| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import time |
| 5 | from collections import defaultdict |
| 6 | from contextlib import contextmanager |
| 7 | from dataclasses import dataclass, field |
| 8 | from typing import Any |
| 9 | |
| 10 | import torch |
| 11 | |
| 12 | import vllm.envs as envs |
| 13 | from vllm.config import CUDAGraphMode, ParallelConfig, VllmConfig |
| 14 | from vllm.logger import init_logger |
| 15 | from vllm.platforms import current_platform |
| 16 | from vllm.v1.attention.backend import AttentionMetadata |
| 17 | from vllm.v1.worker.dp_utils import coordinate_batch_across_dp |
| 18 | from vllm.v1.worker.ubatch_utils import UBatchSlices |
| 19 | |
| 20 | logger = init_logger(__name__) |
| 21 | |
| 22 | track_batchsize: bool = envs.VLLM_LOG_BATCHSIZE_INTERVAL >= 0 |
| 23 | last_logging_time: float = 0 |
| 24 | forward_start_time: float = 0 |
| 25 | batchsize_logging_interval: float = envs.VLLM_LOG_BATCHSIZE_INTERVAL |
| 26 | batchsize_forward_time: defaultdict = defaultdict(list) |
| 27 | |
| 28 | |
| 29 | @dataclass(frozen=True) |
| 30 | class BatchDescriptor: |
| 31 | """Batch descriptor for cudagraph dispatching. We should keep the num of |
| 32 | items as minimal as possible to properly and uniquely describe the padded |
| 33 | batch for cudagraph. |
| 34 | """ |
| 35 | |
| 36 | num_tokens: int |
| 37 | num_reqs: int | None = None |
| 38 | """ |
| 39 | Number of requests in the batch. Can be None for PIECEWISE cudagraphs where |
| 40 | the cudagraphs can handle any number of requests. |
| 41 | """ |
| 42 | uniform: bool = False |
| 43 | """ |
| 44 | True if all the requests in the batch have the same number of tokens. |
| 45 | """ |
| 46 | has_lora: bool = False |
| 47 | """ |
| 48 | Whether this batch has active LoRA adapters. |
| 49 | """ |
| 50 | num_active_loras: int = 0 |
| 51 | """ |
| 52 | Number of distinct active LoRA adapters in this batch. |
| 53 | When cudagraph_specialize_lora_count is enabled, separate CUDA graphs |
| 54 | are captured for each num_active_loras value. This allows kernels |
| 55 | (like fused_moe_lora) whose grid size depends on num_active_loras |
| 56 | to be properly captured. |
| 57 | """ |
| 58 | |
| 59 | |
| 60 | def _compute_sp_num_tokens( |
| 61 | num_tokens_across_dp_cpu: torch.Tensor, |
| 62 | sequence_parallel_size: int, |
| 63 | pcp_size: int = 1, |
| 64 | use_ep: bool = True, |
| 65 | ) -> list[int]: |
| 66 | if pcp_size > 1: |
| 67 | num_tokens_across_dp_cpu = ( |
| 68 | num_tokens_across_dp_cpu.repeat_interleave(pcp_size) |
| 69 | if use_ep |
| 70 | else num_tokens_across_dp_cpu * pcp_size |
| 71 | ) |
| 72 | sp_tokens = ( |
| 73 | num_tokens_across_dp_cpu + sequence_parallel_size - 1 |
| 74 | ) // sequence_parallel_size |
| 75 | |
| 76 | sp_tokens = sp_tokens.repeat_interleave(sequence_parallel_size) |
| 77 | return sp_tokens.tolist() |
| 78 | |
| 79 | |
| 80 | @dataclass |
| 81 | class DPMetadata: |
| 82 | num_tokens_across_dp_cpu: torch.Tensor |
| 83 | |
| 84 | # NOTE: local_sizes should only be set by the chunked_sizes context manager |
| 85 | local_sizes: list[int] | None = None |
| 86 | |
| 87 | @staticmethod |
| 88 | def make( |
| 89 | parallel_config: ParallelConfig, |
| 90 | num_tokens: int, |
| 91 | num_tokens_across_dp_cpu: torch.Tensor, |
| 92 | ) -> "DPMetadata": |
| 93 | assert num_tokens_across_dp_cpu is not None |
| 94 | assert ( |
| 95 | parallel_config.data_parallel_size > 1 |
| 96 | or parallel_config.use_sequence_parallel_moe |
| 97 | ) |
| 98 | assert parallel_config.is_moe_model is not False |
| 99 | dp_rank = parallel_config.data_parallel_rank |
| 100 | batchsize = num_tokens |
| 101 | |
| 102 | # If num_tokens_across_dp is None, it will be computed by all_reduce |
| 103 | # Otherwise, num_tokens_across_dp[dp_rank] should be equal to batchsize |
| 104 | assert num_tokens_across_dp_cpu[dp_rank] == batchsize, ( |
| 105 | f"{num_tokens_across_dp_cpu[dp_rank]} {batchsize}" |
| 106 | ) |
| 107 | return DPMetadata(num_tokens_across_dp_cpu) |
| 108 | |
| 109 | @contextmanager |
| 110 | def sp_local_sizes( |
| 111 | self, sequence_parallel_size: int, pcp_size: int = 1, use_ep: bool = False |
| 112 | ): |
| 113 | """Context manager for setting self.local_sizes. Same as self.chunked_sizes |
| 114 | but without any chunking. |
| 115 | """ |
| 116 | self.local_sizes = _compute_sp_num_tokens( |
| 117 | self.num_tokens_across_dp_cpu, sequence_parallel_size, pcp_size, use_ep |
| 118 | ) |
| 119 | try: |
| 120 | yield self.local_sizes |
| 121 | finally: |
| 122 | self.local_sizes = None |
| 123 | |
| 124 | def get_chunk_sizes_across_dp_rank(self) -> list[int] | None: |
| 125 | assert self.local_sizes is not None |
| 126 | return self.local_sizes |
| 127 | |
| 128 | # Get the cumulative tokens across sequence parallel ranks. |
| 129 | # In this case the input to the MoEs will be distributed w.r.t both |
| 130 | # DP and TP rank. |
| 131 | # When sp_size==1, this is just the cumulative num tokens across DP. |
| 132 | def cu_tokens_across_sp(self, sp_size: int) -> torch.Tensor: |
| 133 | num_tokens_across_sp_cpu = ( |
| 134 | self.num_tokens_across_dp_cpu - 1 + sp_size |
| 135 | ) // sp_size |
| 136 | num_tokens_across_sp_cpu = num_tokens_across_sp_cpu.repeat_interleave(sp_size) |
| 137 | return torch.cumsum(num_tokens_across_sp_cpu, dim=0) |
| 138 | |
| 139 | |
| 140 | @dataclass |
| 141 | class ForwardContext: |
| 142 | # copy from vllm_config.compilation_config.static_forward_context |
| 143 | no_compile_layers: dict[str, Any] |
| 144 | attn_metadata: dict[str, AttentionMetadata] | list[dict[str, AttentionMetadata]] |
| 145 | slot_mapping: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] |
| 146 | """ |
| 147 | Type Dict[str, AttentionMetadata] for v1, map from layer_name of each |
| 148 | attention layer to its attention metadata |
| 149 | Type List[Dict[str, AttentionMetadata]] for DBO. List of size two, one |
| 150 | for each microbatch. |
| 151 | Set dynamically for each forward pass |
| 152 | """ |
| 153 | # set dynamically for each forward pass |
| 154 | dp_metadata: DPMetadata | None = None |
| 155 | # determine the cudagraph style at runtime to be FULL, PIECEWISE, or NONE. |
| 156 | # by default NONE, no cudagraph is used. |
| 157 | cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE |
| 158 | batch_descriptor: BatchDescriptor | None = None |
| 159 | |
| 160 | ubatch_slices: UBatchSlices | None = None |
| 161 | |
| 162 | # Boolean mask over the token axis: True for padding rows that are not real |
| 163 | # tokens. Consumers can use it to skip work for padded tokens. None when |
| 164 | # the producer does not set it. |
| 165 | is_padding: torch.Tensor | None = None |
| 166 | |
| 167 | # If True, bypass the compiled model call, e.g. by using .forward() directly |
| 168 | skip_compiled: bool = False |
| 169 | |
| 170 | # For torch.compile cold start times, we need to avoid hard-coding |
| 171 | # any strings into the graph. Right now, the vllm.moe_forward |
| 172 | # and vllm.moe_forward_shared custom operators hard-code strings into |
| 173 | # the graph. |
| 174 | # |
| 175 | # The workaround is to store a list of the strings that each of those |
| 176 | # custom ops needs in the ForwardContext (all_moe_layers) |
| 177 | # as well as a counter (moe_layer_index). |
| 178 | # The ForwardContext object is alive for the duration of the forward pass. |
| 179 | # When the custom op needs a layer string, get the next string |
| 180 | # from all_moe_layers and increment the counter. |
| 181 | # |
| 182 | # This assumes that the custom operators will always be executed in |
| 183 | # order and that torch.compile will not try to reorder these |
| 184 | # operations with respect to each other. |
| 185 | # |
| 186 | # TODO(https://github.com/vllm-project/vllm/issues/31985): |
| 187 | # There are longer-term solutions, like unwrapping the moe custom operator, |
| 188 | # that aren't ready yet. |
| 189 | # We could also treat the string as a "symbolic input" to the graph but |
| 190 | # the PyTorch-side bits for that aren't ready yet either. |
| 191 | # |
| 192 | # If this value is None (like in some tests), then we end up baking the string |
| 193 | # into the graph. Otherwise, the moe custom ops will pop a string from this list. |
| 194 | all_moe_layers: list[str] | None = None |
| 195 | moe_layer_index: int = 0 |
| 196 | |
| 197 | additional_kwargs: dict[str, Any] = field(default_factory=dict) |
| 198 | |
| 199 | def __post_init__(self): |
| 200 | assert self.cudagraph_runtime_mode.is_valid_runtime_mode(), ( |
| 201 | f"Invalid cudagraph runtime mode: {self.cudagraph_runtime_mode}" |
| 202 | ) |
| 203 | |
| 204 | |
| 205 | _forward_context: ForwardContext | None = None |
| 206 | |
| 207 | |
| 208 | def get_forward_context() -> ForwardContext: |
| 209 | """Get the current forward context.""" |
| 210 | assert _forward_context is not None, ( |
| 211 | "Forward context is not set. " |
| 212 | "Please use `set_forward_context` to set the forward context." |
| 213 | ) |
| 214 | return _forward_context |
| 215 | |
| 216 | |
| 217 | def is_forward_context_available() -> bool: |
| 218 | return _forward_context is not None |
| 219 | |
| 220 | |
| 221 | def in_piecewise_cudagraph() -> bool: |
| 222 | """Whether the current forward runs in piecewise cudagraph mode (graph |
| 223 | segments separated by eager breaks), at capture or replay time.""" |
| 224 | return ( |
| 225 | is_forward_context_available() |
| 226 | and get_forward_context().cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE |
| 227 | ) |
| 228 | |
| 229 | |
| 230 | def create_forward_context( |
| 231 | attn_metadata: Any, |
| 232 | vllm_config: VllmConfig, |
| 233 | dp_metadata: DPMetadata | None = None, |
| 234 | cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE, |
| 235 | batch_descriptor: BatchDescriptor | None = None, |
| 236 | ubatch_slices: UBatchSlices | None = None, |
| 237 | slot_mapping: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None = None, |
| 238 | additional_kwargs: dict[str, Any] | None = None, |
| 239 | skip_compiled: bool = False, |
| 240 | is_padding: torch.Tensor | None = None, |
| 241 | ): |
| 242 | if vllm_config.compilation_config.fast_moe_cold_start: |
| 243 | all_moe_layers = vllm_config.compilation_config.static_all_moe_layers |
| 244 | else: |
| 245 | all_moe_layers = None |
| 246 | |
| 247 | return ForwardContext( |
| 248 | no_compile_layers=vllm_config.compilation_config.static_forward_context, |
| 249 | all_moe_layers=all_moe_layers, |
| 250 | attn_metadata=attn_metadata, |
| 251 | slot_mapping=slot_mapping or {}, |
| 252 | dp_metadata=dp_metadata, |
| 253 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 254 | batch_descriptor=batch_descriptor, |
| 255 | ubatch_slices=ubatch_slices, |
| 256 | skip_compiled=skip_compiled, |
| 257 | additional_kwargs=additional_kwargs or {}, |
| 258 | is_padding=is_padding, |
| 259 | ) |
| 260 | |
| 261 | |
| 262 | @contextmanager |
| 263 | def override_forward_context(forward_context: ForwardContext | None): |
| 264 | """A context manager that overrides the current forward context. |
| 265 | This is used to override the forward context for a specific |
| 266 | forward pass. |
| 267 | """ |
| 268 | global _forward_context |
| 269 | prev_context = _forward_context |
| 270 | _forward_context = forward_context |
| 271 | try: |
| 272 | yield |
| 273 | finally: |
| 274 | _forward_context = prev_context |
| 275 | |
| 276 | |
| 277 | @contextmanager |
| 278 | def set_forward_context( |
| 279 | attn_metadata: Any, |
| 280 | vllm_config: VllmConfig, |
| 281 | num_tokens: int | None = None, |
| 282 | num_tokens_across_dp: torch.Tensor | None = None, |
| 283 | cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE, |
| 284 | batch_descriptor: BatchDescriptor | None = None, |
| 285 | ubatch_slices: UBatchSlices | None = None, |
| 286 | slot_mapping: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None = None, |
| 287 | skip_compiled: bool = False, |
| 288 | is_padding: torch.Tensor | None = None, |
| 289 | ): |
| 290 | """A context manager that stores the current forward context, |
| 291 | can be attention metadata, etc. |
| 292 | Here we can inject common logic for every model forward pass. |
| 293 | """ |
| 294 | global forward_start_time |
| 295 | need_to_track_batchsize = track_batchsize and attn_metadata is not None |
| 296 | if need_to_track_batchsize: |
| 297 | forward_start_time = time.perf_counter() |
| 298 | |
| 299 | dp_metadata: DPMetadata | None = None |
| 300 | if ( |
| 301 | ( |
| 302 | vllm_config.parallel_config.data_parallel_size > 1 |
| 303 | or vllm_config.parallel_config.use_sequence_parallel_moe |
| 304 | ) |
| 305 | and vllm_config.parallel_config.is_moe_model is not False |
| 306 | and (attn_metadata is not None or num_tokens is not None) |
| 307 | ): |
| 308 | # If num_tokens_across_dp hasn't already been initialized, then |
| 309 | # initialize it here. Both DP padding and Microbatching will be |
| 310 | # disabled. |
| 311 | if ( |
| 312 | num_tokens_across_dp is None |
| 313 | and vllm_config.parallel_config.data_parallel_size > 1 |
| 314 | ): |
| 315 | assert ubatch_slices is None |
| 316 | assert num_tokens is not None |
| 317 | _, num_tokens_across_dp, _ = coordinate_batch_across_dp( |
| 318 | num_tokens_unpadded=num_tokens, |
| 319 | parallel_config=vllm_config.parallel_config, |
| 320 | allow_microbatching=False, |
| 321 | ) |
| 322 | assert num_tokens_across_dp is not None |
| 323 | elif num_tokens_across_dp is None: |
| 324 | assert num_tokens is not None |
| 325 | num_tokens_across_dp = torch.tensor([num_tokens], dtype=torch.int32) |
| 326 | dp_metadata = DPMetadata.make( |
| 327 | vllm_config.parallel_config, num_tokens or 0, num_tokens_across_dp |
| 328 | ) |
| 329 | |
| 330 | # Convenience: if cudagraph is used and num_tokens is given, we can just |
| 331 | # create a batch descriptor here if not given (there's no harm since if it |
| 332 | # doesn't match in the wrapper it'll fall through). |
| 333 | if cudagraph_runtime_mode != CUDAGraphMode.NONE and num_tokens is not None: |
| 334 | batch_descriptor = batch_descriptor or BatchDescriptor(num_tokens=num_tokens) |
| 335 | |
| 336 | additional_kwargs = current_platform.set_additional_forward_context( |
| 337 | attn_metadata=attn_metadata, |
| 338 | vllm_config=vllm_config, |
| 339 | dp_metadata=dp_metadata, |
| 340 | num_tokens=num_tokens, |
| 341 | num_tokens_across_dp=num_tokens_across_dp, |
| 342 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 343 | batch_descriptor=batch_descriptor, |
| 344 | ubatch_slices=ubatch_slices, |
| 345 | ) |
| 346 | |
| 347 | forward_context = create_forward_context( |
| 348 | attn_metadata, |
| 349 | vllm_config, |
| 350 | dp_metadata, |
| 351 | cudagraph_runtime_mode, |
| 352 | batch_descriptor, |
| 353 | ubatch_slices, |
| 354 | slot_mapping, |
| 355 | additional_kwargs, |
| 356 | skip_compiled, |
| 357 | is_padding=is_padding, |
| 358 | ) |
| 359 | |
| 360 | try: |
| 361 | with override_forward_context(forward_context): |
| 362 | yield |
| 363 | finally: |
| 364 | global last_logging_time, batchsize_logging_interval |
| 365 | if need_to_track_batchsize: |
| 366 | batchsize = num_tokens |
| 367 | # we use synchronous scheduling right now, |
| 368 | # adding a sync point here should not affect |
| 369 | # scheduling of the next batch |
| 370 | synchronize = current_platform.synchronize |
| 371 | if synchronize is not None: |
| 372 | synchronize() |
| 373 | now = time.perf_counter() |
| 374 | # time measurement is in milliseconds |
| 375 | batchsize_forward_time[batchsize].append((now - forward_start_time) * 1000) |
| 376 | if now - last_logging_time > batchsize_logging_interval: |
| 377 | last_logging_time = now |
| 378 | forward_stats = [] |
| 379 | for bs, times in batchsize_forward_time.items(): |
| 380 | if len(times) <= 1: |
| 381 | # can be cudagraph / profiling run |
| 382 | continue |
| 383 | medium = torch.quantile(torch.tensor(times), q=0.5).item() |
| 384 | medium = round(medium, 2) |
| 385 | forward_stats.append((bs, len(times), medium)) |
| 386 | forward_stats.sort(key=lambda x: x[1], reverse=True) |
| 387 | if forward_stats: |
| 388 | logger.info( |
| 389 | ( |
| 390 | "Batchsize forward time stats " |
| 391 | "(batchsize, count, median_time(ms)): %s" |
| 392 | ), |
| 393 | forward_stats, |
| 394 | ) |