| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import copy |
| 4 | |
| 5 | import torch |
| 6 | import torch.nn.functional as F |
| 7 | from transformers.configuration_utils import PreTrainedConfig |
| 8 | |
| 9 | from vllm.config import ( |
| 10 | VllmConfig, |
| 11 | get_current_vllm_config, |
| 12 | ) |
| 13 | from vllm.distributed import ( |
| 14 | get_tensor_model_parallel_rank, |
| 15 | get_tensor_model_parallel_world_size, |
| 16 | ) |
| 17 | from vllm.forward_context import get_forward_context |
| 18 | from vllm.model_executor.custom_op import PluggableLayer |
| 19 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 20 | from vllm.model_executor.layers.linear import ( |
| 21 | ColumnParallelLinear, |
| 22 | QKVParallelLinear, |
| 23 | RowParallelLinear, |
| 24 | ) |
| 25 | from vllm.model_executor.layers.mamba.linear.base import LinearAttention |
| 26 | from vllm.model_executor.layers.mamba.linear.minimax_linear_attn import ( |
| 27 | MiniMaxText01LinearAttention, |
| 28 | MiniMaxText01LinearKernel, |
| 29 | linear_attention_decode, |
| 30 | ) |
| 31 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 32 | from vllm.third_party.flash_linear_attention.ops.layernorm_guard import ( |
| 33 | RMSNormGated, |
| 34 | layernorm_fn, |
| 35 | ) |
| 36 | from vllm.triton_utils import tl, triton |
| 37 | from vllm.v1.attention.backends.linear_attn import LinearAttentionMetadata |
| 38 | |
| 39 | |
| 40 | def _build_rope_parameters(config: PreTrainedConfig) -> dict | None: |
| 41 | rope_parameters = copy.deepcopy(getattr(config, "rope_parameters", None)) or {} |
| 42 | if "rope_theta" not in rope_parameters and hasattr(config, "rope_theta"): |
| 43 | rope_parameters["rope_theta"] = config.rope_theta |
| 44 | if "partial_rotary_factor" not in rope_parameters and hasattr( |
| 45 | config, "partial_rotary_factor" |
| 46 | ): |
| 47 | rope_parameters["partial_rotary_factor"] = config.partial_rotary_factor |
| 48 | |
| 49 | rope_scaling = getattr(config, "rope_scaling", None) |
| 50 | if isinstance(rope_scaling, dict): |
| 51 | rope_scaling = copy.deepcopy(rope_scaling) |
| 52 | if "type" in rope_scaling and "rope_type" not in rope_scaling: |
| 53 | rope_scaling["rope_type"] = rope_scaling.pop("type") |
| 54 | rope_parameters.update(rope_scaling) |
| 55 | |
| 56 | return rope_parameters or None |
| 57 | |
| 58 | |
| 59 | def clear_linear_attention_cache_for_new_sequences( |
| 60 | kv_cache: torch.Tensor, |
| 61 | state_indices_tensor: torch.Tensor, |
| 62 | attn_metadata: LinearAttentionMetadata, |
| 63 | ) -> None: |
| 64 | num_prefills = getattr(attn_metadata, "num_prefills", 0) |
| 65 | if num_prefills <= 0: |
| 66 | return |
| 67 | |
| 68 | num_decodes = getattr(attn_metadata, "num_decodes", 0) |
| 69 | prefill_state_indices = getattr(attn_metadata, "state_indices_tensor_p", None) |
| 70 | for prefill_idx in range(num_prefills): |
| 71 | if num_decodes + prefill_idx + 1 >= len(attn_metadata.query_start_loc): |
| 72 | break |
| 73 | q_start = attn_metadata.query_start_loc[num_decodes + prefill_idx] |
| 74 | q_end = attn_metadata.query_start_loc[num_decodes + prefill_idx + 1] |
| 75 | query_len = q_end - q_start |
| 76 | context_len = attn_metadata.seq_lens[num_decodes + prefill_idx] - query_len |
| 77 | if context_len == 0: |
| 78 | if prefill_state_indices is not None: |
| 79 | block_to_clear = prefill_state_indices[prefill_idx] |
| 80 | else: |
| 81 | block_to_clear = state_indices_tensor[num_decodes + prefill_idx] |
| 82 | kv_cache[block_to_clear, ...] = 0 |
| 83 | |
| 84 | |
| 85 | def linear_attention_prefill_and_mix( |
| 86 | q: torch.Tensor, |
| 87 | k: torch.Tensor, |
| 88 | v: torch.Tensor, |
| 89 | kv_cache: torch.Tensor, |
| 90 | state_indices_tensor: torch.Tensor, |
| 91 | attn_metadata: LinearAttentionMetadata, |
| 92 | slope_rate: torch.Tensor, |
| 93 | block_size: int, |
| 94 | decode_fn, |
| 95 | prefix_fn, |
| 96 | layer_idx: int | None = None, |
| 97 | ) -> torch.Tensor: |
| 98 | hidden = [] |
| 99 | req_offset = getattr(attn_metadata, "num_decodes", 0) |
| 100 | prefill_state_indices = getattr(attn_metadata, "state_indices_tensor_p", None) |
| 101 | for _prefill_idx in range(getattr(attn_metadata, "num_prefills", 0)): |
| 102 | if req_offset + _prefill_idx + 1 >= len(attn_metadata.query_start_loc): |
| 103 | break |
| 104 | if prefill_state_indices is not None and _prefill_idx >= len( |
| 105 | prefill_state_indices |
| 106 | ): |
| 107 | break |
| 108 | if prefill_state_indices is None and _prefill_idx >= len(state_indices_tensor): |
| 109 | break |
| 110 | _start = attn_metadata.query_start_loc[req_offset + _prefill_idx] |
| 111 | _end = attn_metadata.query_start_loc[req_offset + _prefill_idx + 1] |
| 112 | if prefill_state_indices is not None: |
| 113 | slot_id = prefill_state_indices[_prefill_idx] |
| 114 | else: |
| 115 | slot_id = state_indices_tensor[req_offset + _prefill_idx] |
| 116 | qs = q[_start:_end].transpose(0, 1).contiguous() |
| 117 | ks = k[_start:_end].transpose(0, 1).contiguous() |
| 118 | vs = v[_start:_end].transpose(0, 1).contiguous() |
| 119 | slice_layer_cache = kv_cache[slot_id, ...] |
| 120 | out_slice = prefix_fn( |
| 121 | qs, |
| 122 | ks, |
| 123 | vs, |
| 124 | slice_layer_cache, |
| 125 | slope_rate, |
| 126 | block_size, |
| 127 | layer_idx=layer_idx, |
| 128 | ) |
| 129 | hidden.append(out_slice.contiguous()) |
| 130 | |
| 131 | if attn_metadata.num_decode_tokens > 0: |
| 132 | hidden_decode = decode_fn( |
| 133 | q, k, v, kv_cache, state_indices_tensor, attn_metadata |
| 134 | ) |
| 135 | hidden.insert(0, hidden_decode) |
| 136 | |
| 137 | if not hidden: |
| 138 | return torch.empty((0, q.size(1) * q.size(2)), device=q.device, dtype=q.dtype) |
| 139 | |
| 140 | hidden = torch.concat(hidden, dim=0).contiguous() |
| 141 | return hidden |
| 142 | |
| 143 | |
| 144 | @triton.jit |
| 145 | def _bailing_linear_attn_decode_spec_step_kernel( |
| 146 | q_ptr, |
| 147 | k_ptr, |
| 148 | v_ptr, |
| 149 | kv_cache_ptr, |
| 150 | slope_rate_ptr, |
| 151 | state_indices_ptr, |
| 152 | query_start_loc_ptr, |
| 153 | num_accepted_tokens_ptr, |
| 154 | output_ptr, |
| 155 | q_start: tl.constexpr, |
| 156 | D: tl.constexpr, |
| 157 | q_b_stride, |
| 158 | q_h_stride, |
| 159 | q_d_stride, |
| 160 | k_b_stride, |
| 161 | k_h_stride, |
| 162 | k_d_stride, |
| 163 | v_b_stride, |
| 164 | v_h_stride, |
| 165 | v_d_stride, |
| 166 | cache_b_stride, |
| 167 | cache_h_stride, |
| 168 | cache_d0_stride, |
| 169 | cache_d1_stride, |
| 170 | state_indices_b_stride, |
| 171 | state_indices_t_stride, |
| 172 | output_b_stride, |
| 173 | output_d_stride, |
| 174 | DRAFT_IDX: tl.constexpr, |
| 175 | STATE_WIDTH: tl.constexpr, |
| 176 | BLOCK_SIZE: tl.constexpr, |
| 177 | ): |
| 178 | req_id = tl.program_id(0) |
| 179 | head_id = tl.program_id(1) |
| 180 | block_id = tl.program_id(2) |
| 181 | |
| 182 | req_start = tl.load(query_start_loc_ptr + req_id).to(tl.int64) |
| 183 | req_end = tl.load(query_start_loc_ptr + req_id + 1).to(tl.int64) |
| 184 | query_len = req_end - req_start |
| 185 | if query_len <= DRAFT_IDX: |
| 186 | return |
| 187 | |
| 188 | dst_slot = tl.load( |
| 189 | state_indices_ptr |
| 190 | + req_id * state_indices_b_stride |
| 191 | + DRAFT_IDX * state_indices_t_stride |
| 192 | ).to(tl.int64) |
| 193 | if dst_slot == -1: |
| 194 | return |
| 195 | |
| 196 | if DRAFT_IDX == 0: |
| 197 | accepted_offset = tl.load(num_accepted_tokens_ptr + req_id).to(tl.int64) - 1 |
| 198 | accepted_offset = tl.maximum(accepted_offset, 0) |
| 199 | accepted_offset = tl.minimum(accepted_offset, STATE_WIDTH - 1) |
| 200 | src_slot = tl.load( |
| 201 | state_indices_ptr |
| 202 | + req_id * state_indices_b_stride |
| 203 | + accepted_offset * state_indices_t_stride |
| 204 | ).to(tl.int64) |
| 205 | else: |
| 206 | src_slot = tl.load( |
| 207 | state_indices_ptr |
| 208 | + req_id * state_indices_b_stride |
| 209 | + (DRAFT_IDX - 1) * state_indices_t_stride |
| 210 | ).to(tl.int64) |
| 211 | if src_slot == -1: |
| 212 | return |
| 213 | |
| 214 | token_idx = req_start - q_start + DRAFT_IDX |
| 215 | qk_offsets = tl.arange(0, D) |
| 216 | v_offsets = tl.arange(0, BLOCK_SIZE) + block_id * BLOCK_SIZE |
| 217 | qk_mask = qk_offsets < D |
| 218 | v_mask = v_offsets < D |
| 219 | kv_mask = qk_mask[:, None] & v_mask[None, :] |
| 220 | |
| 221 | q = tl.load( |
| 222 | q_ptr + token_idx * q_b_stride + head_id * q_h_stride + qk_offsets * q_d_stride, |
| 223 | mask=qk_mask, |
| 224 | other=0.0, |
| 225 | ) |
| 226 | k = tl.load( |
| 227 | k_ptr + token_idx * k_b_stride + head_id * k_h_stride + qk_offsets * k_d_stride, |
| 228 | mask=qk_mask, |
| 229 | other=0.0, |
| 230 | ) |
| 231 | v = tl.load( |
| 232 | v_ptr + token_idx * v_b_stride + head_id * v_h_stride + v_offsets * v_d_stride, |
| 233 | mask=v_mask, |
| 234 | other=0.0, |
| 235 | ) |
| 236 | |
| 237 | cache_offsets = ( |
| 238 | qk_offsets[:, None] * cache_d0_stride + v_offsets[None, :] * cache_d1_stride |
| 239 | ) |
| 240 | src_cache_ptr = ( |
| 241 | kv_cache_ptr |
| 242 | + src_slot * cache_b_stride |
| 243 | + head_id * cache_h_stride |
| 244 | + cache_offsets |
| 245 | ) |
| 246 | dst_cache_ptr = ( |
| 247 | kv_cache_ptr |
| 248 | + dst_slot * cache_b_stride |
| 249 | + head_id * cache_h_stride |
| 250 | + cache_offsets |
| 251 | ) |
| 252 | |
| 253 | slope = tl.load(slope_rate_ptr + head_id) |
| 254 | decay = tl.exp(-slope) |
| 255 | kv_old = tl.load(src_cache_ptr, mask=kv_mask, other=0.0) |
| 256 | kv_new = k[:, None] * v[None, :] + decay * kv_old |
| 257 | |
| 258 | output = tl.sum(q[:, None].to(tl.float32) * kv_new, axis=0) |
| 259 | tl.store(dst_cache_ptr, kv_new, mask=kv_mask) |
| 260 | tl.store( |
| 261 | output_ptr |
| 262 | + token_idx * output_b_stride |
| 263 | + (head_id * D + v_offsets) * output_d_stride, |
| 264 | output, |
| 265 | mask=v_mask, |
| 266 | ) |
| 267 | |
| 268 | |
| 269 | def bailing_linear_attention_decode_spec( |
| 270 | q: torch.Tensor, |
| 271 | k: torch.Tensor, |
| 272 | v: torch.Tensor, |
| 273 | kv_cache: torch.Tensor, |
| 274 | slope_rate: torch.Tensor, |
| 275 | state_indices_tensor: torch.Tensor, |
| 276 | query_start_loc: torch.Tensor, |
| 277 | num_accepted_tokens: torch.Tensor, |
| 278 | q_start: int, |
| 279 | q_end: int | None, |
| 280 | slot_start: int, |
| 281 | slot_end: int | None, |
| 282 | block_size: int, |
| 283 | ) -> torch.Tensor: |
| 284 | q_decode = q[q_start:q_end] |
| 285 | k_decode = k[q_start:q_end] |
| 286 | v_decode = v[q_start:q_end] |
| 287 | hidden = torch.empty( |
| 288 | (q_decode.shape[0], q.shape[1] * q.shape[2]), |
| 289 | device=q.device, |
| 290 | dtype=q.dtype, |
| 291 | ) |
| 292 | hidden.zero_() |
| 293 | |
| 294 | state_indices_tensor = state_indices_tensor[slot_start:slot_end] |
| 295 | query_start_loc = query_start_loc.to(device=q.device) |
| 296 | |
| 297 | batch_size = state_indices_tensor.shape[0] |
| 298 | num_heads = q_decode.shape[1] |
| 299 | head_dim = q_decode.shape[2] |
| 300 | assert k_decode.shape == (q_decode.shape[0], num_heads, head_dim) |
| 301 | assert v_decode.shape == (q_decode.shape[0], num_heads, head_dim) |
| 302 | state_width = state_indices_tensor.shape[1] |
| 303 | |
| 304 | grid = (batch_size, num_heads, triton.cdiv(head_dim, block_size)) |
| 305 | for draft_idx in range(state_width): |
| 306 | _bailing_linear_attn_decode_spec_step_kernel[grid]( |
| 307 | q_decode, |
| 308 | k_decode, |
| 309 | v_decode, |
| 310 | kv_cache, |
| 311 | slope_rate, |
| 312 | state_indices_tensor, |
| 313 | query_start_loc, |
| 314 | num_accepted_tokens[:batch_size], |
| 315 | hidden, |
| 316 | q_start, |
| 317 | head_dim, |
| 318 | q_decode.stride(0), |
| 319 | q_decode.stride(1), |
| 320 | q_decode.stride(2), |
| 321 | k_decode.stride(0), |
| 322 | k_decode.stride(1), |
| 323 | k_decode.stride(2), |
| 324 | v_decode.stride(0), |
| 325 | v_decode.stride(1), |
| 326 | v_decode.stride(2), |
| 327 | kv_cache.stride(0), |
| 328 | kv_cache.stride(1), |
| 329 | kv_cache.stride(2), |
| 330 | kv_cache.stride(3), |
| 331 | state_indices_tensor.stride(0), |
| 332 | state_indices_tensor.stride(1), |
| 333 | hidden.stride(0), |
| 334 | hidden.stride(1), |
| 335 | DRAFT_IDX=draft_idx, |
| 336 | STATE_WIDTH=state_width, |
| 337 | BLOCK_SIZE=block_size, |
| 338 | ) |
| 339 | |
| 340 | return hidden |
| 341 | |
| 342 | |
| 343 | class BailingGroupRMSNormGate(RMSNormGated): |
| 344 | def __init__( |
| 345 | self, |
| 346 | hidden_size, |
| 347 | eps=1e-5, |
| 348 | group_size=None, |
| 349 | norm_before_gate=True, |
| 350 | device=None, |
| 351 | dtype=None, |
| 352 | ): |
| 353 | super().__init__( |
| 354 | hidden_size, |
| 355 | eps=eps, |
| 356 | group_size=group_size, |
| 357 | norm_before_gate=norm_before_gate, |
| 358 | device=device, |
| 359 | dtype=dtype, |
| 360 | activation="sigmoid", |
| 361 | ) |
| 362 | # Add custom weight loader for TP sharding |
| 363 | self.weight.weight_loader = self._weight_loader |
| 364 | |
| 365 | @staticmethod |
| 366 | def _weight_loader(param: torch.nn.Parameter, loaded_weight: torch.Tensor) -> None: |
| 367 | """Load weight with TP sharding.""" |
| 368 | tp_size = get_tensor_model_parallel_world_size() |
| 369 | tp_rank = get_tensor_model_parallel_rank() |
| 370 | shard_size = loaded_weight.shape[0] // tp_size |
| 371 | shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size) |
| 372 | param.data.copy_(loaded_weight[shard].contiguous()) |
| 373 | |
| 374 | |
| 375 | # --8<-- [start:bailing_moe_linear_attention] |
| 376 | @PluggableLayer.register("bailing_moe_linear_attention") |
| 377 | class BailingMoELinearAttention(LinearAttention): |
| 378 | """Pluggable Bailing MoE Linear Attention layer which allows OOT backends |
| 379 | to add custom implementations. |
| 380 | |
| 381 | This implements the linear attention mechanism from sglang, adapted for |
| 382 | vLLM's v1 engine with MambaBase interface support. |
| 383 | """ |
| 384 | |
| 385 | # --8<-- [end:bailing_moe_linear_attention] |
| 386 | def __init__( |
| 387 | self, |
| 388 | config: PreTrainedConfig, |
| 389 | vllm_config: VllmConfig, |
| 390 | prefix: str = "linear_attn", |
| 391 | ): |
| 392 | super().__init__(config, vllm_config, prefix) |
| 393 | |
| 394 | self.scaling = self.head_dim**-0.5 |
| 395 | |
| 396 | self.tp_heads = self.num_heads // self.tp_size |
| 397 | |
| 398 | self.max_position_embeddings = config.max_position_embeddings |
| 399 | self.rope_theta = getattr(config, "rope_theta", 600000) |
| 400 | |
| 401 | self.tp_kv_heads = self.num_heads // self.tp_size |
| 402 | self.q_size_per_rank = self.head_dim * self.tp_heads |
| 403 | self.kv_size_per_rank = self.head_dim * self.tp_kv_heads |
| 404 | |
| 405 | self.use_qk_norm = getattr(config, "use_qk_norm", False) |
| 406 | self.linear_backend = "minimax" |
| 407 | self.linear_scale = self.linear_backend == "minimax" |
| 408 | self.linear_rope = getattr(config, "linear_rope", True) |
| 409 | if hasattr(config, "use_linear_silu"): |
| 410 | self.linear_silu = config.use_linear_silu |
| 411 | elif hasattr(config, "linear_silu"): |
| 412 | self.linear_silu = config.linear_silu |
| 413 | else: |
| 414 | self.linear_silu = False |
| 415 | |
| 416 | self.query_key_value = QKVParallelLinear( |
| 417 | self.hidden_size, |
| 418 | self.head_dim, |
| 419 | self.num_heads, |
| 420 | self.num_heads, # MHA: kv_heads = num_heads |
| 421 | bias=(config.use_bias or config.use_qkv_bias), |
| 422 | quant_config=self.quant_config, |
| 423 | prefix=f"{prefix}.query_key_value", |
| 424 | ) |
| 425 | |
| 426 | if self.use_qk_norm: |
| 427 | self.query_layernorm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 428 | self.key_layernorm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 429 | |
| 430 | self.g_proj = ColumnParallelLinear( |
| 431 | self.hidden_size, |
| 432 | self.hidden_inner_size, |
| 433 | bias=False, |
| 434 | quant_config=self.quant_config, |
| 435 | prefix=f"{prefix}.g_proj", |
| 436 | ) |
| 437 | self.dense = RowParallelLinear( |
| 438 | self.hidden_inner_size, |
| 439 | self.hidden_size, |
| 440 | bias=config.use_bias, |
| 441 | quant_config=self.quant_config, |
| 442 | prefix=f"{prefix}.dense", |
| 443 | reduce_results=True, |
| 444 | ) |
| 445 | |
| 446 | self.group_norm_size = getattr(config, "group_norm_size", 1) |
| 447 | self.rms_norm_eps = float(getattr(config, "rms_norm_eps", 1e-5)) |
| 448 | assert self.tp_size <= self.group_norm_size, ( |
| 449 | "tp_size must be <= group_norm_size for local rms norm" |
| 450 | ) |
| 451 | assert self.group_norm_size % self.tp_size == 0, ( |
| 452 | "group_norm_size must be divisible by tp_size" |
| 453 | ) |
| 454 | |
| 455 | # When group_norm_size == 1, group_size equals hidden_size // tp_size |
| 456 | self.g_norm = BailingGroupRMSNormGate( |
| 457 | hidden_size=self.hidden_inner_size // self.tp_size, |
| 458 | eps=self.rms_norm_eps, |
| 459 | group_size=( |
| 460 | self.hidden_inner_size // self.group_norm_size |
| 461 | if self.group_norm_size > 1 |
| 462 | else self.hidden_inner_size // self.tp_size |
| 463 | ), |
| 464 | ) |
| 465 | |
| 466 | # use fp32 rotary embedding |
| 467 | rope_parameters = _build_rope_parameters(config) |
| 468 | |
| 469 | self.rotary_emb = get_rope( |
| 470 | self.head_dim, |
| 471 | max_position=self.max_position_embeddings, |
| 472 | is_neox_style=True, |
| 473 | rope_parameters=rope_parameters or None, |
| 474 | ) |
| 475 | |
| 476 | # Build slope tensor for linear attention decay |
| 477 | slope_rate = MiniMaxText01LinearAttention._build_slope_tensor(self.num_heads) |
| 478 | if self.num_hidden_layers <= 1: |
| 479 | self.slope_rate = slope_rate * (1 + 1e-5) |
| 480 | else: |
| 481 | self.slope_rate = slope_rate * ( |
| 482 | 1 - self.layer_idx / (self.num_hidden_layers - 1) + 1e-5 |
| 483 | ) |
| 484 | self.tp_slope = self.slope_rate[ |
| 485 | self.tp_rank * self.tp_heads : (self.tp_rank + 1) * self.tp_heads |
| 486 | ].contiguous() |
| 487 | |
| 488 | # Register for compilation |
| 489 | compilation_config = get_current_vllm_config().compilation_config |
| 490 | if prefix in compilation_config.static_forward_context: |
| 491 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 492 | compilation_config.static_forward_context[prefix] = self |
| 493 | |
| 494 | def get_attn_backend(self): |
| 495 | from vllm.v1.attention.backends.linear_attn import ( |
| 496 | BailingLinearAttentionBackend, |
| 497 | ) |
| 498 | |
| 499 | return BailingLinearAttentionBackend |
| 500 | |
| 501 | @staticmethod |
| 502 | def weight_direct_load(param: torch.Tensor, loaded_weight: torch.Tensor) -> None: |
| 503 | """Load weight for linear attention layers. |
| 504 | |
| 505 | For FP8 quantized parameters, we need to use the weight_loader if available, |
| 506 | as it handles special cases like tensor parallelism sharding. |
| 507 | """ |
| 508 | # Check if param has a weight_loader (for vLLM ModelWeightParameter) |
| 509 | weight_loader = getattr(param, "weight_loader", None) |
| 510 | if weight_loader is not None: |
| 511 | # Use the weight_loader which handles TP sharding and quantization |
| 512 | weight_loader(param, loaded_weight) |
| 513 | else: |
| 514 | # Fall back to direct copy for standard tensors |
| 515 | assert param.size() == loaded_weight.size(), ( |
| 516 | f"Shape mismatch: {param.shape} vs {loaded_weight.shape}" |
| 517 | ) |
| 518 | param.data.copy_(loaded_weight) |
| 519 | |
| 520 | def forward( |
| 521 | self, |
| 522 | hidden_states: torch.Tensor, |
| 523 | output: torch.Tensor, |
| 524 | positions: torch.Tensor, |
| 525 | ) -> None: |
| 526 | """Forward method called by torch.ops.vllm.linear_attention.""" |
| 527 | torch.ops.vllm.linear_attention( |
| 528 | hidden_states, |
| 529 | output, |
| 530 | positions, |
| 531 | self.prefix, |
| 532 | ) |
| 533 | |
| 534 | def _forward( |
| 535 | self, |
| 536 | hidden_states: torch.Tensor, |
| 537 | output: torch.Tensor, |
| 538 | positions: torch.Tensor, |
| 539 | ) -> None: |
| 540 | """Actual forward implementation.""" |
| 541 | forward_context = get_forward_context() |
| 542 | attn_metadata = forward_context.attn_metadata |
| 543 | if attn_metadata is not None: |
| 544 | assert isinstance(attn_metadata, dict) |
| 545 | attn_metadata = attn_metadata.get(self.prefix) # type: ignore |
| 546 | if attn_metadata is not None: |
| 547 | assert isinstance(attn_metadata, LinearAttentionMetadata) |
| 548 | num_actual_tokens = ( |
| 549 | attn_metadata.num_prefill_tokens + attn_metadata.num_decode_tokens |
| 550 | ) |
| 551 | else: |
| 552 | num_actual_tokens = hidden_states.shape[0] |
| 553 | |
| 554 | # QKV projection |
| 555 | qkv, _ = self.query_key_value(hidden_states[:num_actual_tokens]) |
| 556 | |
| 557 | # use rotary_emb support fp32 |
| 558 | qkv = qkv.to(torch.float32) |
| 559 | if self.linear_silu: |
| 560 | qkv = F.silu(qkv) |
| 561 | |
| 562 | # Split q, k, v |
| 563 | q, k, v = torch.split( |
| 564 | qkv, |
| 565 | [self.q_size_per_rank, self.kv_size_per_rank, self.kv_size_per_rank], |
| 566 | dim=-1, |
| 567 | ) |
| 568 | |
| 569 | # Apply QK norm if needed |
| 570 | if self.use_qk_norm: |
| 571 | q = q.reshape(-1, self.tp_heads, self.head_dim) |
| 572 | k = k.reshape(-1, self.tp_kv_heads, self.head_dim) |
| 573 | q = layernorm_fn( |
| 574 | q, |
| 575 | self.query_layernorm.weight.data, |
| 576 | bias=None, |
| 577 | eps=self.rms_norm_eps, |
| 578 | is_rms_norm=True, |
| 579 | ) |
| 580 | k = layernorm_fn( |
| 581 | k, |
| 582 | self.key_layernorm.weight.data, |
| 583 | bias=None, |
| 584 | eps=self.rms_norm_eps, |
| 585 | is_rms_norm=True, |
| 586 | ) |
| 587 | q = q.reshape(-1, self.q_size_per_rank) |
| 588 | k = k.reshape(-1, self.kv_size_per_rank) |
| 589 | |
| 590 | # Apply rotary embeddings |
| 591 | if self.linear_rope: |
| 592 | q, k = self.rotary_emb(positions[:num_actual_tokens], q, k) |
| 593 | |
| 594 | # Reshape to [batch, heads, seq_len, head_dim] |
| 595 | q = q.view((qkv.shape[0], self.tp_heads, self.head_dim)) |
| 596 | k = k.view((qkv.shape[0], self.tp_kv_heads, self.head_dim)) |
| 597 | v = v.view((qkv.shape[0], self.tp_kv_heads, self.head_dim)) |
| 598 | |
| 599 | # Apply scaling if using minimax backend |
| 600 | if self.linear_scale: |
| 601 | q = q * self.scaling |
| 602 | |
| 603 | # Get KV cache and state indices |
| 604 | if attn_metadata is not None: |
| 605 | kv_cache = self.kv_cache[0] |
| 606 | state_indices_tensor = attn_metadata.state_indices_tensor |
| 607 | clear_linear_attention_cache_for_new_sequences( |
| 608 | kv_cache, state_indices_tensor, attn_metadata |
| 609 | ) |
| 610 | |
| 611 | # Compute attention |
| 612 | decode_only = getattr(attn_metadata, "num_prefills", 0) == 0 |
| 613 | if attn_metadata is None: |
| 614 | hidden = torch.empty( |
| 615 | (q.shape[0], q.shape[1] * q.shape[2]), device=q.device, dtype=q.dtype |
| 616 | ) |
| 617 | else: |
| 618 | if not decode_only: |
| 619 | hidden = self._prefill_and_mix_infer( |
| 620 | q, k, v, kv_cache, state_indices_tensor, attn_metadata |
| 621 | ) |
| 622 | else: |
| 623 | hidden = self._decode_infer( |
| 624 | q, k, v, kv_cache, state_indices_tensor, attn_metadata |
| 625 | ) |
| 626 | |
| 627 | # Apply group norm and gate (matching SGLang behavior) |
| 628 | gate, _ = self.g_proj(hidden_states[:num_actual_tokens]) |
| 629 | |
| 630 | if self.group_norm_size > 1: |
| 631 | hidden = self.g_norm(hidden, gate) |
| 632 | else: |
| 633 | hidden = self.g_norm(hidden) |
| 634 | hidden = F.sigmoid(gate) * hidden |
| 635 | |
| 636 | hidden = hidden.to(hidden_states.dtype) |
| 637 | |
| 638 | # Output projection |
| 639 | dense_out, _ = self.dense(hidden) |
| 640 | output[:num_actual_tokens] = dense_out |
| 641 | |
| 642 | def _prefill_and_mix_infer( |
| 643 | self, q, k, v, kv_cache, state_indices_tensor, attn_metadata |
| 644 | ): |
| 645 | """Handle prefill (mixed with decode if any).""" |
| 646 | return linear_attention_prefill_and_mix( |
| 647 | q=q, |
| 648 | k=k, |
| 649 | v=v, |
| 650 | kv_cache=kv_cache, |
| 651 | state_indices_tensor=state_indices_tensor, |
| 652 | attn_metadata=attn_metadata, |
| 653 | slope_rate=self.tp_slope, |
| 654 | block_size=self.BLOCK, |
| 655 | decode_fn=self._decode_infer, |
| 656 | prefix_fn=MiniMaxText01LinearKernel.jit_linear_forward_prefix, |
| 657 | layer_idx=self.layer_idx, |
| 658 | ) |
| 659 | |
| 660 | def _decode_infer(self, q, k, v, kv_cache, state_indices_tensor, attn_metadata): |
| 661 | """Handle decode (single token per sequence).""" |
| 662 | decode_state_indices = getattr(attn_metadata, "state_indices_tensor_d", None) |
| 663 | num_accepted_tokens = getattr(attn_metadata, "num_accepted_tokens", None) |
| 664 | query_start_loc = getattr(attn_metadata, "query_start_loc_d", None) |
| 665 | if ( |
| 666 | decode_state_indices is not None |
| 667 | and decode_state_indices.dim() > 1 |
| 668 | and num_accepted_tokens is not None |
| 669 | and query_start_loc is not None |
| 670 | ): |
| 671 | return bailing_linear_attention_decode_spec( |
| 672 | q, |
| 673 | k, |
| 674 | v, |
| 675 | kv_cache, |
| 676 | self.tp_slope, |
| 677 | decode_state_indices, |
| 678 | query_start_loc, |
| 679 | num_accepted_tokens, |
| 680 | q_start=0, |
| 681 | q_end=attn_metadata.num_decode_tokens, |
| 682 | slot_start=0, |
| 683 | slot_end=attn_metadata.num_decodes, |
| 684 | block_size=32, |
| 685 | ) |
| 686 | decode_state_indices = ( |
| 687 | state_indices_tensor |
| 688 | if decode_state_indices is None |
| 689 | else decode_state_indices |
| 690 | ) |
| 691 | hidden = linear_attention_decode( |
| 692 | q, |
| 693 | k, |
| 694 | v, |
| 695 | kv_cache, |
| 696 | self.tp_slope, |
| 697 | decode_state_indices, |
| 698 | q_start=0, |
| 699 | q_end=attn_metadata.num_decode_tokens, |
| 700 | slot_start=0, |
| 701 | slot_end=attn_metadata.num_decodes, |
| 702 | block_size=32, |
| 703 | ) |
| 704 | return hidden |