheConfig, |
| 1154 | compress_ratio: int = 1, |
| 1155 | ): |
| 1156 | super().__init__() |
| 1157 | self.kv_cache = torch.tensor([]) |
| 1158 | self.head_dim = head_dim |
| 1159 | self.prefix = prefix |
| 1160 | self.cache_config = cache_config |
| 1161 | self.dtype = dtype |
| 1162 | self.compress_ratio = compress_ratio |
| 1163 | vllm_config = get_current_vllm_config() |
| 1164 | self.sparse_logits = vllm_config.attention_config.indexer_sparse_logits |
| 1165 | # aiter's paged MXFP4 kernels write this cache and read it in place. |
| 1166 | self.rocm_mxfp4 = current_platform.is_rocm() and dsa_indexer_uses_fp4( |
| 1167 | vllm_config |
| 1168 | ) |
| 1169 | compilation_config = vllm_config.compilation_config |
| 1170 | if prefix in compilation_config.static_forward_context: |
| 1171 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 1172 | compilation_config.static_forward_context[prefix] = self |
| 1173 | |
| 1174 | def bind_kv_cache(self, kv_cache: torch.Tensor) -> None: |
| 1175 | # [B, H=1, N, C] -> [B, N, C] |
| 1176 | self.kv_cache = kv_cache.squeeze(1) |
| 1177 | |
| 1178 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec: |
| 1179 | # head_dim already carries the fp8 scale padding |
| 1180 | # tokens_per_state=1 for V3.2, >1 for DeepseekV4; same cache layout. |
| 1181 | # nvfp4_ds_mla is packed too: its compressed record is NVFP4 but the |
| 1182 | # sliding-window record stays the V4.1 MXFP8 one, so the indexer page |
| 1183 | # takes the same alignment either way. |
| 1184 | uses_fp8_ds_mla_layout = vllm_config.cache_config.cache_dtype in ( |
| 1185 | "fp8_ds_mla", |
| 1186 | "nvfp4_ds_mla", |
| 1187 | ) |
| 1188 | page_alignment = ( |
| 1189 | 576 if uses_fp8_ds_mla_layout and not _use_v41_mxfp8_kv_record() else 512 |
| 1190 | ) |
| 1191 | return MLAAttentionSpec( |
| 1192 | block_size=self.cache_config.block_size, |
| 1193 | num_kv_heads=1, |
| 1194 | head_size=self.head_dim, |
| 1195 | dtype=self.dtype, |
| 1196 | tokens_per_state=self.compress_ratio, |
| 1197 | # The indexer page shares a physical block with the main KV cache |
| 1198 | # (the block stride is the sum of every page in the group), so it |
| 1199 | # has to carry the same padding: the packed record's TMA stride, |
| 1200 | # else 512B for FlashInfer sparse (#44577). |
| 1201 | alignment=page_alignment, |
| 1202 | # DeepGEMM's paged sparse MQA logits address pages by the block |
| 1203 | # stride, which packs every layer's page in this layout, and need |
| 1204 | # it 512B-aligned; the main KV record read from the same blocks |
| 1205 | # needs its own TMA stride, so keep both. |
| 1206 | block_stride_alignment=( |
| 1207 | math.lcm(512, page_alignment) |
| 1208 | if self.sparse_logits and not self.rocm_mxfp4 |
| 1209 | else None |
| 1210 | ), |
| 1211 | ) |
| 1212 | |
| 1213 | def forward(self): ... |
| 1214 | |
| 1215 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 1216 | if self.rocm_mxfp4: |
| 1217 | from vllm.v1.attention.backends.mla.rocm_paged_mxfp4_indexer import ( |
| 1218 | DeepseekV41RocmMxfp4IndexerBackend, |
| 1219 | ) |
| 1220 | |
| 1221 | return DeepseekV41RocmMxfp4IndexerBackend |
| 1222 | if self.sparse_logits: |
| 1223 | return DeepseekV41SparseIndexerBackend |
| 1224 | return DeepseekV41IndexerBackend |
| 1225 | |
| 1226 | |
| 1227 | class DeepseekV4Indexer(nn.Module): |
| 1228 | """DeepSeek V4.1 sparse-attention indexer. |
| 1229 | |
| 1230 | Exists only on ``index_source_layer_ids``; consumers reuse the topk |
| 1231 | indices it publishes into the shared ``topk_indices_buffer``. Unlike v4.0 |
| 1232 | the index key is derived from the kv-source layer's compressor latent |
| 1233 | (``k = k_norm(wk(latent))``, ``owns_k``) instead of an indexer-local |
| 1234 | compressor over hidden states, so there is no hidden-state K GEMM here. |
| 1235 | Non-owning index sources share the kv source's paged K cache. |
| 1236 | |
| 1237 | Two-level candidate filtering: the indexer at ``candidate_source_layer_id`` |
| 1238 | additionally publishes the top ``candidate_topk_blocks`` blocks of |
| 1239 | ``candidate_block_size`` compressed positions (``candidate_write``); |
| 1240 | later indexers mask their scores to those blocks before their own top-k. |
| 1241 | """ |
| 1242 | |
| 1243 | # The scoring layers; a platform subclass swaps in its own. |
| 1244 | mqa_cls: ClassVar[type["SparseMQAIndexer | RocmSparseMQAIndexer"]] = ( |
| 1245 | SparseMQAIndexer |
| 1246 | ) |
| 1247 | attn_cls: ClassVar[type[SparseAttnIndexer]] = SparseAttnIndexer |
| 1248 | |
| 1249 | def __init__( |
| 1250 | self, |
| 1251 | vllm_config: VllmConfig, |
| 1252 | config: DeepseekV2Config | DeepseekV3Config, |
| 1253 | hidden_size: int, |
| 1254 | q_lora_rank: int, |
| 1255 | quant_config: QuantizationConfig | None, |
| 1256 | cache_config: CacheConfig | None, |
| 1257 | topk_indices_buffer: torch.Tensor | None, |
| 1258 | compress_ratio: int, |
| 1259 | prefix: str, |
| 1260 | *, |
| 1261 | owns_k: bool, |
| 1262 | k_cache: DeepseekV4IndexerCache, |
| 1263 | main_head_dim: int, |
| 1264 | candidate_block_buffer: torch.Tensor | None = None, |
| 1265 | candidate_block_size: int = 0, |
| 1266 | candidate_write: bool = False, |
| 1267 | ): |
| 1268 | super().__init__() |
| 1269 | self.vllm_config = vllm_config |
| 1270 | self.config = config |
| 1271 | self.quant_config = quant_config |
| 1272 | self.topk_tokens = config.index_topk |
| 1273 | self.n_head = config.index_n_heads # 32 |
| 1274 | self.head_dim = config.index_head_dim # 128 |
| 1275 | self.rope_dim = config.qk_rope_head_dim # 64 |
| 1276 | self.q_lora_rank = q_lora_rank # 1280 |
| 1277 | self.compress_ratio = compress_ratio |
| 1278 | self.owns_k = owns_k |
| 1279 | self.use_fp4_kv = dsa_indexer_uses_fp4(vllm_config) |
| 1280 | logger.info_once( |
| 1281 | "Using %s indexer cache for Lightning Indexer.", |
| 1282 | "MXFP4" if self.use_fp4_kv else "FP8", |
| 1283 | ) |
| 1284 | |
| 1285 | # no tensor parallel, just replicated |
| 1286 | self.wq_b = ReplicatedLinear( |
| 1287 | self.q_lora_rank, |
| 1288 | self.head_dim * self.n_head, |
| 1289 | bias=False, |
| 1290 | quant_config=quant_config, |
| 1291 | prefix=f"{prefix}.wq_b", |
| 1292 | ) |
| 1293 | self.weights_proj = ReplicatedLinear( |
| 1294 | hidden_size, |
| 1295 | self.n_head, |
| 1296 | bias=False, |
| 1297 | quant_config=None, |
| 1298 | prefix=f"{prefix}.weights_proj", |
| 1299 | ) |
| 1300 | self.softmax_scale = self.head_dim**-0.5 |
| 1301 | |
| 1302 | self.scale_fmt = "ue8m0" |
| 1303 | self.quant_block_size = 128 # TODO: get from config |
| 1304 | self.topk_indices_buffer = topk_indices_buffer |
| 1305 | |
| 1306 | self.max_model_len = ( |
| 1307 | vllm_config.model_config.max_model_len // self.compress_ratio |
| 1308 | ) |
| 1309 | self.prefix = prefix |
| 1310 | |
| 1311 | self.max_total_seq_len = ( |
| 1312 | get_max_prefill_buffer_size(vllm_config) // self.compress_ratio |
| 1313 | ) |
| 1314 | |
| 1315 | assert cache_config is not None, "Deepseek V4 indexer requires cache_config" |
| 1316 | if owns_k: |
| 1317 | # wk maps the main compressor's (pre-RoPE) latent to index keys. |
| 1318 | # The checkpoint stores it in bf16 with no quantization scales. |
| 1319 | self.wk = ReplicatedLinear( |
| 1320 | main_head_dim, |
| 1321 | self.head_dim, |
| 1322 | bias=False, |
| 1323 | quant_config=None, |
| 1324 | prefix=f"{prefix}.wk", |
| 1325 | ) |
| 1326 | self.k_norm = RMSNorm(self.head_dim, config.rms_norm_eps) |
| 1327 | self.k_cache = k_cache |
| 1328 | |
| 1329 | # Candidate consumers can score only the candidate blocks (opt-in); |
| 1330 | # the candidate source and pre-candidate indexers stay dense. |
| 1331 | use_sparse_logits = ( |
| 1332 | vllm_config.attention_config.indexer_sparse_logits |
| 1333 | and candidate_block_buffer is not None |
| 1334 | and not candidate_write |
| 1335 | ) |
| 1336 | self.indexer_op: SparseAttnIndexer | SparseMQAIndexer | RocmSparseMQAIndexer |
| 1337 | if use_sparse_logits: |
| 1338 | if not self.use_fp4_kv: |
| 1339 | raise ValueError( |
| 1340 | "attention_config.indexer_sparse_logits requires the MXFP4 " |
| 1341 | "indexer cache (attention_config.indexer_kv_dtype='mxfp4')." |
| 1342 | ) |
| 1343 | assert candidate_block_buffer is not None |
| 1344 | assert topk_indices_buffer is not None |
| 1345 | self.indexer_op = self.mqa_cls( |
| 1346 | self.k_cache, |
| 1347 | self.topk_tokens, |
| 1348 | self.head_dim, |
| 1349 | self.max_total_seq_len, |
| 1350 | topk_indices_buffer, |
| 1351 | candidate_block_buffer, |
| 1352 | candidate_block_size, |
| 1353 | ) |
| 1354 | else: |
| 1355 | self.indexer_op = self.attn_cls( |
| 1356 | self.k_cache, |
| 1357 | self.quant_block_size, |
| 1358 | self.scale_fmt, |
| 1359 | self.topk_tokens, |
| 1360 | self.head_dim, |
| 1361 | self.max_model_len, |
| 1362 | self.max_total_seq_len, |
| 1363 | self.topk_indices_buffer, |
| 1364 | skip_k_cache_insert=True, |
| 1365 | use_fp4_cache=self.use_fp4_kv, |
| 1366 | compress_ratio=self.compress_ratio, |
| 1367 | candidate_blocks=candidate_block_buffer, |
| 1368 | candidate_block_size=candidate_block_size, |
| 1369 | candidate_write=candidate_write, |
| 1370 | ) |
| 1371 | # The fused Q kernel writes the per-head weights in the dtype the |
| 1372 | # scoring kernels take, so no cast runs per step. |
| 1373 | self.indexer_weights_dtype = ( |
| 1374 | self.mqa_cls.weights_dtype if use_sparse_logits else torch.float32 |
| 1375 | ) |
| 1376 | |
| 1377 | def _produce_k( |
| 1378 | self, |
| 1379 | latent: torch.Tensor | None, |
| 1380 | positions: torch.Tensor, |
| 1381 | rotary_emb: nn.Module, |
| 1382 | ) -> None: |
| 1383 | """Turn the compressor's pre-RoPE latent into paged index keys. |
| 1384 | |
| 1385 | ``k_norm(wk(latent))`` at group-boundary tokens, RoPE'd at the group |
| 1386 | position and MXFP4/FP8-quantized into the indexer K cache. |
| 1387 | """ |
| 1388 | attn_metadata = get_forward_context().attn_metadata |
| 1389 | if not isinstance(attn_metadata, dict) or latent is None: |
| 1390 | # Profile run: the indexer K cache is not bound yet and the |
| 1391 | # compressor skipped its latent output. |
| 1392 | return |
| 1393 | assert self.owns_k |
| 1394 | indexer_metadata = cast(Any, attn_metadata[self.k_cache.prefix]) |
| 1395 | # ReplicatedLinear returns (output, bias); bias is None. Rows at |
| 1396 | # non-boundary tokens hold garbage latent and are skipped by the |
| 1397 | # store kernel. |
| 1398 | k_pre, _ = self.wk(latent) |
| 1399 | indexer_k_norm_rope_store( |
| 1400 | k_pre, |
| 1401 | positions, |
| 1402 | rotary_emb.cos_sin_cache, |
| 1403 | self.k_norm.weight, |
| 1404 | self.k_norm.variance_epsilon, |
| 1405 | self.k_cache.kv_cache, |
| 1406 | indexer_metadata.slot_mapping, |
| 1407 | self.compress_ratio, |
| 1408 | self.use_fp4_kv, |
| 1409 | ) |
| 1410 | |
| 1411 | def forward_q( |
| 1412 | self, |
| 1413 | qr: torch.Tensor | QuantizedActivation, |
| 1414 | qr_scale: torch.Tensor | None, |
| 1415 | indexer_weights: torch.Tensor, |
| 1416 | positions: torch.Tensor, |
| 1417 | rotary_emb: nn.Module, |
| 1418 | ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor | None]: |
| 1419 | """Build the indexer queries: wq_b over qr plus fused RoPE/quant. |
| 1420 | |
| 1421 | Split out so the ROCm layer can schedule it apart from |
| 1422 | ``_produce_k`` (e.g. on an aux stream once ``qr`` is ready). |
| 1423 | """ |
| 1424 | q = self._wq_b_proj(qr, qr_scale) |
| 1425 | q = q.view(-1, self.n_head, self.head_dim) |
| 1426 | q_quant, weights = fused_indexer_q_rope_quant( |
| 1427 | positions, |
| 1428 | q, |
| 1429 | rotary_emb.cos_sin_cache, |
| 1430 | indexer_weights, |
| 1431 | self.softmax_scale, |
| 1432 | self.n_head**-0.5, |
| 1433 | use_fp4=self.use_fp4_kv, |
| 1434 | weights_out_dtype=self.indexer_weights_dtype, |
| 1435 | ) |
| 1436 | if isinstance(q_quant, tuple): |
| 1437 | q, q_scale = q_quant |
| 1438 | else: |
| 1439 | q, q_scale = q_quant, None |
| 1440 | return q, q_scale, weights |
| 1441 | |
| 1442 | def forward( |
| 1443 | self, |
| 1444 | qr: torch.Tensor | QuantizedActivation, |
| 1445 | latent: torch.Tensor | None, |
| 1446 | indexer_weights: torch.Tensor, |
| 1447 | positions: torch.Tensor, |
| 1448 | rotary_emb: nn.Module, |
| 1449 | qr_scale: torch.Tensor | None = None, |
| 1450 | ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor | None]: |
| 1451 | attn_metadata = get_forward_context().attn_metadata |
| 1452 | if isinstance(attn_metadata, dict): |
| 1453 | indexer_metadata = cast(Any, attn_metadata[self.k_cache.prefix]) |
| 1454 | if ( |
| 1455 | indexer_metadata.max_seq_len // self.compress_ratio <= self.topk_tokens |
| 1456 | and not torch.cuda.is_current_stream_capturing() |
| 1457 | ): |
| 1458 | # candidates num smaller than topk, every candidate is selected |
| 1459 | # but we still need to build k cache |
| 1460 | if self.owns_k: |
| 1461 | self._produce_k(latent, positions, rotary_emb) |
| 1462 | assert self.topk_indices_buffer is not None |
| 1463 | num_tokens = ( |
| 1464 | indexer_metadata.num_decode_tokens |
| 1465 | + indexer_metadata.num_prefill_tokens |
| 1466 | ) |
| 1467 | if num_tokens > 0: |
| 1468 | _fill_short_context_topk_indices[(num_tokens,)]( |
| 1469 | self.topk_indices_buffer, |
| 1470 | positions, |
| 1471 | TOP_K=self.topk_tokens, |
| 1472 | COMPRESS_RATIO=self.compress_ratio, |
| 1473 | PADDED_TOP_K=triton.next_power_of_2(self.topk_tokens), |
| 1474 | num_warps=8, |
| 1475 | ) |
| 1476 | return None, None, None |
| 1477 | |
| 1478 | if self.owns_k: |
| 1479 | # K write must land before indexer_op reads the cache |
| 1480 | # (skip_k_cache_insert=True). |
| 1481 | self._produce_k(latent, positions, rotary_emb) |
| 1482 | |
| 1483 | return self.forward_q(qr, qr_scale, indexer_weights, positions, rotary_emb) |
| 1484 | |
| 1485 | def _wq_b_proj( |
| 1486 | self, |
| 1487 | qr: torch.Tensor | QuantizedActivation, |
| 1488 | qr_scale: torch.Tensor | None = None, |
| 1489 | ) -> torch.Tensor: |
| 1490 | """Project normalized Q with the main attention's shared quantization.""" |
| 1491 | assert qr_scale is None, "ROCm-only path" |
| 1492 | # ReplicatedLinear returns (output, bias); bias is None. |
| 1493 | q, _ = self.wq_b(qr) |
| 1494 | return q |