| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """NVIDIA QSA owner with Triton kernels.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | from typing import ClassVar, cast |
| 8 | |
| 9 | import torch |
| 10 | from torch import nn |
| 11 | from transformers import Qwen4ExpTextConfig |
| 12 | |
| 13 | from vllm.compilation.breakable_cudagraph import eager_break_during_capture |
| 14 | from vllm.config import VllmConfig |
| 15 | from vllm.config.cache import CacheDType |
| 16 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 17 | from vllm.forward_context import get_forward_context |
| 18 | from vllm.model_executor.layers.attention.attention import ( |
| 19 | set_default_quant_scales, |
| 20 | ) |
| 21 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 22 | from vllm.model_executor.layers.layernorm import GemmaRMSNorm |
| 23 | from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear |
| 24 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 25 | from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding, get_rope |
| 26 | from vllm.model_executor.models.qwen3_next import Qwen3NextAttention |
| 27 | from vllm.platforms import current_platform |
| 28 | from vllm.platforms.interface import DeviceCapability |
| 29 | from vllm.utils.torch_utils import ( |
| 30 | kv_cache_dtype_str_to_dtype, |
| 31 | ) |
| 32 | from vllm.v1.attention.backend import ( |
| 33 | AttentionBackend, |
| 34 | AttentionCGSupport, |
| 35 | AttentionType, |
| 36 | MultipleOf, |
| 37 | ) |
| 38 | from vllm.v1.attention.backends.fa_utils import is_flash_attn_varlen_func_available |
| 39 | from vllm.v1.attention.backends.flash_attn import ( |
| 40 | FlashAttentionBackend, |
| 41 | FlashAttentionImpl, |
| 42 | FlashAttentionMetadata, |
| 43 | FlashAttentionMetadataBuilder, |
| 44 | ) |
| 45 | from vllm.v1.kv_cache_interface import ( |
| 46 | FullAttentionSpec, |
| 47 | KVCacheSpec, |
| 48 | get_kv_quant_mode, |
| 49 | ) |
| 50 | |
| 51 | from ..common.qsa_cache import QSAForwardMetadata |
| 52 | from . import model |
| 53 | from .indexer_qsa import QSAIndexer |
| 54 | |
| 55 | |
| 56 | class Qwen4ExpQSAMetadataBuilder(FlashAttentionMetadataBuilder): |
| 57 | """Flash metadata supporting uniform decode and target-verify graphs.""" |
| 58 | |
| 59 | _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH |
| 60 | |
| 61 | |
| 62 | class Qwen4ExpQSAFlashAttentionBackend(FlashAttentionBackend): |
| 63 | """FullAttentionSpec backend used by the merged QSA owner.""" |
| 64 | |
| 65 | supported_dtypes: ClassVar[list[torch.dtype]] = [torch.bfloat16] |
| 66 | # fp8/fp8_e4m3: e4m3 bytes in a uint8 cache, written by reshape_and_cache |
| 67 | # with the layer's per-tensor scales and dequantized on load inside the QSA |
| 68 | # Triton kernel. flash-attn never runs over this cache, so its fp8 probe |
| 69 | # does not apply (see supports_kv_cache_dtype and the impl constructor). |
| 70 | supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [ |
| 71 | "auto", |
| 72 | "bfloat16", |
| 73 | "fp8", |
| 74 | "fp8_e4m3", |
| 75 | ] |
| 76 | |
| 77 | @classmethod |
| 78 | def supports_kv_cache_dtype(cls, kv_cache_dtype: CacheDType | None) -> bool: |
| 79 | return kv_cache_dtype is None or kv_cache_dtype in cls.supported_kv_cache_dtypes |
| 80 | |
| 81 | @classmethod |
| 82 | def supports_combination( |
| 83 | cls, |
| 84 | head_size: int, |
| 85 | dtype: torch.dtype, |
| 86 | kv_cache_dtype: CacheDType | None, |
| 87 | block_size: int | None, |
| 88 | use_mla: bool, |
| 89 | has_sink: bool, |
| 90 | use_sparse: bool, |
| 91 | use_mm_prefix: bool, |
| 92 | device_capability: DeviceCapability, |
| 93 | ) -> str | None: |
| 94 | # QSA dequantizes the fp8 KV in its own Triton kernel and never runs |
| 95 | # flash-attn over the quantized cache, so the parent's fp8-KV rejection |
| 96 | # does not apply and every combination it is handed is accepted here. |
| 97 | return None |
| 98 | |
| 99 | @staticmethod |
| 100 | def get_name() -> str: |
| 101 | return "QWEN4_EXP_QSA_TRITON" |
| 102 | |
| 103 | @staticmethod |
| 104 | def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]: |
| 105 | # QSA consumes manager pages directly and does not use FA4 paged attention. |
| 106 | return [MultipleOf(16)] |
| 107 | |
| 108 | @staticmethod |
| 109 | def get_impl_cls() -> type[Qwen4ExpQSAFlashAttentionImpl]: |
| 110 | return Qwen4ExpQSAFlashAttentionImpl |
| 111 | |
| 112 | @staticmethod |
| 113 | def get_builder_cls() -> type[Qwen4ExpQSAMetadataBuilder]: |
| 114 | return Qwen4ExpQSAMetadataBuilder |
| 115 | |
| 116 | @classmethod |
| 117 | def is_sparse(cls) -> bool: |
| 118 | return True |
| 119 | |
| 120 | @classmethod |
| 121 | def supports_kv_connector(cls) -> bool: |
| 122 | return False |
| 123 | |
| 124 | |
| 125 | class Qwen4ExpQSAFlashAttentionImpl(FlashAttentionImpl): |
| 126 | """Run paged sparse GQA with the QSA Triton kernel.""" |
| 127 | |
| 128 | supports_dcp: bool = False |
| 129 | supports_pcp: bool = False |
| 130 | |
| 131 | def __init__( |
| 132 | self, |
| 133 | num_heads: int, |
| 134 | head_size: int, |
| 135 | scale: float, |
| 136 | num_kv_heads: int, |
| 137 | alibi_slopes: list[float] | None, |
| 138 | sliding_window: int | None, |
| 139 | kv_cache_dtype: str, |
| 140 | logits_soft_cap: float | None = None, |
| 141 | attn_type: AttentionType = AttentionType.DECODER, |
| 142 | kv_sharing_target_layer_name: str | None = None, |
| 143 | sinks: torch.Tensor | None = None, |
| 144 | ) -> None: |
| 145 | # The parent constructor probes flash-attn for quantized-KV support and |
| 146 | # raises where it is unavailable (sm120), but QSA dequantizes fp8 inside |
| 147 | # its own Triton kernel and never runs flash-attn over the cache. Hand |
| 148 | # the parent "auto" for that probe and restore the real dtype afterwards: |
| 149 | # the parent only uses it there, and do_kv_cache_update reads the |
| 150 | # attribute at call time. |
| 151 | real_kv_cache_dtype = kv_cache_dtype |
| 152 | if kv_cache_dtype in ("fp8", "fp8_e4m3"): |
| 153 | kv_cache_dtype = "auto" |
| 154 | super().__init__( |
| 155 | num_heads, |
| 156 | head_size, |
| 157 | scale, |
| 158 | num_kv_heads, |
| 159 | alibi_slopes, |
| 160 | sliding_window, |
| 161 | kv_cache_dtype, |
| 162 | logits_soft_cap, |
| 163 | attn_type, |
| 164 | kv_sharing_target_layer_name, |
| 165 | sinks, |
| 166 | ) |
| 167 | self.kv_cache_dtype = real_kv_cache_dtype |
| 168 | if not is_flash_attn_varlen_func_available(): |
| 169 | raise NotImplementedError("Qwen4Exp QSA requires FlashAttention") |
| 170 | if self.dcp_world_size != 1: |
| 171 | raise NotImplementedError( |
| 172 | "Qwen4Exp QSA does not support decode context parallelism" |
| 173 | ) |
| 174 | if self.kv_cache_dtype not in ("auto", "bfloat16", "fp8", "fp8_e4m3"): |
| 175 | raise NotImplementedError( |
| 176 | "Qwen4Exp QSA requires a BF16 or FP8-e4m3 main KV cache" |
| 177 | ) |
| 178 | self.supports_quant_query_input = False |
| 179 | |
| 180 | def forward_qsa( |
| 181 | self, |
| 182 | layer: torch.nn.Module, |
| 183 | query: torch.Tensor, |
| 184 | key: torch.Tensor, |
| 185 | value: torch.Tensor, |
| 186 | kv_cache: torch.Tensor, |
| 187 | attn_metadata: FlashAttentionMetadata, |
| 188 | output: torch.Tensor, |
| 189 | token_to_req: torch.Tensor, |
| 190 | use_prefill_config: bool, |
| 191 | output_gate: torch.Tensor, |
| 192 | output_scale: torch.Tensor | None = None, |
| 193 | output_block_scale: torch.Tensor | None = None, |
| 194 | ) -> torch.Tensor: |
| 195 | del key, value |
| 196 | if output_scale is not None or output_block_scale is not None: |
| 197 | raise NotImplementedError("QSA does not support fused output quantization") |
| 198 | if self.alibi_slopes is not None or self.sinks is not None: |
| 199 | raise NotImplementedError("QSA does not support ALiBi or attention sinks") |
| 200 | if self.sliding_window != (-1, -1): |
| 201 | raise NotImplementedError("QSA does not support sliding-window attention") |
| 202 | |
| 203 | num_tokens = attn_metadata.num_actual_tokens |
| 204 | output.zero_() |
| 205 | if num_tokens == 0: |
| 206 | return output |
| 207 | |
| 208 | topk_buffer = getattr(layer, "topk_indices_buffer", None) |
| 209 | if topk_buffer is None: |
| 210 | raise RuntimeError("QSA owner did not provide its top-k buffer") |
| 211 | logical_indices = topk_buffer[:num_tokens] |
| 212 | token_to_req = token_to_req[:num_tokens] |
| 213 | key_cache, value_cache = kv_cache.transpose(1, 2).split(self.head_size, dim=-1) |
| 214 | k_scale = v_scale = None |
| 215 | if self.kv_cache_dtype in ("fp8", "fp8_e4m3"): |
| 216 | # The cache is allocated as uint8; reinterpret the e4m3 bytes |
| 217 | # (same itemsize, so shape and strides are preserved). |
| 218 | key_cache = key_cache.view(torch.float8_e4m3fn) |
| 219 | value_cache = value_cache.view(torch.float8_e4m3fn) |
| 220 | # Host-side per-tensor dequant scales (Python floats), as used by |
| 221 | # other host-scale backends; folded into the kernel's scales. |
| 222 | k_scale = layer._k_scale_float |
| 223 | v_scale = layer._v_scale_float |
| 224 | if query.dtype != torch.bfloat16 or key_cache.dtype not in ( |
| 225 | torch.bfloat16, |
| 226 | torch.float8_e4m3fn, |
| 227 | ): |
| 228 | raise NotImplementedError( |
| 229 | "Qwen4Exp QSA requires BF16 Q and BF16 or FP8-e4m3 K/V" |
| 230 | ) |
| 231 | |
| 232 | from .ops.qsa import qsa_sparse_paged_attention |
| 233 | |
| 234 | qsa_sparse_paged_attention( |
| 235 | query[:num_tokens], |
| 236 | key_cache, |
| 237 | value_cache, |
| 238 | logical_indices, |
| 239 | attn_metadata.block_table, |
| 240 | token_to_req, |
| 241 | use_prefill_config, |
| 242 | output[:num_tokens], |
| 243 | k_scale=k_scale, |
| 244 | v_scale=v_scale, |
| 245 | output_gate=output_gate[:num_tokens], |
| 246 | ) |
| 247 | return output |
| 248 | |
| 249 | |
| 250 | class Qwen4ExpQSAAttention(Qwen3NextAttention, AttentionLayerBase): |
| 251 | """Merged Qwen full-attention owner with a QSA index side branch.""" |
| 252 | |
| 253 | supports_dcp = False |
| 254 | |
| 255 | def __init__( |
| 256 | self, |
| 257 | *, |
| 258 | vllm_config: VllmConfig, |
| 259 | config: Qwen4ExpTextConfig, |
| 260 | layer_id: int, |
| 261 | quant_config: QuantizationConfig | None = None, |
| 262 | reduce_results: bool = True, |
| 263 | prefix: str = "", |
| 264 | ) -> None: |
| 265 | nn.Module.__init__(self) |
| 266 | cache_config = vllm_config.cache_config |
| 267 | model_config = vllm_config.model_config |
| 268 | if cache_config is None: |
| 269 | raise ValueError("Qwen4Exp QSA requires a paged KV cache") |
| 270 | if model_config.dtype != torch.bfloat16: |
| 271 | raise NotImplementedError("Qwen4Exp QSA currently requires BF16") |
| 272 | if cache_config.cache_dtype not in ( |
| 273 | "auto", |
| 274 | "bfloat16", |
| 275 | "fp8", |
| 276 | "fp8_e4m3", |
| 277 | ): |
| 278 | raise NotImplementedError( |
| 279 | "Qwen4Exp QSA requires a BF16 or FP8-e4m3 main KV cache" |
| 280 | ) |
| 281 | if getattr(quant_config, "kv_cache_scheme", None) is not None: |
| 282 | raise NotImplementedError("Qwen4Exp QSA does not support KV quantization") |
| 283 | parallel_config = vllm_config.parallel_config |
| 284 | if ( |
| 285 | parallel_config.prefill_context_parallel_size > 1 |
| 286 | or parallel_config.decode_context_parallel_size > 1 |
| 287 | ): |
| 288 | raise NotImplementedError( |
| 289 | "Qwen4Exp QSA does not support context parallelism" |
| 290 | ) |
| 291 | if not getattr(config, "is_causal", True): |
| 292 | raise NotImplementedError("Qwen4Exp QSA requires causal decoder attention") |
| 293 | |
| 294 | self.config = config |
| 295 | self.hidden_size = int(config.hidden_size) |
| 296 | tp_size = get_tensor_model_parallel_world_size() |
| 297 | self.total_num_heads = int(config.num_attention_heads) |
| 298 | if self.total_num_heads % tp_size: |
| 299 | raise ValueError("QSA attention heads must be divisible by TP size") |
| 300 | self.num_heads = self.total_num_heads // tp_size |
| 301 | # Decode/verify batches have at most 1 + num_spec query tokens per |
| 302 | # request; use_prefill_config (max_query_len > this) steers the |
| 303 | # config table. Shorter batches take the decode profile — harmless, |
| 304 | # the difference is tile-shape tuning, not correctness. |
| 305 | self._max_decode_query_len = 1 + vllm_config.num_speculative_tokens |
| 306 | self.total_num_kv_heads = int(config.num_key_value_heads) |
| 307 | if self.total_num_kv_heads >= tp_size: |
| 308 | if self.total_num_kv_heads % tp_size: |
| 309 | raise ValueError("QSA KV heads must be divisible by TP size") |
| 310 | elif tp_size % self.total_num_kv_heads: |
| 311 | raise ValueError("TP size must be divisible by replicated QSA KV heads") |
| 312 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 313 | self.head_dim = int(config.head_dim or self.hidden_size // self.num_heads) |
| 314 | self.q_size = self.num_heads * self.head_dim |
| 315 | self.kv_size = self.num_kv_heads * self.head_dim |
| 316 | self.scaling = self.head_dim**-0.5 |
| 317 | self.dual_chunk_attention_config = getattr( |
| 318 | config, "dual_chunk_attention_config", None |
| 319 | ) |
| 320 | if self.dual_chunk_attention_config is not None: |
| 321 | raise NotImplementedError("Qwen4Exp QSA does not support dual-chunk RoPE") |
| 322 | # Qwen4Exp full-attention checkpoints always pack a sigmoid output |
| 323 | # gate next to Q, even when an inherited config default says otherwise. |
| 324 | self.attn_output_gate = True |
| 325 | |
| 326 | self.qkv_proj = QKVParallelLinear( |
| 327 | self.hidden_size, |
| 328 | self.head_dim, |
| 329 | self.total_num_heads * (1 + self.attn_output_gate), |
| 330 | self.total_num_kv_heads, |
| 331 | bias=False, |
| 332 | quant_config=model.without_modelopt_fp4(quant_config), |
| 333 | prefix=f"{prefix}.qkv_proj", |
| 334 | ) |
| 335 | self.o_proj = RowParallelLinear( |
| 336 | self.total_num_heads * self.head_dim, |
| 337 | self.hidden_size, |
| 338 | bias=False, |
| 339 | reduce_results=reduce_results, |
| 340 | quant_config=quant_config, |
| 341 | prefix=f"{prefix}.o_proj", |
| 342 | ) |
| 343 | self.rotary_emb = get_rope( |
| 344 | head_size=self.head_dim, |
| 345 | max_position=config.max_position_embeddings, |
| 346 | rope_parameters=config.rope_parameters, |
| 347 | ) |
| 348 | self.q_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 349 | self.k_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 350 | |
| 351 | mm_config = model_config.multimodal_config |
| 352 | text_only = mm_config is None or mm_config.language_model_only |
| 353 | mrope_section = getattr(self.rotary_emb, "mrope_section", None) |
| 354 | supports_mrope = bool( |
| 355 | type(self.rotary_emb) is MRotaryEmbedding |
| 356 | and mrope_section |
| 357 | and len(mrope_section) == 3 |
| 358 | and sum(mrope_section) == self.rotary_emb.rotary_dim // 2 |
| 359 | and getattr(self.rotary_emb, "mrope_interleaved", False) |
| 360 | ) |
| 361 | supports_dtype = getattr(self.rotary_emb, "dtype", None) in ( |
| 362 | torch.float16, |
| 363 | torch.bfloat16, |
| 364 | ) |
| 365 | self.use_fused_qk_norm_rope_gate = ( |
| 366 | self.attn_output_gate |
| 367 | and getattr(self.rotary_emb, "is_neox_style", False) |
| 368 | and current_platform.is_cuda() |
| 369 | and supports_dtype |
| 370 | and (text_only or supports_mrope) |
| 371 | ) |
| 372 | |
| 373 | self.layer_name = f"{prefix}.attn" |
| 374 | self.attn_type = AttentionType.DECODER |
| 375 | self.kv_cache_dtype = cache_config.cache_dtype |
| 376 | self.kv_cache_torch_dtype = kv_cache_dtype_str_to_dtype( |
| 377 | self.kv_cache_dtype, model_config |
| 378 | ) |
| 379 | if self.kv_cache_torch_dtype not in (torch.bfloat16, torch.uint8): |
| 380 | raise NotImplementedError( |
| 381 | "Qwen4Exp QSA requires BF16 or FP8-e4m3 (uint8) cache storage" |
| 382 | ) |
| 383 | self.kv_sharing_target_layer_name = None |
| 384 | self.kv_cache = torch.tensor([]) |
| 385 | set_default_quant_scales(self, register_buffer=True) |
| 386 | |
| 387 | self.attn_backend = Qwen4ExpQSAFlashAttentionBackend |
| 388 | self.impl = Qwen4ExpQSAFlashAttentionImpl( |
| 389 | self.num_heads, |
| 390 | self.head_dim, |
| 391 | self.scaling, |
| 392 | self.num_kv_heads, |
| 393 | None, |
| 394 | None, |
| 395 | self.kv_cache_dtype, |
| 396 | None, |
| 397 | AttentionType.DECODER, |
| 398 | None, |
| 399 | ) |
| 400 | self.indexer = QSAIndexer( |
| 401 | vllm_config=vllm_config, |
| 402 | config=config, |
| 403 | layer_id=layer_id, |
| 404 | rotary_emb=self.rotary_emb, |
| 405 | quant_config=quant_config, |
| 406 | prefix=f"{prefix}.indexer", |
| 407 | ) |
| 408 | # One launch does the indexer prepare, the main QK-norm/RoPE/gate and |
| 409 | # the main K/V cache write (see QSAIndexer.forward); otherwise all of |
| 410 | # them take the separate kernels. |
| 411 | self.use_fused_qsa_prepare = ( |
| 412 | self.use_fused_qk_norm_rope_gate and self.indexer.use_fused_pre_indexer |
| 413 | ) |
| 414 | max_tokens = vllm_config.scheduler_config.max_num_batched_tokens |
| 415 | # PACKED selection buffer: the trailing column holds each row's |
| 416 | # valid-entry count (written by the expand kernel) — never a token |
| 417 | # index; the sparse attention kernel reads it as its loop bound. |
| 418 | # MTP skip_topk steps reuse rows frozen from step 0; the count is |
| 419 | # a row column, so compaction/reuse keep it paired with the content. |
| 420 | self.register_buffer( |
| 421 | "topk_indices_buffer", |
| 422 | torch.empty( |
| 423 | max_tokens, |
| 424 | self.indexer.packed_output_width, |
| 425 | dtype=torch.int32, |
| 426 | ), |
| 427 | persistent=False, |
| 428 | ) |
| 429 | |
| 430 | static_context = vllm_config.compilation_config.static_forward_context |
| 431 | if self.layer_name in static_context: |
| 432 | raise ValueError(f"Duplicate layer name: {self.layer_name}") |
| 433 | static_context[self.layer_name] = self |
| 434 | |
| 435 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 436 | return self.attn_backend |
| 437 | |
| 438 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec: |
| 439 | return FullAttentionSpec( |
| 440 | block_size=vllm_config.cache_config.block_size, |
| 441 | num_kv_heads=self.num_kv_heads, |
| 442 | head_size=self.head_dim, |
| 443 | head_size_v=self.head_dim, |
| 444 | dtype=self.kv_cache_torch_dtype, |
| 445 | kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype), |
| 446 | ) |
| 447 | |
| 448 | @eager_break_during_capture |
| 449 | def _run_qsa( |
| 450 | self, |
| 451 | projected_qk: torch.Tensor, |
| 452 | positions: torch.Tensor, |
| 453 | query: torch.Tensor | None, |
| 454 | key: torch.Tensor | None, |
| 455 | value: torch.Tensor | None, |
| 456 | output: torch.Tensor, |
| 457 | output_gate: torch.Tensor | None, |
| 458 | qkv: torch.Tensor, |
| 459 | ) -> None: |
| 460 | # query/key/value/output_gate are None when the fused prepare runs |
| 461 | # inside the indexer launch. |
| 462 | metadata = get_forward_context().attn_metadata |
| 463 | if isinstance(metadata, list): |
| 464 | metadata = metadata[0] |
| 465 | if not isinstance(metadata, dict): |
| 466 | output.zero_() |
| 467 | return |
| 468 | main_metadata = cast(FlashAttentionMetadata, metadata[self.layer_name]) |
| 469 | if self.kv_cache.numel() == 0: |
| 470 | raise RuntimeError("QSA main K/V cache is not bound") |
| 471 | |
| 472 | num_tokens = main_metadata.num_actual_tokens |
| 473 | side_metadata = cast( |
| 474 | QSAForwardMetadata, |
| 475 | metadata[self.indexer.raw_key_cache.prefix], |
| 476 | ) |
| 477 | if side_metadata.num_actual_tokens != num_tokens: |
| 478 | raise RuntimeError("QSA main and side metadata token counts disagree") |
| 479 | selected, main_outputs = self.indexer( |
| 480 | projected_qk, |
| 481 | positions, |
| 482 | self.topk_indices_buffer[:num_tokens], |
| 483 | attn=self, |
| 484 | qkv=qkv, |
| 485 | slot_mapping=main_metadata.slot_mapping, |
| 486 | ) |
| 487 | if selected.shape != (num_tokens, self.indexer.packed_output_width): |
| 488 | raise RuntimeError("QSA indexer returned an invalid selection shape") |
| 489 | impl = cast(Qwen4ExpQSAFlashAttentionImpl, self.impl) |
| 490 | if main_outputs is None: |
| 491 | assert key is not None and value is not None |
| 492 | impl.do_kv_cache_update( |
| 493 | self, |
| 494 | key, |
| 495 | value, |
| 496 | self.kv_cache, |
| 497 | main_metadata.slot_mapping, |
| 498 | ) |
| 499 | else: |
| 500 | query, output_gate = main_outputs |
| 501 | assert query is not None and output_gate is not None |
| 502 | impl.forward_qsa( |
| 503 | self, |
| 504 | query, |
| 505 | key, |
| 506 | value, |
| 507 | self.kv_cache, |
| 508 | main_metadata, |
| 509 | output, |
| 510 | token_to_req=side_metadata.token_to_req, |
| 511 | use_prefill_config=main_metadata.max_query_len > self._max_decode_query_len, |
| 512 | output_gate=output_gate, |
| 513 | ) |
| 514 | |
| 515 | def forward( |
| 516 | self, |
| 517 | positions: torch.Tensor, |
| 518 | hidden_states: torch.Tensor, |
| 519 | ) -> torch.Tensor: |
| 520 | qkv, _ = self.qkv_proj(hidden_states) |
| 521 | num_tokens = hidden_states.shape[0] |
| 522 | if not self.use_fused_qsa_prepare: |
| 523 | q, k, v, gate = self._project_qkv_gate(qkv, positions) |
| 524 | assert gate is not None |
| 525 | query = q.view(num_tokens, self.num_heads, self.head_dim) |
| 526 | key = k.view(num_tokens, self.num_kv_heads, self.head_dim) |
| 527 | value = v.view(num_tokens, self.num_kv_heads, self.head_dim) |
| 528 | else: |
| 529 | # Norm/RoPE/gate and the K/V cache write happen inside _run_qsa. |
| 530 | query = key = value = gate = None |
| 531 | attn_output = qkv.new_empty(num_tokens, self.num_heads, self.head_dim) |
| 532 | # Keep the index projection outside the eager break. |
| 533 | projected_qk, _ = self.indexer.index_qk_proj(hidden_states) |
| 534 | self._run_qsa( |
| 535 | projected_qk, |
| 536 | positions, |
| 537 | query, |
| 538 | key, |
| 539 | value, |
| 540 | attn_output, |
| 541 | gate, |
| 542 | qkv, |
| 543 | ) |
| 544 | flat_output = attn_output.view(num_tokens, -1) |
| 545 | output, _ = self.o_proj(flat_output) |
| 546 | return output |
| 547 | |
| 548 | |
| 549 | __all__ = [ |
| 550 | "QSAIndexer", |
| 551 | "Qwen4ExpQSAAttention", |
| 552 | "Qwen4ExpQSAFlashAttentionBackend", |
| 553 | "Qwen4ExpQSAFlashAttentionImpl", |
| 554 | ] |