| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import copy |
| 5 | import getpass |
| 6 | import json |
| 7 | import os |
| 8 | import tempfile |
| 9 | import threading |
| 10 | import time |
| 11 | from collections.abc import Iterable |
| 12 | from contextlib import contextmanager |
| 13 | from dataclasses import is_dataclass |
| 14 | from datetime import datetime |
| 15 | from enum import IntEnum |
| 16 | from functools import lru_cache |
| 17 | from pathlib import Path |
| 18 | from typing import TYPE_CHECKING, Any, Literal, TypeVar, get_args |
| 19 | |
| 20 | import torch |
| 21 | from pydantic import ConfigDict, Field, model_validator |
| 22 | |
| 23 | import vllm.envs as envs |
| 24 | from vllm.logger import enable_trace_function_call, init_logger |
| 25 | from vllm.transformers_utils.runai_utils import is_runai_obj_uri |
| 26 | from vllm.triton_utils import HAS_TRITON |
| 27 | from vllm.utils import random_uuid |
| 28 | from vllm.utils.hashing import safe_hash |
| 29 | |
| 30 | from .attention import AttentionConfig, HiSparseConfig |
| 31 | from .aux_output import AuxOutputConfig |
| 32 | from .cache import CacheConfig |
| 33 | from .compilation import CompilationConfig, CompilationMode, CUDAGraphMode |
| 34 | from .device import DeviceConfig |
| 35 | from .diffusion import DiffusionConfig |
| 36 | from .ec_manager_config import EncoderCacheManagerConfig |
| 37 | from .ec_transfer import ECTransferConfig |
| 38 | from .engram import EngramConfig, model_has_engram_layers |
| 39 | from .kernel import KernelConfig |
| 40 | from .kv_events import KVEventsConfig |
| 41 | from .kv_transfer import KVTransferConfig |
| 42 | from .load import LoadConfig |
| 43 | from .logging import LoggingConfig |
| 44 | from .lora import LoRAConfig |
| 45 | from .mamba import MambaBackendEnum, MambaConfig |
| 46 | from .model import ModelConfig |
| 47 | from .observability import ObservabilityConfig |
| 48 | from .offload import OffloadConfig |
| 49 | from .parallel import ParallelConfig |
| 50 | from .profiler import ProfilerConfig |
| 51 | from .reasoning import ReasoningConfig |
| 52 | from .scheduler import SchedulerConfig |
| 53 | from .speculative import EagleModelTypes, NgramGPUTypes, SpeculativeConfig |
| 54 | from .structured_outputs import StructuredOutputsConfig |
| 55 | from .utils import SupportsHash, config, get_field, replace |
| 56 | from .watermarking import WatermarkConfig |
| 57 | from .weight_transfer import WeightTransferConfig |
| 58 | |
| 59 | if TYPE_CHECKING: |
| 60 | from transformers import PreTrainedConfig |
| 61 | |
| 62 | from vllm.model_executor.layers.quantization.base_config import QuantizationConfig |
| 63 | from vllm.v1.kv_cache_interface import KVCacheConfig |
| 64 | else: |
| 65 | PreTrainedConfig = Any |
| 66 | |
| 67 | QuantizationConfig = Any |
| 68 | |
| 69 | KVCacheConfig = Any |
| 70 | |
| 71 | logger = init_logger(__name__) |
| 72 | |
| 73 | # TODO(rocm): These models are either unsupported by MRV2 or slower with |
| 74 | # MRV2 on AMD GPUs. |
| 75 | ROCM_DEFAULT_MRV1_ARCHITECTURES = frozenset( |
| 76 | {"DeepseekV32ForCausalLM", "DeepseekV4ForCausalLM", "GlmMoeDsaForCausalLM"} |
| 77 | ) |
| 78 | |
| 79 | DEFAULT_BREAKABLE_CUDAGRAPH_ARCHITECTURES = frozenset( |
| 80 | { |
| 81 | "DeepseekV32MTPModel", |
| 82 | "DeepseekV32ForCausalLM", |
| 83 | "DeepseekV4ForCausalLM", |
| 84 | "DeepseekV4ForConditionalGeneration", |
| 85 | "DeepSeekV4MTPModel", |
| 86 | "DeepseekV41ForCausalLM", |
| 87 | "Dots3NoteForCausalLM", |
| 88 | "Dots3NoteMTPModel", |
| 89 | "Glm5NextForCausalLM", |
| 90 | "Glm5NextForConditionalGeneration", |
| 91 | "Glm5NextMTPModel", |
| 92 | "GlmMoeDsaForCausalLM", |
| 93 | "HYV4ForCausalLM", |
| 94 | "HYV4MTPModel", |
| 95 | "InklingForCausalLM", |
| 96 | "InklingForConditionalGeneration", |
| 97 | "KimiK3ForConditionalGeneration", |
| 98 | "KimiK3MTPModel", |
| 99 | "KimiLinearForCausalLM", |
| 100 | "MiniMaxM3SparseForCausalLM", |
| 101 | "MiniMaxM3SparseForConditionalGeneration", |
| 102 | "Qwen4ExpForCausalLM", |
| 103 | "Qwen4ExpForConditionalGeneration", |
| 104 | "Qwen4ExpMTP", |
| 105 | } |
| 106 | ) |
| 107 | |
| 108 | |
| 109 | @lru_cache |
| 110 | def default_breakable_cudagraph_architectures() -> frozenset[str]: |
| 111 | """Architectures defaulting to breakable CUDA graphs on this platform.""" |
| 112 | from vllm.platforms import current_platform |
| 113 | |
| 114 | if current_platform.is_rocm(): |
| 115 | # Breakable CUDA graphs currently regress performance on ROCm for |
| 116 | # models that can use torch.compile piecewise graphs instead. Do not |
| 117 | # opt those in by default. Users can still force them with |
| 118 | # VLLM_USE_BREAKABLE_CUDAGRAPH=1. |
| 119 | # |
| 120 | # DeepseekV41ForCausalLM cannot torch.compile, and the ROCm sparse |
| 121 | # SWA backend only reports AttentionCGSupport.UNIFORM_BATCH. Default |
| 122 | # FULL_AND_PIECEWISE then dies at capture unless breakable CUDA |
| 123 | # graphs are on. Enable this architecture so the published AMD |
| 124 | # recipe can start. |
| 125 | return frozenset({"DeepseekV41ForCausalLM"}) |
| 126 | return DEFAULT_BREAKABLE_CUDAGRAPH_ARCHITECTURES |
| 127 | |
| 128 | |
| 129 | class OptimizationLevel(IntEnum): |
| 130 | """Optimization level enum.""" |
| 131 | |
| 132 | O0 = 0 |
| 133 | """O0 : No optimization. no compilation, no cudagraphs, no other |
| 134 | optimization, just starting up immediately""" |
| 135 | O1 = 1 |
| 136 | """O1: Quick optimizations. Dynamo+Inductor compilation and Piecewise |
| 137 | cudagraphs""" |
| 138 | O2 = 2 |
| 139 | """O2: Full optimizations. -O1 as well as Full and Piecewise cudagraphs.""" |
| 140 | O3 = 3 |
| 141 | """O3: Currently the same as -O2s.""" |
| 142 | |
| 143 | |
| 144 | PerformanceMode = Literal["balanced", "interactivity", "throughput"] |
| 145 | |
| 146 | IS_QUANTIZED = False |
| 147 | IS_DENSE = False |
| 148 | # The optimizations that depend on these properties currently set to False |
| 149 | # in all cases. |
| 150 | # if model_config is not None: |
| 151 | # IS_QUANTIZED = lambda c: c.model_config.is_quantized() |
| 152 | # IS_DENSE = lambda c: not c.model_config.is_model_moe() |
| 153 | # See https://github.com/vllm-project/vllm/issues/25689. |
| 154 | |
| 155 | |
| 156 | def enable_norm_fusion(cfg: "VllmConfig") -> bool: |
| 157 | """Enable if either RMS norm or quant FP8 custom op is active; |
| 158 | otherwise Inductor handles fusion.""" |
| 159 | return ( |
| 160 | cfg.compilation_config.is_custom_op_enabled("rms_norm") |
| 161 | or cfg.compilation_config.is_custom_op_enabled("quant_fp8") |
| 162 | or cfg.kernel_config.ir_op_priority.rms_norm[0] != "native" |
| 163 | ) |
| 164 | |
| 165 | |
| 166 | def enable_act_fusion(cfg: "VllmConfig") -> bool: |
| 167 | """Enable if either SiLU+Mul or quant FP8 custom op is active; |
| 168 | otherwise Inductor handles fusion. |
| 169 | Also enable for FP4 models as FP4 quant is always custom so Inductor cannot fuse it. |
| 170 | """ |
| 171 | return ( |
| 172 | cfg.compilation_config.is_custom_op_enabled("silu_and_mul") |
| 173 | or cfg.compilation_config.is_custom_op_enabled("quant_fp8") |
| 174 | or (cfg.model_config is not None and cfg.model_config.is_nvfp4_quantized()) |
| 175 | ) |
| 176 | |
| 177 | |
| 178 | def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool: |
| 179 | """Enable if TP > 1 and Hopper/Blackwell and flashinfer installed.""" |
| 180 | from vllm.platforms import current_platform |
| 181 | from vllm.utils.flashinfer import has_flashinfer |
| 182 | |
| 183 | # The fused all-reduce + RMSNorm path is not batch-invariant |
| 184 | if envs.VLLM_BATCH_INVARIANT: |
| 185 | return False |
| 186 | |
| 187 | if current_platform.is_rocm(): |
| 188 | from vllm._aiter_ops import rocm_aiter_ops |
| 189 | |
| 190 | return ( |
| 191 | rocm_aiter_ops.is_enabled() and cfg.parallel_config.tensor_parallel_size > 1 |
| 192 | ) |
| 193 | |
| 194 | return ( |
| 195 | cfg.parallel_config.tensor_parallel_size > 1 |
| 196 | and current_platform.is_cuda() |
| 197 | and has_flashinfer() |
| 198 | and ( |
| 199 | current_platform.is_device_capability_family(100) |
| 200 | or current_platform.is_device_capability(90) |
| 201 | ) |
| 202 | ) |
| 203 | |
| 204 | |
| 205 | def enable_rope_kvcache_fusion(cfg: "VllmConfig") -> bool: |
| 206 | """Enable if rotary embedding custom op is active and |
| 207 | use_inductor_graph_partition is enabled. |
| 208 | """ |
| 209 | from vllm._aiter_ops import rocm_aiter_ops |
| 210 | |
| 211 | return ( |
| 212 | rocm_aiter_ops.is_enabled() |
| 213 | and cfg.compilation_config.is_custom_op_enabled("rotary_embedding") |
| 214 | and ( |
| 215 | cfg.compilation_config.use_inductor_graph_partition |
| 216 | or not cfg.compilation_config.splitting_ops_contain_kv_cache_update() |
| 217 | ) |
| 218 | ) |
| 219 | |
| 220 | |
| 221 | def enable_rope_kvcache_mla_fusion(cfg: "VllmConfig") -> bool: |
| 222 | """Enable if use_inductor_graph_partition is enabled.""" |
| 223 | return ( |
| 224 | cfg.compilation_config.use_inductor_graph_partition |
| 225 | or not cfg.compilation_config.splitting_ops_contain_kv_cache_update() |
| 226 | ) |
| 227 | |
| 228 | |
| 229 | def enable_norm_pad_fusion(cfg: "VllmConfig") -> bool: |
| 230 | """Enable if using AITER RMSNorm and hidden size is 2880 i.e. gpt-oss.""" |
| 231 | return ( |
| 232 | cfg.kernel_config.ir_op_priority.fused_add_rms_norm[0] == "aiter" |
| 233 | and cfg.model_config is not None |
| 234 | and cfg.model_config.get_hidden_size() == 2880 |
| 235 | ) |
| 236 | |
| 237 | |
| 238 | def enable_mla_dual_rms_norm_fusion(cfg: "VllmConfig") -> bool: |
| 239 | """Enable MLA dual RMS norm fusion on ROCm with AITER.""" |
| 240 | from vllm._aiter_ops import rocm_aiter_ops |
| 241 | |
| 242 | return rocm_aiter_ops.is_enabled() |
| 243 | |
| 244 | |
| 245 | def enable_qk_norm_rope_kvcache(cfg: "VllmConfig") -> bool: |
| 246 | """Enable fused QK-norm + RoPE/MRoPE + KV cache update with AITER.""" |
| 247 | from vllm._aiter_ops import rocm_aiter_ops |
| 248 | |
| 249 | if not rocm_aiter_ops.is_enabled(): |
| 250 | return False |
| 251 | return cfg.compilation_config.is_custom_op_enabled("rotary_embedding") |
| 252 | |
| 253 | |
| 254 | OPTIMIZATION_LEVEL_00 = { |
| 255 | "compilation_config": { |
| 256 | "pass_config": { |
| 257 | "fuse_norm_quant": False, |
| 258 | "fuse_act_quant": False, |
| 259 | "fuse_allreduce_rms": False, |
| 260 | "fuse_attn_quant": False, |
| 261 | "enable_sp": False, |
| 262 | "fuse_gemm_comms": False, |
| 263 | "fuse_act_padding": False, |
| 264 | "fuse_mla_dual_rms_norm": False, |
| 265 | "fuse_rope_kvcache": False, |
| 266 | "fuse_qk_norm_rope_kvcache": False, |
| 267 | "enable_qk_norm_rope_fusion": False, |
| 268 | "fuse_rope_kvcache_cat_mla": False, |
| 269 | }, |
| 270 | "cudagraph_mode": CUDAGraphMode.NONE, |
| 271 | "use_inductor_graph_partition": False, |
| 272 | }, |
| 273 | "kernel_config": { |
| 274 | "enable_flashinfer_autotune": False, |
| 275 | }, |
| 276 | } |
| 277 | OPTIMIZATION_LEVEL_01 = { |
| 278 | "compilation_config": { |
| 279 | "pass_config": { |
| 280 | "fuse_norm_quant": enable_norm_fusion, |
| 281 | "fuse_act_quant": enable_act_fusion, |
| 282 | "fuse_allreduce_rms": False, |
| 283 | "fuse_attn_quant": False, |
| 284 | "enable_sp": False, |
| 285 | "fuse_gemm_comms": False, |
| 286 | "fuse_act_padding": enable_norm_pad_fusion, |
| 287 | "fuse_mla_dual_rms_norm": enable_mla_dual_rms_norm_fusion, |
| 288 | "fuse_rope_kvcache": False, |
| 289 | "fuse_qk_norm_rope_kvcache": False, |
| 290 | "enable_qk_norm_rope_fusion": False, |
| 291 | "fuse_rope_kvcache_cat_mla": False, |
| 292 | }, |
| 293 | "cudagraph_mode": CUDAGraphMode.PIECEWISE, |
| 294 | "use_inductor_graph_partition": False, |
| 295 | }, |
| 296 | "kernel_config": { |
| 297 | "enable_flashinfer_autotune": True, |
| 298 | }, |
| 299 | } |
| 300 | OPTIMIZATION_LEVEL_02 = { |
| 301 | "compilation_config": { |
| 302 | "pass_config": { |
| 303 | "fuse_norm_quant": enable_norm_fusion, |
| 304 | "fuse_act_quant": enable_act_fusion, |
| 305 | "fuse_allreduce_rms": enable_allreduce_rms_fusion, |
| 306 | "fuse_attn_quant": IS_QUANTIZED, |
| 307 | "enable_sp": IS_DENSE, |
| 308 | "fuse_gemm_comms": IS_DENSE, |
| 309 | "fuse_act_padding": enable_norm_pad_fusion, |
| 310 | "fuse_mla_dual_rms_norm": enable_mla_dual_rms_norm_fusion, |
| 311 | "fuse_rope_kvcache": enable_rope_kvcache_fusion, |
| 312 | "fuse_qk_norm_rope_kvcache": enable_qk_norm_rope_kvcache, |
| 313 | "enable_qk_norm_rope_fusion": False, |
| 314 | "fuse_rope_kvcache_cat_mla": enable_rope_kvcache_mla_fusion, |
| 315 | }, |
| 316 | "cudagraph_mode": CUDAGraphMode.FULL_AND_PIECEWISE, |
| 317 | "use_inductor_graph_partition": False, |
| 318 | }, |
| 319 | "kernel_config": { |
| 320 | "enable_flashinfer_autotune": True, |
| 321 | }, |
| 322 | } |
| 323 | OPTIMIZATION_LEVEL_03 = { |
| 324 | "compilation_config": { |
| 325 | "pass_config": { |
| 326 | "fuse_norm_quant": enable_norm_fusion, |
| 327 | "fuse_act_quant": enable_act_fusion, |
| 328 | "fuse_allreduce_rms": enable_allreduce_rms_fusion, |
| 329 | "fuse_attn_quant": IS_QUANTIZED, |
| 330 | "enable_sp": IS_DENSE, |
| 331 | "fuse_gemm_comms": IS_DENSE, |
| 332 | "fuse_act_padding": enable_norm_pad_fusion, |
| 333 | "fuse_mla_dual_rms_norm": enable_mla_dual_rms_norm_fusion, |
| 334 | "fuse_rope_kvcache": enable_rope_kvcache_fusion, |
| 335 | "fuse_qk_norm_rope_kvcache": enable_qk_norm_rope_kvcache, |
| 336 | "enable_qk_norm_rope_fusion": False, |
| 337 | "fuse_rope_kvcache_cat_mla": enable_rope_kvcache_mla_fusion, |
| 338 | }, |
| 339 | "cudagraph_mode": CUDAGraphMode.FULL_AND_PIECEWISE, |
| 340 | "use_inductor_graph_partition": False, |
| 341 | }, |
| 342 | "kernel_config": { |
| 343 | "enable_flashinfer_autotune": True, |
| 344 | }, |
| 345 | } |
| 346 | |
| 347 | OPTIMIZATION_LEVEL_TO_CONFIG = { |
| 348 | OptimizationLevel.O0: OPTIMIZATION_LEVEL_00, |
| 349 | OptimizationLevel.O1: OPTIMIZATION_LEVEL_01, |
| 350 | OptimizationLevel.O2: OPTIMIZATION_LEVEL_02, |
| 351 | OptimizationLevel.O3: OPTIMIZATION_LEVEL_03, |
| 352 | } |
| 353 | |
| 354 | |
| 355 | @config(config=ConfigDict(arbitrary_types_allowed=True)) |
| 356 | class VllmConfig: |
| 357 | """Dataclass which contains all vllm-related configuration. This |
| 358 | simplifies passing around the distinct configurations in the codebase. |
| 359 | """ |
| 360 | |
| 361 | # TODO: use default_factory once default constructing ModelConfig doesn't |
| 362 | # try to download a model |
| 363 | model_config: ModelConfig = None # type: ignore[assignment] |
| 364 | """Model configuration.""" |
| 365 | cache_config: CacheConfig = Field(default_factory=CacheConfig) |
| 366 | """Cache configuration.""" |
| 367 | parallel_config: ParallelConfig = Field(default_factory=ParallelConfig) |
| 368 | """Parallel configuration.""" |
| 369 | scheduler_config: SchedulerConfig = Field( |
| 370 | default_factory=SchedulerConfig.default_factory, |
| 371 | ) |
| 372 | """Scheduler configuration.""" |
| 373 | device_config: DeviceConfig = Field(default_factory=DeviceConfig) |
| 374 | """Device configuration.""" |
| 375 | load_config: LoadConfig = Field(default_factory=LoadConfig) |
| 376 | """Load configuration.""" |
| 377 | offload_config: OffloadConfig = Field(default_factory=OffloadConfig) |
| 378 | """Model weight offloading configuration.""" |
| 379 | attention_config: AttentionConfig = Field(default_factory=AttentionConfig) |
| 380 | """Attention configuration.""" |
| 381 | aux_output_config: AuxOutputConfig = Field(default_factory=AuxOutputConfig) |
| 382 | """Execution auxiliary output configuration.""" |
| 383 | engram_config: EngramConfig | None = None |
| 384 | """N-gram embedding storage and sharding settings.""" |
| 385 | mamba_config: MambaConfig = Field(default_factory=MambaConfig) |
| 386 | """Mamba configuration.""" |
| 387 | kernel_config: KernelConfig = Field(default_factory=KernelConfig) |
| 388 | """Kernel configuration.""" |
| 389 | lora_config: LoRAConfig | None = None |
| 390 | """LoRA configuration.""" |
| 391 | speculative_config: SpeculativeConfig | None = None |
| 392 | """Speculative decoding configuration.""" |
| 393 | watermark_config: WatermarkConfig | None = None |
| 394 | """Text watermarking configuration.""" |
| 395 | diffusion_config: DiffusionConfig | None = None |
| 396 | """Diffusion LLM (dLLM) configuration.""" |
| 397 | |
| 398 | structured_outputs_config: StructuredOutputsConfig = Field( |
| 399 | default_factory=StructuredOutputsConfig |
| 400 | ) |
| 401 | """Structured outputs configuration.""" |
| 402 | observability_config: ObservabilityConfig = Field( |
| 403 | default_factory=ObservabilityConfig |
| 404 | ) |
| 405 | """Observability configuration.""" |
| 406 | logging_config: LoggingConfig = Field(default_factory=LoggingConfig) |
| 407 | """Logging configuration.""" |
| 408 | quant_config: QuantizationConfig | None = None |
| 409 | """Quantization configuration.""" |
| 410 | compilation_config: CompilationConfig = Field(default_factory=CompilationConfig) |
| 411 | """`torch.compile` and cudagraph capture configuration for the model. |
| 412 | |
| 413 | As a shorthand, one can append compilation arguments via |
| 414 | -cc.parameter=argument such as `-cc.mode=3` (same as `-cc='{"mode":3}'`). |
| 415 | |
| 416 | You can specify the full compilation config like so: |
| 417 | `{"mode": 3, "cudagraph_capture_sizes": [1, 2, 4, 8]}` |
| 418 | """ |
| 419 | profiler_config: ProfilerConfig = Field(default_factory=ProfilerConfig) |
| 420 | """Profiling configuration.""" |
| 421 | kv_transfer_config: KVTransferConfig | None = None |
| 422 | """The configurations for distributed KV cache transfer.""" |
| 423 | kv_events_config: KVEventsConfig | None = None |
| 424 | """The configurations for event publishing.""" |
| 425 | ec_transfer_config: ECTransferConfig | None = None |
| 426 | """The configurations for distributed EC cache transfer.""" |
| 427 | ec_manager_config: EncoderCacheManagerConfig = Field( |
| 428 | default_factory=EncoderCacheManagerConfig |
| 429 | ) |
| 430 | """The configurations for custom encoder cache manager.""" |
| 431 | reasoning_config: ReasoningConfig | None = None |
| 432 | """The configurations for reasoning model.""" |
| 433 | # some opaque config, only used to provide additional information |
| 434 | # for the hash computation, mainly used for testing, debugging or out of |
| 435 | # tree config registration. |
| 436 | additional_config: dict | SupportsHash = Field(default_factory=dict) |
| 437 | """Additional config for specified platform. Different platforms may |
| 438 | support different configs. Make sure the configs are valid for the platform |
| 439 | you are using. Contents must be hashable.""" |
| 440 | instance_id: str = "" |
| 441 | """The ID of the vLLM instance.""" |
| 442 | optimization_level: OptimizationLevel = OptimizationLevel.O2 |
| 443 | """The optimization level. These levels trade startup time cost for |
| 444 | performance, with -O0 having the best startup time and -O3 having the best |
| 445 | performance. -O2 is used by default. See OptimizationLevel for full |
| 446 | description.""" |
| 447 | |
| 448 | performance_mode: PerformanceMode = "balanced" |
| 449 | """Performance mode for runtime behavior, 'balanced' is the default. |
| 450 | 'interactivity' favors low end-to-end per-request latency at small batch |
| 451 | sizes (fine-grained CUDA graphs, latency-oriented kernels). |
| 452 | 'throughput' favors aggregate tokens/sec at high concurrency (larger CUDA |
| 453 | graphs, more aggressive batching, throughput-oriented kernels).""" |
| 454 | |
| 455 | weight_transfer_config: WeightTransferConfig | None = None |
| 456 | """The configurations for weight transfer during RL training.""" |
| 457 | |
| 458 | shutdown_timeout: int = Field(default=0, ge=0) |
| 459 | """Shutdown grace period for in-flight requests. Shutdown will be delayed for |
| 460 | up to this amount of time to allow already-running requests to complete. Any |
| 461 | remaining requests are aborted once the timeout is reached. |
| 462 | """ |
| 463 | |
| 464 | def compute_hash(self, include_version: bool = True) -> str: |
| 465 | """WARNING: Whenever a new field is added to this config, |
| 466 | ensure that it is included in the factors list if |
| 467 | it affects the computation graph. |
| 468 | |
| 469 | Provide a hash that uniquely identifies all the configs |
| 470 | that affect the structure of the computation |
| 471 | graph from input ids/embeddings to the final hidden states, |
| 472 | excluding anything before input ids/embeddings and after |
| 473 | the final hidden states. |
| 474 | |
| 475 | Args: |
| 476 | include_version: Include the vLLM version in the hash. |
| 477 | |
| 478 | """ |
| 479 | factors: list[Any] = [] |
| 480 | |
| 481 | # summarize vllm config |
| 482 | vllm_factors: list[Any] = [] |
| 483 | if include_version: |
| 484 | from vllm import __version__ |
| 485 | |
| 486 | vllm_factors.append(__version__) |
| 487 | if self.model_config: |
| 488 | vllm_factors.append(self.model_config.compute_hash()) |
| 489 | if ( |
| 490 | self.compilation_config |
| 491 | and getattr(self.compilation_config, "compile_mm_encoder", False) |
| 492 | and self.model_config.multimodal_config |
| 493 | ): |
| 494 | vllm_factors.append(self.model_config.multimodal_config.compute_hash()) |
| 495 | else: |
| 496 | vllm_factors.append("None") |
| 497 | if self.cache_config: |
| 498 | vllm_factors.append(self.cache_config.compute_hash()) |
| 499 | else: |
| 500 | vllm_factors.append("None") |
| 501 | if self.parallel_config: |
| 502 | vllm_factors.append(self.parallel_config.compute_hash()) |
| 503 | else: |
| 504 | vllm_factors.append("None") |
| 505 | if self.scheduler_config: |
| 506 | vllm_factors.append(self.scheduler_config.compute_hash()) |
| 507 | else: |
| 508 | vllm_factors.append("None") |
| 509 | if self.device_config: |
| 510 | vllm_factors.append(self.device_config.compute_hash()) |
| 511 | else: |
| 512 | vllm_factors.append("None") |
| 513 | if self.load_config: |
| 514 | vllm_factors.append(self.load_config.compute_hash()) |
| 515 | else: |
| 516 | vllm_factors.append("None") |
| 517 | if self.offload_config: |
| 518 | vllm_factors.append(self.offload_config.compute_hash()) |
| 519 | else: |
| 520 | vllm_factors.append("None") |
| 521 | if self.attention_config: |
| 522 | vllm_factors.append(self.attention_config.compute_hash()) |
| 523 | else: |
| 524 | vllm_factors.append("None") |
| 525 | vllm_factors.append( |
| 526 | self.engram_config.compute_hash() |
| 527 | if self.engram_config is not None |
| 528 | else "None" |
| 529 | ) |
| 530 | if self.lora_config: |
| 531 | vllm_factors.append(self.lora_config.compute_hash()) |
| 532 | else: |
| 533 | vllm_factors.append("None") |
| 534 | if self.speculative_config: |
| 535 | vllm_factors.append(self.speculative_config.compute_hash()) |
| 536 | else: |
| 537 | vllm_factors.append("None") |
| 538 | if self.structured_outputs_config: |
| 539 | vllm_factors.append(self.structured_outputs_config.compute_hash()) |
| 540 | if self.profiler_config: |
| 541 | vllm_factors.append(self.profiler_config.compute_hash()) |
| 542 | else: |
| 543 | vllm_factors.append("None") |
| 544 | vllm_factors.append(self.observability_config.compute_hash()) |
| 545 | if self.quant_config: |
| 546 | pass # should be captured by model_config.quantization |
| 547 | if self.compilation_config: |
| 548 | vllm_factors.append(self.compilation_config.compute_hash()) |
| 549 | else: |
| 550 | vllm_factors.append("None") |
| 551 | if self.kernel_config: |
| 552 | vllm_factors.append(self.kernel_config.compute_hash()) |
| 553 | else: |
| 554 | vllm_factors.append(None) |
| 555 | if self.kv_transfer_config: |
| 556 | vllm_factors.append(self.kv_transfer_config.compute_hash()) |
| 557 | else: |
| 558 | vllm_factors.append("None") |
| 559 | if self.ec_transfer_config: |
| 560 | vllm_factors.append(self.ec_transfer_config.compute_hash()) |
| 561 | else: |
| 562 | vllm_factors.append("None") |
| 563 | vllm_factors.append(self.aux_output_config.compute_hash()) |
| 564 | if self.additional_config: |
| 565 | if isinstance(additional_config := self.additional_config, dict): |
| 566 | additional_config_hash = safe_hash( |
| 567 | json.dumps(additional_config, sort_keys=True).encode(), |
| 568 | usedforsecurity=False, |
| 569 | ).hexdigest() |
| 570 | else: |
| 571 | additional_config_hash = additional_config.compute_hash() |
| 572 | vllm_factors.append(additional_config_hash) |
| 573 | else: |
| 574 | vllm_factors.append("None") |
| 575 | factors.append(vllm_factors) |
| 576 | |
| 577 | hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()[ |
| 578 | :10 |
| 579 | ] |
| 580 | return hash_str |
| 581 | |
| 582 | @property |
| 583 | def is_mm_encoder_only(self) -> bool: |
| 584 | mm_config = ( |
| 585 | self.model_config.multimodal_config |
| 586 | if self.model_config is not None |
| 587 | else None |
| 588 | ) |
| 589 | return bool(mm_config and mm_config.mm_encoder_only) |
| 590 | |
| 591 | @property |
| 592 | def max_concurrent_batches(self) -> int: |
| 593 | # PP requires PP-size concurrent batches to fill the pipeline. |
| 594 | # Async scheduling requires 2 concurrent batches to overlap. |
| 595 | pp_size = self.parallel_config.pipeline_parallel_size |
| 596 | if self.scheduler_config.async_scheduling: |
| 597 | if self.use_v2_model_runner: |
| 598 | return pp_size + 1 |
| 599 | # V1 Model Runner does not fully support async scheduling with PP. |
| 600 | if pp_size <= 1: |
| 601 | return 2 |
| 602 | return pp_size |
| 603 | |
| 604 | @property |
| 605 | def max_in_flight_tokens(self) -> int: |
| 606 | # Upper bound on tokens that are scheduled but not yet settled (freed): |
| 607 | # every concurrent batch may hold up to a full `max_num_batched_tokens`. |
| 608 | # Recycling-aware KV cache specs (sliding-window, chunked-local) reserve |
| 609 | # for this because out-of-window blocks are freed on the processed-token |
| 610 | # basis, so in-flight steps transiently keep their blocks. |
| 611 | return ( |
| 612 | self.max_concurrent_batches * self.scheduler_config.max_num_batched_tokens |
| 613 | ) |
| 614 | |
| 615 | @property |
| 616 | def num_speculative_tokens(self) -> int: |
| 617 | if ( |
| 618 | self.speculative_config is not None |
| 619 | and self.speculative_config.num_speculative_tokens is not None |
| 620 | ): |
| 621 | return self.speculative_config.num_speculative_tokens |
| 622 | if ( |
| 623 | self.diffusion_config is not None |
| 624 | and self.diffusion_config.canvas_length is not None |
| 625 | ): |
| 626 | return self.diffusion_config.canvas_length |
| 627 | return 0 |
| 628 | |
| 629 | @property |
| 630 | def num_lookahead_tokens(self) -> int: |
| 631 | """KV slots to reserve past the tokens the target model is scheduled for. |
| 632 | |
| 633 | The drafter writes KV for positions beyond the target model's query |
| 634 | range, so every component that reserves blocks must add this margin: |
| 635 | the scheduler through `allocate_slots`, and the worker warmup, which |
| 636 | builds its own `SchedulerOutput`s. Consumers must read this property |
| 637 | rather than re-deriving their own per-method lookahead, so the |
| 638 | scheduler and warmup cannot drift apart. |
| 639 | """ |
| 640 | speculative_config = self.speculative_config |
| 641 | if speculative_config is None: |
| 642 | return 0 |
| 643 | if speculative_config.use_dflash(): |
| 644 | # DFlash requires an extra lookahead slot since it uses in-fill-style |
| 645 | # decoding instead of standard next-token sampling, so it has a query |
| 646 | # for the last sampled token plus queries for each draft token. |
| 647 | return self.num_speculative_tokens + 1 |
| 648 | if speculative_config.use_eagle() or speculative_config.uses_draft_model(): |
| 649 | # DSpark (covered by use_eagle) drafts a block of num_speculative_tokens |
| 650 | # query tokens in which the anchor itself is the first prediction |
| 651 | # position (no separate bonus query), so it needs exactly |
| 652 | # num_speculative_tokens lookahead slots. |
| 653 | return self.num_speculative_tokens |
| 654 | return 0 |
| 655 | |
| 656 | @property |
| 657 | def num_prefill_lookahead_tokens(self) -> int: |
| 658 | """Prefill tokens past the computed range that the drafter reads. |
| 659 | |
| 660 | Mid-prefill the drafter consumes tokens the target model has not been |
| 661 | scheduled for yet, so every component that has to keep them available |
| 662 | must apply this margin: the scheduler, which never ends a chunk within |
| 663 | it and shifts encoder scheduling by it, and the KV cache manager, which |
| 664 | treats the trailing `this - 1` tokens as re-prefillable rather than |
| 665 | finalized. Consumers must read this property rather than re-deriving |
| 666 | their own per-method lookahead, so those components cannot drift apart. |
| 667 | """ |
| 668 | speculative_config = self.speculative_config |
| 669 | if speculative_config is None or not speculative_config.use_eagle(): |
| 670 | return 0 |
| 671 | if speculative_config.use_multi_module_mtp(): |
| 672 | # Each MTP module reads one token further ahead than the one before |
| 673 | # it, so the chain needs num_speculative_tokens of runway at a |
| 674 | # chunked-prefill boundary. |
| 675 | return self.num_speculative_tokens |
| 676 | # Eagle-family drafters read only the immediate next token. |
| 677 | return 1 |
| 678 | |
| 679 | @property |
| 680 | def uniform_decode_query_len(self) -> int: |
| 681 | """Query length of every request in a uniform decode batch. |
| 682 | |
| 683 | A decode step submits one query for the newly sampled token plus one |
| 684 | for each draft token, so the widest uniform decode batch the scheduler |
| 685 | can build is `max_num_seqs * uniform_decode_query_len` tokens. Anything |
| 686 | that has to cover a decode batch reads this, so the sizing rule cannot |
| 687 | drift between the places that apply it. |
| 688 | |
| 689 | This deliberately does not derive from the KV slots a drafter reserves |
| 690 | past the target's query range, which is a *reservation* contract rather |
| 691 | than a query-length one. The two do not differ by a constant: DFlash |
| 692 | reserves `num_speculative_tokens + 1` slots yet still verifies `1 + |
| 693 | num_speculative_tokens` queries, while EAGLE reserves |
| 694 | `num_speculative_tokens` and verifies the same `1 + n`. Deriving one |
| 695 | from the other would under-size EAGLE by a full request width, which is |
| 696 | the failure this property exists to prevent. |
| 697 | """ |
| 698 | return 1 + self.num_speculative_tokens |
| 699 | |
| 700 | @property |
| 701 | def use_v2_model_runner(self) -> bool: |
| 702 | if self.attention_config.hisparse_config is not None: |
| 703 | if envs.VLLM_USE_V2_MODEL_RUNNER is False: |
| 704 | raise ValueError( |
| 705 | "HiSparse requires Model Runner V2; remove " |
| 706 | "VLLM_USE_V2_MODEL_RUNNER=0." |
| 707 | ) |
| 708 | return True |
| 709 | |
| 710 | if getattr(self, "watermark_config", None) is not None: |
| 711 | if envs.VLLM_USE_V2_MODEL_RUNNER is False: |
| 712 | logger.info_once( |
| 713 | "Watermarking requires Model Runner V2 and overrides " |
| 714 | "VLLM_USE_V2_MODEL_RUNNER=0." |
| 715 | ) |
| 716 | return True |
| 717 | |
| 718 | use_v2_model_runner = envs.VLLM_USE_V2_MODEL_RUNNER |
| 719 | if use_v2_model_runner is not None: |
| 720 | return use_v2_model_runner |
| 721 | |
| 722 | from vllm.platforms import current_platform |
| 723 | |
| 724 | model_config = self.model_config |
| 725 | if model_config is not None and current_platform.is_rocm(): |
| 726 | architectures = getattr(model_config, "architectures", ()) |
| 727 | if any(arch in ROCM_DEFAULT_MRV1_ARCHITECTURES for arch in architectures): |
| 728 | # This default is a speed preference, not a claim that V1 can |
| 729 | # serve the config, so it yields where V1 cannot. It yields by |
| 730 | # falling through to the checks below, not by selecting V2. |
| 731 | v1_unsupported = self._get_v1_model_runner_unsupported_features() |
| 732 | if not v1_unsupported: |
| 733 | logger.warning_once( |
| 734 | "Defaulting to V1 model runner on ROCm for model " |
| 735 | "architectures: %s", |
| 736 | ", ".join(architectures), |
| 737 | ) |
| 738 | return False |
| 739 | logger.warning_once( |
| 740 | "Skipping the ROCm V1 model runner default for %s: V1 does " |
| 741 | "not support %s.", |
| 742 | ", ".join(architectures), |
| 743 | ", ".join(v1_unsupported), |
| 744 | ) |
| 745 | |
| 746 | if not HAS_TRITON: |
| 747 | logger.warning_once( |
| 748 | "Model Runner V2 requires Triton; using the V1 model runner instead." |
| 749 | ) |
| 750 | return False |
| 751 | |
| 752 | unsupported = self._get_v2_model_runner_unsupported_features() |
| 753 | if unsupported: |
| 754 | logger.warning_once( |
| 755 | "Model Runner V2 does not yet support %s; using the V1 model " |
| 756 | "runner instead.", |
| 757 | ", ".join(unsupported), |
| 758 | ) |
| 759 | return False |
| 760 | |
| 761 | return True |
| 762 | |
| 763 | def _is_dflash_candidate_draft(self) -> bool: |
| 764 | """Whether the DFlash draft has a candidate head, by the architecture the |
| 765 | speculator selects on (v1/worker/gpu/spec_decode/__init__.py).""" |
| 766 | spec = self.speculative_config |
| 767 | if spec is None or spec.method != "dflash": |
| 768 | return False |
| 769 | draft_config = getattr(spec, "draft_model_config", None) |
| 770 | if draft_config is None: |
| 771 | return False |
| 772 | return bool( |
| 773 | {"DFlash2DraftModel", "LiLiCorrDraftModel"}.intersection( |
| 774 | draft_config.architectures or [] |
| 775 | ) |
| 776 | ) |
| 777 | |
| 778 | def _dflash_needs_multi_kv_group(self) -> bool: |
| 779 | """Whether a DFlash draft mixes sliding-window and full attention.""" |
| 780 | spec = self.speculative_config |
| 781 | if spec is None or spec.method != "dflash": |
| 782 | return False |
| 783 | draft_config = getattr(spec, "draft_model_config", None) |
| 784 | if draft_config is None: |
| 785 | return False |
| 786 | layer_types = getattr(draft_config.hf_config, "layer_types", None) or [] |
| 787 | num_sliding = sum(lt == "sliding_attention" for lt in layer_types) |
| 788 | return 0 < num_sliding < len(layer_types) |
| 789 | |
| 790 | def _uses_breakable_cudagraph_by_default(self) -> bool: |
| 791 | model_config = self.model_config |
| 792 | if model_config is None: |
| 793 | return False |
| 794 | |
| 795 | architectures = set(model_config.architectures) |
| 796 | return bool(architectures & default_breakable_cudagraph_architectures()) |
| 797 | |
| 798 | def _uses_breakable_cudagraph_for_batch_invariance(self) -> bool: |
| 799 | """Avoid freezing runtime-M tile lookup in compiled forward (#54243). |
| 800 | Breakable graphs look up tuned bf16, unquantized qkv/o/gate_up/down tiles |
| 801 | at capture; lm_head runs outside compiled forward and does not benefit.""" |
| 802 | from vllm.model_executor.determinism import batch_invariant_configs as bi |
| 803 | from vllm.platforms import current_platform |
| 804 | |
| 805 | model = self.model_config |
| 806 | if ( |
| 807 | not envs.VLLM_BATCH_INVARIANT |
| 808 | or model is None |
| 809 | or model.enforce_eager |
| 810 | or model.dtype != torch.bfloat16 |
| 811 | or model.quantization is not None |
| 812 | or not current_platform.is_cuda() |
| 813 | # Sequence parallelism / async TP are torch.compile passes. |
| 814 | or self.compilation_config.pass_config.enable_sp |
| 815 | or self.compilation_config.pass_config.fuse_gemm_comms |
| 816 | ): |
| 817 | return False |
| 818 | family = bi._get_tuned_matmul_arch_family( |
| 819 | current_platform.get_device_capability() |
| 820 | ) |
| 821 | if family is None or family not in bi._BATCH_INVARIANT_MATMUL_TUNED_CONFIGS: |
| 822 | return False |
| 823 | table = bi._BATCH_INVARIANT_MATMUL_TUNED_CONFIGS[family] |
| 824 | parallel = self.parallel_config |
| 825 | tp = parallel.tensor_parallel_size |
| 826 | hidden = model.get_hidden_size() |
| 827 | head = model.get_head_size() |
| 828 | heads = model.get_num_attention_heads(parallel) |
| 829 | kv_heads = model.get_num_kv_heads(parallel) |
| 830 | shapes = [((heads + 2 * kv_heads) * head, hidden), (hidden, heads * head)] |
| 831 | intermediate = getattr(model.hf_text_config, "intermediate_size", None) |
| 832 | # Per-layer sizes (e.g. Gemma3n) are not modeled. |
| 833 | if isinstance(intermediate, int): |
| 834 | shapes += [(2 * intermediate // tp, hidden), (hidden, intermediate // tp)] |
| 835 | return any(shape in table for shape in shapes) |
| 836 | |
| 837 | def _maybe_enable_breakable_cudagraph(self) -> bool: |
| 838 | if ( |
| 839 | "VLLM_USE_BREAKABLE_CUDAGRAPH" not in os.environ |
| 840 | and self._uses_breakable_cudagraph_by_default() |
| 841 | ): |
| 842 | os.environ["VLLM_USE_BREAKABLE_CUDAGRAPH"] = "1" |
| 843 | logger.info_once( |
| 844 | "Auto-enabling VLLM_USE_BREAKABLE_CUDAGRAPH=1. " |
| 845 | "Set VLLM_USE_BREAKABLE_CUDAGRAPH=0 to opt out." |
| 846 | ) |
| 847 | elif ( |
| 848 | "VLLM_USE_BREAKABLE_CUDAGRAPH" not in os.environ |
| 849 | and envs.VLLM_BATCH_INVARIANT |
| 850 | and self._uses_breakable_cudagraph_for_batch_invariance() |
| 851 | ): |
| 852 | os.environ["VLLM_USE_BREAKABLE_CUDAGRAPH"] = "1" |
| 853 | logger.info_once( |
| 854 | "VLLM_BATCH_INVARIANT=1: auto-enabling VLLM_USE_BREAKABLE_CUDAGRAPH=1. " |
| 855 | "Set VLLM_USE_BREAKABLE_CUDAGRAPH=0 to opt out." |
| 856 | ) |
| 857 | |
| 858 | from vllm.compilation.breakable_cudagraph import ( |
| 859 | is_breakable_cudagraph_enabled, |
| 860 | ) |
| 861 | |
| 862 | enabled = is_breakable_cudagraph_enabled() |
| 863 | if enabled: |
| 864 | self.compilation_config.mode = CompilationMode.NONE |
| 865 | return enabled |
| 866 | |
| 867 | @property |
| 868 | def needs_dp_coordinator(self) -> bool: |
| 869 | """Determine if the DPCoordinator process is needed. |
| 870 | |
| 871 | The DPCoordinator is needed in two cases: |
| 872 | 1. For MoE models with DP > 1: to handle wave coordination |
| 873 | (even in external LB mode, since wave coordination runs in the coordinator) |
| 874 | 2. For non-MoE models in internal/hybrid LB mode: to collect and publish |
| 875 | queue stats for load balancing across DP ranks |
| 876 | |
| 877 | Returns: |
| 878 | True if DPCoordinator process is needed, False otherwise. |
| 879 | |
| 880 | """ |
| 881 | # For non-MoE models, only need coordinator in internal/hybrid LB mode |
| 882 | # (for stats collection). |
| 883 | return self.parallel_config.data_parallel_size > 1 and ( |
| 884 | self.model_config is None |
| 885 | or self.model_config.is_moe |
| 886 | or not self.parallel_config.data_parallel_external_lb |
| 887 | ) |
| 888 | |
| 889 | def enable_trace_function_call_for_thread(self) -> None: |
| 890 | """Set up function tracing for the current thread, |
| 891 | if enabled via the `VLLM_TRACE_FUNCTION` environment variable. |
| 892 | """ |
| 893 | if envs.VLLM_TRACE_FUNCTION: |
| 894 | tmp_dir = tempfile.gettempdir() |
| 895 | # add username to tmp_dir to avoid permission issues |
| 896 | tmp_dir = os.path.join(tmp_dir, getpass.getuser()) |
| 897 | filename = ( |
| 898 | f"VLLM_TRACE_FUNCTION_for_process_{os.getpid()}" |
| 899 | f"_thread_{threading.get_ident()}_at_{datetime.now()}.log" |
| 900 | ).replace(" ", "_") |
| 901 | log_path = os.path.join( |
| 902 | tmp_dir, |
| 903 | "vllm", |
| 904 | f"vllm-instance-{self.instance_id}", |
| 905 | filename, |
| 906 | ) |
| 907 | os.makedirs(os.path.dirname(log_path), exist_ok=True) |
| 908 | enable_trace_function_call(log_path) |
| 909 | |
| 910 | @staticmethod |
| 911 | def _get_quantization_config( |
| 912 | model_config: ModelConfig, load_config: LoadConfig |
| 913 | ) -> QuantizationConfig | None: |
| 914 | """Get the quantization config.""" |
| 915 | from vllm.platforms import current_platform |
| 916 | |
| 917 | if model_config.quantization is not None: |
| 918 | from vllm.model_executor.model_loader.weight_utils import get_quant_config |
| 919 | |
| 920 | quant_config = get_quant_config(model_config, load_config) |
| 921 | capability_tuple = current_platform.get_device_capability() |
| 922 | |
| 923 | if capability_tuple is not None: |
| 924 | capability = capability_tuple.to_int() |
| 925 | if capability < quant_config.get_min_capability(): |
| 926 | raise ValueError( |
| 927 | f"The quantization method {model_config.quantization} " |
| 928 | "is not supported for the current GPU. Minimum " |
| 929 | f"capability: {quant_config.get_min_capability()}. " |
| 930 | f"Current capability: {capability}." |
| 931 | ) |
| 932 | supported_dtypes = quant_config.get_supported_act_dtypes() |
| 933 | if model_config.dtype not in supported_dtypes: |
| 934 | raise ValueError( |
| 935 | f"{model_config.dtype} is not supported for quantization " |
| 936 | f"method {model_config.quantization}. Supported dtypes: " |
| 937 | f"{supported_dtypes}" |
| 938 | ) |
| 939 | quant_config.maybe_update_config( |
| 940 | model_config.model, |
| 941 | hf_config=model_config.hf_config, |
| 942 | revision=model_config.revision, |
| 943 | ) |
| 944 | return quant_config |
| 945 | return None |
| 946 | |
| 947 | @staticmethod |
| 948 | def get_quantization_config( |
| 949 | model_config: ModelConfig, load_config: LoadConfig |
| 950 | ) -> QuantizationConfig | None: |
| 951 | import copy |
| 952 | |
| 953 | # For some reason, the _ version of this modifies the model_config |
| 954 | # object, so using deepcopy to avoid this problem. |
| 955 | return VllmConfig._get_quantization_config( |
| 956 | copy.deepcopy(model_config), load_config |
| 957 | ) |
| 958 | |
| 959 | def with_hf_config( |
| 960 | self, |
| 961 | hf_config: PreTrainedConfig, |
| 962 | architectures: list[str] | None = None, |
| 963 | ) -> "VllmConfig": |
| 964 | if architectures is not None: |
| 965 | hf_config = copy.deepcopy(hf_config) |
| 966 | hf_config.architectures = architectures |
| 967 | elif hf_config.architectures is None: |
| 968 | from transformers.models.auto.modeling_auto import ( |
| 969 | MODEL_FOR_CAUSAL_LM_MAPPING_NAMES, |
| 970 | ) |
| 971 | |
| 972 | if hf_config.model_type in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES: |
| 973 | hf_config = copy.deepcopy(hf_config) |
| 974 | hf_config.architectures = [ |
| 975 | MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[hf_config.model_type] |
| 976 | ] |
| 977 | |
| 978 | model_config = copy.deepcopy(self.model_config) |
| 979 | |
| 980 | # In Transformers v5, tie_word_embeddings belongs to the config of the class |
| 981 | # that can see both layers to be tied. For example: |
| 982 | # |
| 983 | # SomeVLModel: |
| 984 | # self.language_model = SomeLanguageModel(SomeVLTextConfig) |
| 985 | # self.vision_model = SomeVisionModel(SomeVLVisionConfig) |
| 986 | # |
| 987 | # SomeVLModelForMultimodalLM: |
| 988 | # self.model = SomeVLModel(SomeVLConfig) |
| 989 | # self.lm_head = nn.Linear() |
| 990 | # |
| 991 | # Therefore, tie_word_embeddings is defined in SomeVLConfig and is not present |
| 992 | # in SomeVLTextConfig*. In vLLM, the lm_head belongs to the language_model, so |
| 993 | # we must ensure that tie_word_embeddings is set in the language_model's config. |
| 994 | # |
| 995 | # *For some models, SomeVLTextConfig may also have a tie_word_embeddings field. |
| 996 | # This is only the case if SomeVLTextConfig is also used for a text only version |
| 997 | # of the same model. For example: |
| 998 | # |
| 999 | # SomeVLModelForCausalLM: |
| 1000 | # self.model = SomeLanguageModel(SomeVLTextConfig) |
| 1001 | # self.lm_head = nn.Linear() |
| 1002 | # |
| 1003 | # Therefore, the presence of tie_word_embeddings in SomeVLTextConfig cannot |
| 1004 | # be used as a signal for whether tie_word_embeddings should be copied from |
| 1005 | # hf_config to the language_model config. |
| 1006 | if model_config.is_multimodal_model and hasattr( |
| 1007 | model_config.hf_config, "tie_word_embeddings" |
| 1008 | ): |
| 1009 | tie_word_embeddings = model_config.hf_config.tie_word_embeddings |
| 1010 | hf_config.get_text_config().tie_word_embeddings = tie_word_embeddings |
| 1011 | |
| 1012 | model_config.hf_config = hf_config |
| 1013 | model_config.model_arch_config = model_config.get_model_arch_config() |
| 1014 | model_config.is_submodel_config = True |
| 1015 | |
| 1016 | return replace(self, model_config=model_config) |
| 1017 | |
| 1018 | def _set_config_default(self, config_obj: Any, key: str, value: Any) -> None: |
| 1019 | """Set config attribute to default if not already set by user. |
| 1020 | |
| 1021 | Args: |
| 1022 | config_obj: Configuration object to update. |
| 1023 | key: Attribute name. |
| 1024 | value: Default value (static or callable). |
| 1025 | |
| 1026 | """ |
| 1027 | if getattr(config_obj, key) is None: |
| 1028 | # Some config values are known before initialization and are |
| 1029 | # hard coded. |
| 1030 | # Other values depend on the user given configuration, so they are |
| 1031 | # implemented with lambda functions and decided at run time. |
| 1032 | setattr(config_obj, key, value(self) if callable(value) else value) |
| 1033 | |
| 1034 | def _apply_optimization_level_defaults(self, defaults: dict[str, Any]) -> None: |
| 1035 | """Apply optimization level defaults using self as root. |
| 1036 | |
| 1037 | Recursively applies values from defaults into nested config objects. |
| 1038 | Only fields present in defaults are overwritten. |
| 1039 | |
| 1040 | If the user configuration does not specify a value for a default field |
| 1041 | and if the default field is still None after all user selections are |
| 1042 | applied, then default values will be applied to the field. User specified |
| 1043 | fields will not be overridden by the default. |
| 1044 | |
| 1045 | Args: |
| 1046 | defaults: Dictionary of default values to apply. |
| 1047 | |
| 1048 | """ |
| 1049 | |
| 1050 | def apply_recursive(config_obj: Any, config_defaults: dict[str, Any]) -> None: |
| 1051 | """Recursively apply defaults to config_obj, using self as root.""" |
| 1052 | for key, value in config_defaults.items(): |
| 1053 | if not hasattr(config_obj, key): |
| 1054 | continue |
| 1055 | |
| 1056 | current = getattr(config_obj, key) |
| 1057 | if isinstance(value, dict) and is_dataclass(current): |
| 1058 | apply_recursive(current, value) |
| 1059 | else: |
| 1060 | self._set_config_default(config_obj, key, value) |
| 1061 | |
| 1062 | apply_recursive(self, defaults) |
| 1063 | |
| 1064 | def _maybe_override_dynamic_sd_cudagraph_mode(self) -> None: |
| 1065 | speculative_config = self.speculative_config |
| 1066 | if ( |
| 1067 | speculative_config is None |
| 1068 | or not speculative_config.uses_dynamic_speculative_decoding() |
| 1069 | or not self.compilation_config.cudagraph_mode.has_full_cudagraphs() |
| 1070 | or self.use_v2_model_runner |
| 1071 | ): |
| 1072 | return |
| 1073 | |
| 1074 | logger.warning_once( |
| 1075 | "Dynamic speculative decoding changes the target verification " |
| 1076 | "length at runtime. Overriding cudagraph_mode from %s to " |
| 1077 | "PIECEWISE for reliability. Use VLLM_USE_V2_MODEL_RUNNER=1 " |
| 1078 | "if you want to use full CUDA graphs.", |
| 1079 | self.compilation_config.cudagraph_mode.name, |
| 1080 | ) |
| 1081 | self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE |
| 1082 | |
| 1083 | def _maybe_disable_dynamic_sd_for_data_parallel(self) -> None: |
| 1084 | speculative_config = self.speculative_config |
| 1085 | if ( |
| 1086 | speculative_config is None |
| 1087 | or not speculative_config.uses_dynamic_speculative_decoding() |
| 1088 | or self.parallel_config.data_parallel_size <= 1 |
| 1089 | ): |
| 1090 | return |
| 1091 | |
| 1092 | logger.warning_once( |
| 1093 | "Dynamic speculative decoding is not supported with data " |
| 1094 | "parallelism because data-parallel ranks can select different " |
| 1095 | "speculative-token counts, causing DP divergence and deadlocks. " |
| 1096 | "Disabling num_speculative_tokens_per_batch_size and falling back " |
| 1097 | "to static num_speculative_tokens=%d.", |
| 1098 | speculative_config.num_speculative_tokens, |
| 1099 | ) |
| 1100 | speculative_config.num_speculative_tokens_per_batch_size = None |
| 1101 | |
| 1102 | def _normalize_piecewise_cudagraph_mode( |
| 1103 | self, *, breakable_cudagraph_enabled: bool |
| 1104 | ) -> None: |
| 1105 | compilation_config = self.compilation_config |
| 1106 | if ( |
| 1107 | compilation_config.cudagraph_mode.requires_piecewise_compilation() |
| 1108 | and compilation_config.mode != CompilationMode.VLLM_COMPILE |
| 1109 | and not breakable_cudagraph_enabled |
| 1110 | ): |
| 1111 | fallback_mode = compilation_config.cudagraph_mode.without_piecewise() |
| 1112 | logger.info_once( |
| 1113 | "Cudagraph mode %s is not compatible with compilation mode %s. " |
| 1114 | "Overriding to %s.", |
| 1115 | compilation_config.cudagraph_mode, |
| 1116 | compilation_config.mode, |
| 1117 | fallback_mode, |
| 1118 | ) |
| 1119 | compilation_config.cudagraph_mode = fallback_mode |
| 1120 | if fallback_mode == CUDAGraphMode.NONE: |
| 1121 | compilation_config.max_cudagraph_capture_size = 0 |
| 1122 | compilation_config.cudagraph_capture_sizes = [] |
| 1123 | |
| 1124 | def _post_init_kv_transfer_config(self) -> None: |
| 1125 | """Update KVTransferConfig based on top-level configs in VllmConfig. |
| 1126 | |
| 1127 | Right now, this function reads the offloading settings from |
| 1128 | CacheConfig and configures the KVTransferConfig accordingly. |
| 1129 | """ |
| 1130 | # KV offloading is only activated when kv_offloading_size is set. |
| 1131 | if (kv_offloading_size := self.cache_config.kv_offloading_size) is None: |
| 1132 | return |
| 1133 | |
| 1134 | kv_offloading_backend = self.cache_config.kv_offloading_backend |
| 1135 | |
| 1136 | # If no KVTransferConfig is provided, create a default one. |
| 1137 | if self.kv_transfer_config is None: |
| 1138 | self.kv_transfer_config = KVTransferConfig() |
| 1139 | |
| 1140 | if kv_offloading_backend == "native": |
| 1141 | if envs.VLLM_USE_SIMPLE_KV_OFFLOAD: |
| 1142 | config_connector = "SimpleCPUOffloadConnector" |
| 1143 | else: |
| 1144 | config_connector = "OffloadingConnector" |
| 1145 | self.kv_transfer_config.kv_connector = config_connector |
| 1146 | self.kv_transfer_config.kv_connector_extra_config.update( |
| 1147 | {"cpu_bytes_to_use": kv_offloading_size * (1 << 30)} |
| 1148 | ) |
| 1149 | elif kv_offloading_backend == "lmcache": |
| 1150 | # Default to LMCache multi-process (MP) mode. The actual KV |
| 1151 | # storage capacity is managed by the standalone LMCache server |
| 1152 | # process, so ``kv_offloading_size`` is not propagated here. |
| 1153 | # ``LMCacheMPConnector`` falls back to ``tcp://localhost:5555`` |
| 1154 | # when host/port are not provided via extra_config. |
| 1155 | self.kv_transfer_config.kv_connector = "LMCacheMPConnector" |
| 1156 | |
| 1157 | # This is the same for all backends |
| 1158 | self.kv_transfer_config.kv_role = "kv_both" |
| 1159 | |
| 1160 | def _verify_aux_output_compatibility(self) -> None: |
| 1161 | """Reject configurations unsupported by enabled auxiliary outputs.""" |
| 1162 | if not self.aux_output_config.enabled: |
| 1163 | return |
| 1164 | from vllm.platforms import current_platform |
| 1165 | |
| 1166 | # In-tree platforms only wire AuxOutput to MRV2. TPU and out-of-tree |
| 1167 | # platforms bring their own model runners and validate AuxOutput |
| 1168 | # support themselves. |
| 1169 | if not self.use_v2_model_runner and not ( |
| 1170 | current_platform.is_tpu() or current_platform.is_out_of_tree() |
| 1171 | ): |
| 1172 | raise ValueError( |
| 1173 | "AuxOutput Connector requires Model Runner V2; set " |
| 1174 | "VLLM_USE_V2_MODEL_RUNNER=1." |
| 1175 | ) |
| 1176 | if self.model_config.runner_type != "generate": |
| 1177 | raise ValueError("AuxOutput Connector only supports generate runners.") |
| 1178 | if not self.model_config.is_moe: |
| 1179 | raise ValueError("AuxOutput Connector only supports MoE models.") |
| 1180 | if not self.cache_config.enable_prefix_caching: |
| 1181 | raise ValueError("AuxOutput Connector requires prefix caching.") |
| 1182 | if ( |
| 1183 | self.speculative_config is not None |
| 1184 | and self.speculative_config.enable_adaptive_verification |
| 1185 | ): |
| 1186 | raise ValueError( |
| 1187 | "--enable-return-routed-experts is incompatible with " |
| 1188 | "adaptive speculative verification." |
| 1189 | ) |
| 1190 | if self.parallel_config.pipeline_parallel_size > 1: |
| 1191 | raise ValueError( |
| 1192 | "--enable-return-routed-experts is incompatible with " |
| 1193 | "pipeline parallelism (PP > 1)." |
| 1194 | ) |
| 1195 | if ( |
| 1196 | self.parallel_config.decode_context_parallel_size > 1 |
| 1197 | or self.parallel_config.prefill_context_parallel_size > 1 |
| 1198 | ): |
| 1199 | raise ValueError( |
| 1200 | "--enable-return-routed-experts is incompatible with " |
| 1201 | "context parallelism (DCP/PCP > 1)." |
| 1202 | ) |
| 1203 | |
| 1204 | kv_transfer_config = self.kv_transfer_config |
| 1205 | if kv_transfer_config is not None: |
| 1206 | for connector_name in ( |
| 1207 | "NixlConnector", |
| 1208 | "NixlPullConnector", |
| 1209 | "NixlPushConnector", |
| 1210 | "MoRIIOConnector", |
| 1211 | "MooncakeConnector", |
| 1212 | ): |
| 1213 | if kv_transfer_config.has_connector(connector_name): |
| 1214 | raise ValueError( |
| 1215 | "--enable-return-routed-experts is incompatible with " |
| 1216 | f"{connector_name}; PD auxiliary output is not supported." |
| 1217 | ) |
| 1218 | |
| 1219 | def _verify_kv_transfer_compat(self) -> None: |
| 1220 | """Reject configurations that silently corrupt KV transfers.""" |
| 1221 | if ( |
| 1222 | self.kv_transfer_config is None |
| 1223 | or self.kv_transfer_config.kv_connector is None |
| 1224 | ): |
| 1225 | return |
| 1226 | |
| 1227 | # PyTorch's expandable_segments allocator uses CUDA VMM, which can |
| 1228 | # remap a virtual address range to different physical pages over the |
| 1229 | # engine's lifetime. KV connectors that pin KV cache memory (e.g. |
| 1230 | # NixlConnector via ibv_reg_mr, MooncakeConnector) end up with their |
| 1231 | # registrations pointing at stale physical pages after any remap, |
| 1232 | # producing RDMA failures like IBV_WC_REM_ACCESS_ERR / |
| 1233 | # NIXL_ERR_REMOTE_DISCONNECT at the first inter-node KV transfer. |
| 1234 | # We can't enumerate every in-tree and out-of-tree connector that |
| 1235 | # pins memory, so we conservatively reject the combination whenever |
| 1236 | # any KV connector is configured. |
| 1237 | # |
| 1238 | # CuMem allocator is exempt: CuMemAllocator.use_memory_pool toggles |
| 1239 | # expandable_segments off around its pool (see #40812), so the KV |
| 1240 | # cache allocated within that context lands on stable physical pages |
| 1241 | # even when the env var is set. |
| 1242 | if "expandable_segments:True" not in os.environ.get( |
| 1243 | "PYTORCH_CUDA_ALLOC_CONF", "" |
| 1244 | ): |
| 1245 | return |
| 1246 | if self.model_config is not None and (self.model_config.enable_cumem_allocator): |
| 1247 | return |
| 1248 | |
| 1249 | raise ValueError( |
| 1250 | f"KV connector {self.kv_transfer_config.kv_connector} is " |
| 1251 | "incompatible with PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True " |
| 1252 | "unless enable_cumem_allocator is also enabled. PyTorch's CUDA VMM " |
| 1253 | "allocator can remap KV cache virtual addresses to different " |
| 1254 | "physical pages, invalidating any pinned/registered KV memory " |
| 1255 | "(e.g. IB memory regions registered by NIXL or Mooncake). Either " |
| 1256 | "unset expandable_segments:True or enable the cumem allocator " |
| 1257 | "(sleep mode does this automatically and also " |
| 1258 | "routes KV allocations through CuMemAllocator's pool, where " |
| 1259 | "expandable_segments is automatically disabled)." |
| 1260 | ) |
| 1261 | |
| 1262 | def _verify_sampling_replay_config(self) -> None: |
| 1263 | model_config = self.model_config |
| 1264 | if model_config is None or not model_config.return_sampling_mask: |
| 1265 | return |
| 1266 | if not self.use_v2_model_runner: |
| 1267 | raise ValueError("sampling distribution replay requires Model Runner V2") |
| 1268 | speculative_config = self.speculative_config |
| 1269 | if ( |
| 1270 | speculative_config is not None |
| 1271 | and speculative_config.enable_adaptive_verification |
| 1272 | ): |
| 1273 | raise ValueError( |
| 1274 | "sampling distribution replay with speculative decoding " |
| 1275 | "requires fixed verification boundaries; disable adaptive " |
| 1276 | "verification" |
| 1277 | ) |
| 1278 | if model_config.is_diffusion: |
| 1279 | raise ValueError( |
| 1280 | "sampling distribution replay does not support diffusion models" |
| 1281 | ) |
| 1282 | if model_config.logits_processors: |
| 1283 | raise ValueError( |
| 1284 | "sampling distribution replay does not support custom logits processors" |
| 1285 | ) |
| 1286 | if model_config.logprobs_mode != "processed_logprobs": |
| 1287 | raise ValueError( |
| 1288 | "sampling distribution replay requires " |
| 1289 | "logprobs_mode='processed_logprobs' so that returned logprobs " |
| 1290 | "are normalized over the same nucleus as the sampling mask" |
| 1291 | ) |
| 1292 | |
| 1293 | def _verify_trace_replay_config(self) -> None: |
| 1294 | model_config = self.model_config |
| 1295 | if model_config is None or not model_config.enable_trace_replay: |
| 1296 | return |
| 1297 | if not self.use_v2_model_runner: |
| 1298 | raise ValueError("trace replay requires Model Runner V2") |
| 1299 | |
| 1300 | def _check_watermarking_unsupported( |
| 1301 | self, |
| 1302 | *, |
| 1303 | beam_search: bool = False, |
| 1304 | custom_sampler: bool = False, |
| 1305 | ) -> None: |
| 1306 | watermark_config = getattr(self, "watermark_config", None) |
| 1307 | if watermark_config is None: |
| 1308 | return |
| 1309 | if self.speculative_config is not None: |
| 1310 | speculative_config = self.speculative_config |
| 1311 | if speculative_config.draft_sample_method != "probabilistic": |
| 1312 | raise ValueError( |
| 1313 | "Speculative decoding with watermarking requires " |
| 1314 | "draft_sample_method='probabilistic'." |
| 1315 | ) |
| 1316 | if speculative_config.rejection_sample_method != "standard": |
| 1317 | raise ValueError( |
| 1318 | "Speculative decoding with watermarking requires " |
| 1319 | "rejection_sample_method='standard'." |
| 1320 | ) |
| 1321 | if ( |
| 1322 | speculative_config.parallel_drafting |
| 1323 | and speculative_config.method != "dspark" |
| 1324 | ): |
| 1325 | raise ValueError( |
| 1326 | "Parallel speculative drafting is not supported with watermarking." |
| 1327 | ) |
| 1328 | if speculative_config.method not in ("dspark", "eagle", "eagle3", "mtp"): |
| 1329 | raise ValueError( |
| 1330 | "Watermarking supports only autoregressive model-based " |
| 1331 | "speculative decoding." |
| 1332 | ) |
| 1333 | if ( |
| 1334 | not watermark_config.allow_target_only_watermarking |
| 1335 | and not watermark_config.supports_speculative_decoding |
| 1336 | ): |
| 1337 | raise ValueError( |
| 1338 | f"The '{watermark_config.algorithm}' watermarking algorithm " |
| 1339 | "does not support speculative decoding. Set " |
| 1340 | "allow_target_only_watermarking=true to leave draft tokens " |
| 1341 | "unwatermarked." |
| 1342 | ) |
| 1343 | if ( |
| 1344 | watermark_config.allow_target_only_watermarking |
| 1345 | and not watermark_config.supports_speculative_decoding |
| 1346 | ): |
| 1347 | logger.warning_once( |
| 1348 | "Target-only watermarking leaves accepted draft tokens " |
| 1349 | "unwatermarked, weakening detectability in proportion to the " |
| 1350 | "share of output tokens supplied by accepted drafts.", |
| 1351 | scope="global", |
| 1352 | ) |
| 1353 | if watermark_config.algorithm == "dual_key_gumbel" and ( |
| 1354 | watermark_config.alpha != get_field(WatermarkConfig, "alpha").default |
| 1355 | ): |
| 1356 | logger.warning_once( |
| 1357 | "Speculative decoding selects the watermark key by token role: " |
| 1358 | "draft tokens use key A, recovery and bonus tokens use key B. " |
| 1359 | "The configured alpha=%s is not used.", |
| 1360 | watermark_config.alpha, |
| 1361 | scope="global", |
| 1362 | ) |
| 1363 | if beam_search: |
| 1364 | raise ValueError("Beam search is not supported with watermarking.") |
| 1365 | if custom_sampler: |
| 1366 | raise ValueError( |
| 1367 | "Model-specific custom samplers are not supported with watermarking." |
| 1368 | ) |
| 1369 | |
| 1370 | def _resolve_and_verify_engram_config(self) -> None: |
| 1371 | """Resolve defaults and validate n-gram embedding settings.""" |
| 1372 | model_config = self.model_config |
| 1373 | speculative_config = self.speculative_config |
| 1374 | # Draft configs inherit the target's communication groups and settings. |
| 1375 | # Validate the target because the draft may disable n-gram embeddings. |
| 1376 | if ( |
| 1377 | speculative_config is not None |
| 1378 | and model_config is speculative_config.draft_model_config |
| 1379 | ): |
| 1380 | model_config = speculative_config.target_model_config |
| 1381 | if ( |
| 1382 | model_config is not None |
| 1383 | and model_config.architecture == "DeepseekV41ForCausalLM" |
| 1384 | and getattr(model_config.hf_text_config, "engram_layer_ids", None) |
| 1385 | and self.parallel_config.use_ubatching |
| 1386 | ): |
| 1387 | raise ValueError( |
| 1388 | "DeepSeek V4.1 Engram does not support DBO or microbatching. " |
| 1389 | "Disable --enable-dbo and set --ubatch-size to 0." |
| 1390 | ) |
| 1391 | if self.engram_config is None: |
| 1392 | if not model_has_engram_layers(model_config): |
| 1393 | return |
| 1394 | self.engram_config = EngramConfig() |
| 1395 | self.engram_config.verify_model_config(model_config) |
| 1396 | self.engram_config.resolve_dp_shared_memory(self.parallel_config) |
| 1397 | self.engram_config.verify_parallel_config(self.parallel_config) |
| 1398 | logger.info_once("Resolved Engram configuration: %s", str(self.engram_config)) |
| 1399 | |
| 1400 | def __post_init__(self): |
| 1401 | """Verify configs are valid & consistent with each other.""" |
| 1402 | # To give each torch profile run a unique instance name. |
| 1403 | self.instance_id = f"{time.time_ns()}" |
| 1404 | |
| 1405 | if self.model_config is not None and self.model_config.is_submodel_config: |
| 1406 | # with_hf_config() view: the parent config was already validated, |
| 1407 | # and this view's empty architecture list makes the model-dependent checks |
| 1408 | # below unsafe (e.g. use_mla resolves the architecture registry). |
| 1409 | return |
| 1410 | |
| 1411 | self._resolve_mm_encoder_only() |
| 1412 | |
| 1413 | if self.is_mm_encoder_only and self.cache_config.enable_prefix_caching: |
| 1414 | # Such an instance publishes encoder embeddings and runs no language |
| 1415 | # model, so it holds no KV cache for prefix caching to reuse and its |
| 1416 | # coordinator would have no group to manage. Disable before |
| 1417 | # `try_verify_and_update_config` so model config hooks (e.g. the |
| 1418 | # hybrid mamba hook setting `mamba_block_size`) already see prefix |
| 1419 | # caching as disabled. |
| 1420 | logger.info( |
| 1421 | "Disabling prefix caching: this instance runs the " |
| 1422 | "multi-modal encoder only." |
| 1423 | ) |
| 1424 | self.cache_config.enable_prefix_caching = False |
| 1425 | |
| 1426 | if self.performance_mode != "balanced": |
| 1427 | logger.info_once("Performance mode set to '%s'.", self.performance_mode) |
| 1428 | |
| 1429 | self.try_verify_and_update_config() |
| 1430 | self._resolve_and_verify_engram_config() |
| 1431 | |
| 1432 | self._check_watermarking_unsupported() |
| 1433 | # Models may have supplied their own DCP defaults above; anything still |
| 1434 | # unset falls back to the stock ones. |
| 1435 | self.parallel_config.set_dcp_defaults() |
| 1436 | |
| 1437 | if self.model_config is not None: |
| 1438 | self.model_config.verify_with_parallel_config(self.parallel_config) |
| 1439 | self.model_config.verify_dual_chunk_attention_config(self.load_config) |
| 1440 | |
| 1441 | self.parallel_config.is_moe_model = self.model_config.is_moe |
| 1442 | |
| 1443 | if ( |
| 1444 | self.model_config is not None |
| 1445 | and self.model_config.multimodal_config is not None |
| 1446 | and self.model_config.multimodal_config.language_model_only |
| 1447 | and self.compilation_config.cudagraph_mm_encoder |
| 1448 | ): |
| 1449 | raise ValueError( |
| 1450 | "--language-model-only is incompatible with " |
| 1451 | "cudagraph_mm_encoder=True, since it disables all multimodal " |
| 1452 | "inputs and the multimodal encoder is never run. Please " |
| 1453 | "disable one of them." |
| 1454 | ) |
| 1455 | |
| 1456 | self._verify_sampling_replay_config() |
| 1457 | self._verify_trace_replay_config() |
| 1458 | |
| 1459 | # A NIXL side is either fully replicated or fully DCP-sharded; MLA only. |
| 1460 | if ( |
| 1461 | self.kv_transfer_config is not None |
| 1462 | and self.kv_transfer_config.has_connector("NixlConnector") |
| 1463 | ): |
| 1464 | dcp_size = self.parallel_config.decode_context_parallel_size |
| 1465 | transfer_tp_size = max( |
| 1466 | self.parallel_config.tensor_parallel_size, |
| 1467 | self.parallel_config.prefill_context_parallel_size, |
| 1468 | ) |
| 1469 | assert dcp_size in (1, transfer_tp_size), ( |
| 1470 | f"decode_context_parallel_size={dcp_size} must be 1 or equal " |
| 1471 | f"to the NIXL transfer parallel size={transfer_tp_size}." |
| 1472 | ) |
| 1473 | if self.model_config is not None: |
| 1474 | assert self.model_config.use_mla or dcp_size == 1, ( |
| 1475 | "PD with decode_context_parallel_size > 1 is only " |
| 1476 | "supported for MLA models." |
| 1477 | ) |
| 1478 | if self.lora_config is not None: |
| 1479 | self.lora_config.verify_with_model_config(self.model_config) |
| 1480 | |
| 1481 | if ( |
| 1482 | self.mamba_config.enable_stochastic_rounding |
| 1483 | and self.cache_config.mamba_ssm_cache_dtype != "float16" |
| 1484 | ): |
| 1485 | raise ValueError( |
| 1486 | "Stochastic rounding for Mamba cache requires " |
| 1487 | "the SSM cache to be float16. Please set it explicitly, " |
| 1488 | "by specifying `--mamba-ssm-cache-dtype float16`, or disable " |
| 1489 | "stochastic rounding by not specifying " |
| 1490 | "`--enable-mamba-cache-stochastic-rounding`." |
| 1491 | ) |
| 1492 | |
| 1493 | if self.quant_config is None and self.model_config is not None: |
| 1494 | self.quant_config = VllmConfig._get_quantization_config( |
| 1495 | self.model_config, self.load_config |
| 1496 | ) |
| 1497 | |
| 1498 | # "dummy" reads no weights at all, and the sharded formats read a vLLM |
| 1499 | # state dict, which stores tied word embeddings under the lm_head only. |
| 1500 | # Neither can tell us what the original checkpoint contained. |
| 1501 | if self.model_config is not None and self.load_config.load_format not in ( |
| 1502 | "dummy", |
| 1503 | "sharded_state", |
| 1504 | "runai_streamer_sharded", |
| 1505 | ): |
| 1506 | self.model_config.maybe_untie_word_embeddings() |
| 1507 | |
| 1508 | if ( |
| 1509 | self.quant_config is not None |
| 1510 | and self.model_config is not None |
| 1511 | and hasattr(self.quant_config, "use_deep_gemm") |
| 1512 | and self.quant_config.use_deep_gemm is None |
| 1513 | ): |
| 1514 | from vllm.utils.deep_gemm import should_auto_disable_deep_gemm |
| 1515 | |
| 1516 | model_type = getattr(self.model_config.hf_text_config, "model_type", None) |
| 1517 | if should_auto_disable_deep_gemm(model_type): |
| 1518 | self.quant_config.use_deep_gemm = False |
| 1519 | logger.warning_once( |
| 1520 | "Auto-disabled DeepGemm for model_type=%s on Blackwell. " |
| 1521 | "DeepGemm E8M0 scale format causes accuracy degradation " |
| 1522 | "for this architecture. Falling back to CUTLASS. " |
| 1523 | "To disable DeepGemm globally, set VLLM_USE_DEEP_GEMM=0.", |
| 1524 | model_type, |
| 1525 | ) |
| 1526 | |
| 1527 | from vllm.platforms import current_platform |
| 1528 | from vllm.v1.executor.abstract import Executor |
| 1529 | |
| 1530 | executor_backend = self.parallel_config.distributed_executor_backend |
| 1531 | executor_class = Executor.get_class(self) |
| 1532 | executor_supports_async_sched = executor_class.supports_async_scheduling() |
| 1533 | uses_rocm_deepep_ht_dbo = ( |
| 1534 | current_platform.is_rocm() |
| 1535 | and self.parallel_config.enable_dbo |
| 1536 | and self.parallel_config.all2all_backend == "deepep_high_throughput" |
| 1537 | ) |
| 1538 | |
| 1539 | if self.scheduler_config.async_scheduling: |
| 1540 | # Async scheduling explicitly enabled, hard fail any incompatibilities. |
| 1541 | # Currently, async scheduling only support eagle speculative |
| 1542 | # decoding. |
| 1543 | if uses_rocm_deepep_ht_dbo: |
| 1544 | raise ValueError( |
| 1545 | "Async scheduling is not compatible with ROCm DeepEP " |
| 1546 | "high-throughput DBO. Please use --no-async-scheduling or " |
| 1547 | "select a different all2all backend." |
| 1548 | ) |
| 1549 | if self.speculative_config is not None: |
| 1550 | if ( |
| 1551 | self.speculative_config.method not in get_args(EagleModelTypes) |
| 1552 | and self.speculative_config.method not in get_args(NgramGPUTypes) |
| 1553 | and self.speculative_config.method != "draft_model" |
| 1554 | and self.speculative_config.method != "dspark" |
| 1555 | and self.speculative_config.method != "dflash" |
| 1556 | ): |
| 1557 | raise ValueError( |
| 1558 | "Currently, async scheduling is only supported " |
| 1559 | "with EAGLE/MTP/Draft Model/NGram GPU/DSpark/DFlash " |
| 1560 | "kind of speculative decoding" |
| 1561 | ) |
| 1562 | if self.speculative_config.disable_padded_drafter_batch: |
| 1563 | raise ValueError( |
| 1564 | "Async scheduling is not compatible with " |
| 1565 | "disable_padded_drafter_batch=True." |
| 1566 | ) |
| 1567 | if not executor_supports_async_sched: |
| 1568 | raise ValueError( |
| 1569 | f"`{executor_backend}` does not support async scheduling yet." |
| 1570 | ) |
| 1571 | elif self.scheduler_config.async_scheduling is None: |
| 1572 | # Enable async scheduling unless there is an incompatible option. |
| 1573 | if ( |
| 1574 | self.model_config is not None |
| 1575 | and self.model_config.runner_type == "pooling" |
| 1576 | ): |
| 1577 | # The current implementation of asynchronous scheduling negatively |
| 1578 | # impacts performance of pooling models, so we disable by default. |
| 1579 | logger.debug( |
| 1580 | "Disabling asynchronous scheduling by default for pooling model." |
| 1581 | ) |
| 1582 | self.scheduler_config.async_scheduling = False |
| 1583 | elif ( |
| 1584 | self.speculative_config is not None |
| 1585 | and self.speculative_config.method not in get_args(EagleModelTypes) |
| 1586 | and self.speculative_config.method not in get_args(NgramGPUTypes) |
| 1587 | and self.speculative_config.method != "draft_model" |
| 1588 | and self.speculative_config.method != "dspark" |
| 1589 | and self.speculative_config.method != "dflash" |
| 1590 | ): |
| 1591 | logger.warning_once( |
| 1592 | "Async scheduling not supported with %s-based " |
| 1593 | "speculative decoding and will be disabled.", |
| 1594 | self.speculative_config.method, |
| 1595 | ) |
| 1596 | self.scheduler_config.async_scheduling = False |
| 1597 | elif ( |
| 1598 | self.speculative_config is not None |
| 1599 | and self.speculative_config.disable_padded_drafter_batch |
| 1600 | ): |
| 1601 | logger.warning_once( |
| 1602 | "Async scheduling is not compatible with " |
| 1603 | "disable_padded_drafter_batch=True and will be disabled.", |
| 1604 | ) |
| 1605 | self.scheduler_config.async_scheduling = False |
| 1606 | elif not executor_supports_async_sched: |
| 1607 | logger.warning_once( |
| 1608 | "Async scheduling will be disabled because it is not supported " |
| 1609 | "with the `%s` distributed executor backend. ", |
| 1610 | executor_backend, |
| 1611 | ) |
| 1612 | self.scheduler_config.async_scheduling = False |
| 1613 | elif uses_rocm_deepep_ht_dbo: |
| 1614 | logger.warning_once( |
| 1615 | "Async scheduling is disabled for ROCm DeepEP " |
| 1616 | "high-throughput DBO because that combination can corrupt " |
| 1617 | "DP+EP generation accuracy." |
| 1618 | ) |
| 1619 | self.scheduler_config.async_scheduling = False |
| 1620 | elif ( |
| 1621 | self.parallel_config.pipeline_parallel_size > 1 |
| 1622 | and not self.use_v2_model_runner |
| 1623 | ): |
| 1624 | logger.warning_once( |
| 1625 | "Async scheduling is disabled because the V1 model runner " |
| 1626 | "does not support it with pipeline parallelism." |
| 1627 | ) |
| 1628 | self.scheduler_config.async_scheduling = False |
| 1629 | else: |
| 1630 | self.scheduler_config.async_scheduling = True |
| 1631 | |
| 1632 | if self.parallel_config.disable_nccl_for_dp_synchronization is None: |
| 1633 | if self.scheduler_config.async_scheduling: |
| 1634 | if self.parallel_config.data_parallel_size > 1 and ( |
| 1635 | self.model_config is None or self.model_config.is_moe |
| 1636 | ): |
| 1637 | logger.info_once( |
| 1638 | "Disabling NCCL for DP synchronization " |
| 1639 | "when using async scheduling.", |
| 1640 | ) |
| 1641 | self.parallel_config.disable_nccl_for_dp_synchronization = True |
| 1642 | else: |
| 1643 | self.parallel_config.disable_nccl_for_dp_synchronization = False |
| 1644 | |
| 1645 | if ( |
| 1646 | self.speculative_config is not None |
| 1647 | and self.scheduler_config.async_scheduling |
| 1648 | and self.model_config is not None |
| 1649 | and not self.model_config.disable_cascade_attn |
| 1650 | ): |
| 1651 | logger.warning_once( |
| 1652 | "Disabling cascade attention (not yet compatible with " |
| 1653 | "async speculative decoding).", |
| 1654 | ) |
| 1655 | self.model_config.disable_cascade_attn = True |
| 1656 | |
| 1657 | if ( |
| 1658 | self.observability_config.per_request_spec_decode_metrics != "none" |
| 1659 | and self.speculative_config is None |
| 1660 | ): |
| 1661 | raise ValueError( |
| 1662 | "--per-request-spec-decode-metrics requires speculative decoding " |
| 1663 | "to be enabled (via --speculative-config)." |
| 1664 | ) |
| 1665 | |
| 1666 | if ( |
| 1667 | self.model_config is not None |
| 1668 | and self.model_config.multimodal_config is not None |
| 1669 | and self.model_config.multimodal_config.mm_tensor_ipc == "torch_shm" |
| 1670 | and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD") != "spawn" |
| 1671 | ): |
| 1672 | raise ValueError( |
| 1673 | "torch_shm is known to fail without " |
| 1674 | "VLLM_WORKER_MULTIPROC_METHOD set to spawn" |
| 1675 | ) |
| 1676 | |
| 1677 | if ( |
| 1678 | self.model_config is not None |
| 1679 | and self.scheduler_config.enable_chunked_prefill |
| 1680 | and self.model_config.dtype == torch.float32 |
| 1681 | and current_platform.get_device_capability() == (7, 5) |
| 1682 | ): |
| 1683 | logger.warning_once( |
| 1684 | "Turing devices tensor cores do not support float32 matmul. " |
| 1685 | "To workaround this limitation, vLLM will set 'ieee' input " |
| 1686 | "precision for chunked prefill triton kernels." |
| 1687 | ) |
| 1688 | |
| 1689 | if self.model_config is not None and self.model_config.enforce_eager: |
| 1690 | self.compilation_config.mode = CompilationMode.NONE |
| 1691 | self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE |
| 1692 | if self.parallel_config.enable_fault_tolerance: |
| 1693 | # Keep JIT warmup: in-inference Triton compilation latency |
| 1694 | # spikes can delay peer-fault detection past its deadline. |
| 1695 | logger.warning_once( |
| 1696 | "Enforce eager set, disabling torch.compile and CUDAGraphs. " |
| 1697 | "This is equivalent to setting -cc.mode=none " |
| 1698 | "-cc.cudagraph_mode=none" |
| 1699 | ) |
| 1700 | else: |
| 1701 | logger.warning_once( |
| 1702 | "Enforce eager set, disabling torch.compile, CUDAGraphs, and " |
| 1703 | "JIT kernel warmup. This is equivalent to setting " |
| 1704 | "-cc.mode=none -cc.cudagraph_mode=none and " |
| 1705 | "--kernel_config.enable_jit_warmup=False" |
| 1706 | ) |
| 1707 | self.kernel_config.enable_jit_warmup = False |
| 1708 | |
| 1709 | if os.environ.get("TORCH_COMPILE_DISABLE") == "1": |
| 1710 | logger.warning_once( |
| 1711 | "TORCH_COMPILE_DISABLE is set, disabling torch.compile. " |
| 1712 | "This is equivalent to setting -cc.mode=none" |
| 1713 | ) |
| 1714 | self.compilation_config.mode = CompilationMode.NONE |
| 1715 | |
| 1716 | # TODO: This is a stopgap that turns sequence parallelism / async TP |
| 1717 | # off under batch invariance. Make the sequence-parallel reduce-scatter |
| 1718 | # path batch-invariant instead so they can stay enabled (#56370). |
| 1719 | # Sequence parallelism / async TP rewrite all_reduce + rms_norm into a |
| 1720 | # reduce-scatter whose reduction order depends on the batch, so they |
| 1721 | # are not batch-invariant. Decide this before the breakable-CUDA-graph |
| 1722 | # default below, which declines to auto-enable when they are on. |
| 1723 | pass_config = self.compilation_config.pass_config |
| 1724 | if envs.VLLM_BATCH_INVARIANT and ( |
| 1725 | pass_config.enable_sp or pass_config.fuse_gemm_comms |
| 1726 | ): |
| 1727 | logger.warning_once( |
| 1728 | "Disabling sequence parallelism and async TP " |
| 1729 | "(pass_config.enable_sp / fuse_gemm_comms) when " |
| 1730 | "VLLM_BATCH_INVARIANT is enabled: the reduce-scatter path " |
| 1731 | "is not batch-invariant (see vllm-project/vllm#56370)." |
| 1732 | ) |
| 1733 | pass_config.enable_sp = False |
| 1734 | pass_config.fuse_gemm_comms = False |
| 1735 | |
| 1736 | breakable_cudagraph_enabled = self._maybe_enable_breakable_cudagraph() |
| 1737 | |
| 1738 | if not breakable_cudagraph_enabled and ( |
| 1739 | self.compilation_config.backend == "eager" |
| 1740 | or ( |
| 1741 | self.compilation_config.mode is not None |
| 1742 | and self.compilation_config.mode != CompilationMode.VLLM_COMPILE |
| 1743 | ) |
| 1744 | ): |
| 1745 | logger.warning_once( |
| 1746 | "Inductor compilation was disabled by user settings, " |
| 1747 | "optimizations settings that are only active during " |
| 1748 | "inductor compilation will be ignored." |
| 1749 | ) |
| 1750 | |
| 1751 | def has_blocked_weights(): |
| 1752 | if self.quant_config is not None: |
| 1753 | if hasattr(self.quant_config, "weight_block_size"): |
| 1754 | return self.quant_config.weight_block_size is not None |
| 1755 | elif hasattr(self.quant_config, "has_blocked_weights"): |
| 1756 | return self.quant_config.has_blocked_weights() |
| 1757 | return False |
| 1758 | |
| 1759 | # Enable quant_fp8 CUDA ops (TODO disable in follow up) |
| 1760 | # On H100 the CUDA kernel is faster than |
| 1761 | # native implementation |
| 1762 | # https://github.com/vllm-project/vllm/issues/25094 |
| 1763 | if has_blocked_weights(): |
| 1764 | custom_ops = self.compilation_config.custom_ops |
| 1765 | if "-quant_fp8" not in custom_ops: |
| 1766 | custom_ops.append("+quant_fp8") |
| 1767 | |
| 1768 | current_platform.apply_config_platform_defaults(self) |
| 1769 | |
| 1770 | if self.compilation_config.mode is None: |
| 1771 | if self.optimization_level > OptimizationLevel.O0: |
| 1772 | self.compilation_config.mode = CompilationMode.VLLM_COMPILE |
| 1773 | else: |
| 1774 | self.compilation_config.mode = CompilationMode.NONE |
| 1775 | |
| 1776 | # By default, enable torch wrapping only when using custom Inductor lowering |
| 1777 | if self.compilation_config.ir_enable_torch_wrap is None: |
| 1778 | self.compilation_config.ir_enable_torch_wrap = ( |
| 1779 | self.compilation_config.mode == CompilationMode.VLLM_COMPILE |
| 1780 | and self.compilation_config.backend == "inductor" |
| 1781 | ) |
| 1782 | |
| 1783 | if all(s not in self.compilation_config.custom_ops for s in ("all", "none")): |
| 1784 | if ( |
| 1785 | self.compilation_config.backend == "inductor" |
| 1786 | and self.compilation_config.mode != CompilationMode.NONE |
| 1787 | ): |
| 1788 | self.compilation_config.custom_ops.append("none") |
| 1789 | else: |
| 1790 | self.compilation_config.custom_ops.append("all") |
| 1791 | |
| 1792 | # This populates IR op priorities, |
| 1793 | # must happen after compilation mode and backend are decided, |
| 1794 | # but before fusion defaults are applied as those may depend on op priority. |
| 1795 | self.kernel_config.set_platform_defaults(self) |
| 1796 | |
| 1797 | default_config = OPTIMIZATION_LEVEL_TO_CONFIG[self.optimization_level] |
| 1798 | self._apply_optimization_level_defaults(default_config) |
| 1799 | if self.kernel_config.enable_flashinfer_autotune is None: |
| 1800 | raise ValueError( |
| 1801 | "KernelConfig.enable_flashinfer_autotune must be set after applying " |
| 1802 | "optimization level defaults." |
| 1803 | ) |
| 1804 | |
| 1805 | self._maybe_disable_dynamic_sd_for_data_parallel() |
| 1806 | self._maybe_override_dynamic_sd_cudagraph_mode() |
| 1807 | |
| 1808 | if ( |
| 1809 | self.attention_config.hisparse_config is None |
| 1810 | and self.kv_transfer_config is not None |
| 1811 | and self.kv_transfer_config.has_connector("HiSparseConnector") |
| 1812 | ): |
| 1813 | self.attention_config.hisparse_config = HiSparseConfig() |
| 1814 | |
| 1815 | if self.attention_config.hisparse_config is not None: |
| 1816 | if not current_platform.is_cuda(): |
| 1817 | raise ValueError("HiSparse currently requires NVIDIA CUDA.") |
| 1818 | if self.parallel_config.pipeline_parallel_size > 1: |
| 1819 | raise ValueError("HiSparse does not support pipeline parallelism.") |
| 1820 | if self.parallel_config.decode_context_parallel_size > 1: |
| 1821 | raise ValueError( |
| 1822 | "HiSparse does not support decode context parallelism." |
| 1823 | ) |
| 1824 | if not self.scheduler_config.scheduler_reserve_full_isl: |
| 1825 | # Without it, async loads admitted against free host blocks can |
| 1826 | # each wait on host pages the others hold, and waiting requests |
| 1827 | # are never preempted to free them. |
| 1828 | raise ValueError( |
| 1829 | "HiSparse requires --scheduler-reserve-full-isl; remove " |
| 1830 | "--no-scheduler-reserve-full-isl." |
| 1831 | ) |
| 1832 | if self.model_config is not None and not hasattr( |
| 1833 | self.model_config.hf_config, "index_topk" |
| 1834 | ): |
| 1835 | raise ValueError( |
| 1836 | "HiSparse is only supported for DSA models with index_topk." |
| 1837 | ) |
| 1838 | if self.kv_transfer_config is not None and ( |
| 1839 | self.kv_transfer_config.kv_connector |
| 1840 | not in ( |
| 1841 | None, |
| 1842 | "NixlConnector", |
| 1843 | "MooncakeStoreConnector", |
| 1844 | "MultiConnector", |
| 1845 | ) |
| 1846 | ): |
| 1847 | logger.warning( |
| 1848 | "HiSparse host-resident KV is configured with connector " |
| 1849 | "%s. NixlConnector (GPU-staged host imports) and " |
| 1850 | "MooncakeStoreConnector (shared-store offload) are the " |
| 1851 | "validated paths; other connectors are treated as " |
| 1852 | "debug/fallback paths.", |
| 1853 | self.kv_transfer_config.kv_connector, |
| 1854 | ) |
| 1855 | |
| 1856 | self._normalize_piecewise_cudagraph_mode( |
| 1857 | breakable_cudagraph_enabled=breakable_cudagraph_enabled |
| 1858 | ) |
| 1859 | |
| 1860 | # async tp is built on top of sequence parallelism and requires it. |
| 1861 | pass_config = self.compilation_config.pass_config |
| 1862 | if pass_config.fuse_gemm_comms: |
| 1863 | pass_config.enable_sp = True |
| 1864 | if pass_config.enable_sp: |
| 1865 | if self.parallel_config.tensor_parallel_size == 1: |
| 1866 | logger.warning_once("Sequence Parallelism requires TP>1, disabling") |
| 1867 | pass_config.enable_sp = False |
| 1868 | pass_config.fuse_gemm_comms = False |
| 1869 | else: |
| 1870 | if pass_config.sp_min_token_num is None: |
| 1871 | from vllm.compilation.passes.fusion.sequence_parallelism import ( |
| 1872 | get_sequence_parallelism_threshold, |
| 1873 | ) |
| 1874 | |
| 1875 | tp_size = self.parallel_config.tensor_parallel_size |
| 1876 | hidden_size = self.model_config.get_hidden_size() |
| 1877 | assert isinstance(self.model_config.dtype, torch.dtype) |
| 1878 | element_size = self.model_config.dtype.itemsize |
| 1879 | pass_config.sp_min_token_num = get_sequence_parallelism_threshold( |
| 1880 | hidden_size, tp_size, element_size |
| 1881 | ) |
| 1882 | |
| 1883 | if pass_config.sp_min_token_num is None: |
| 1884 | logger.warning_once( |
| 1885 | "Model hidden_size too small for the SP " |
| 1886 | "threshold heuristic, disabling. To force SP, " |
| 1887 | "set pass_config.sp_min_token_num manually." |
| 1888 | ) |
| 1889 | pass_config.enable_sp = False |
| 1890 | pass_config.fuse_gemm_comms = False |
| 1891 | |
| 1892 | from vllm.utils.torch_utils import HAS_OPAQUE_TYPE |
| 1893 | |
| 1894 | if HAS_OPAQUE_TYPE: |
| 1895 | # On torch >= 2.11 the hoisted OpaqueObject approach supersedes |
| 1896 | # fast_moe_cold_start, so force it off. |
| 1897 | self.compilation_config.fast_moe_cold_start = False |
| 1898 | elif self.compilation_config.fast_moe_cold_start is None: |
| 1899 | # resolve default behavior: try to be as safe as possible |
| 1900 | # this config is unsafe if any spec decoding draft model has a MOE. |
| 1901 | # We'll conservatively turn it off if we see spec decoding. |
| 1902 | self.compilation_config.fast_moe_cold_start = ( |
| 1903 | self.speculative_config is None |
| 1904 | ) |
| 1905 | |
| 1906 | self._set_max_num_scheduled_tokens() |
| 1907 | |
| 1908 | if current_platform.support_static_graph_mode(): |
| 1909 | # if cudagraph_mode has full cudagraphs, we need to check support |
| 1910 | if model_config := self.model_config: |
| 1911 | if ( |
| 1912 | self.compilation_config.cudagraph_mode.has_full_cudagraphs() |
| 1913 | and model_config.pooler_config is not None |
| 1914 | ): |
| 1915 | logger.warning_once( |
| 1916 | "Pooling models do not support full cudagraphs. " |
| 1917 | "Overriding cudagraph_mode to PIECEWISE." |
| 1918 | ) |
| 1919 | self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE |
| 1920 | elif ( |
| 1921 | model_config.is_encoder_decoder |
| 1922 | and self.compilation_config.cudagraph_mode |
| 1923 | not in (CUDAGraphMode.NONE, CUDAGraphMode.FULL_DECODE_ONLY) |
| 1924 | ): |
| 1925 | logger.info_once( |
| 1926 | "Encoder-decoder models do not support %s. " |
| 1927 | "Overriding cudagraph_mode to FULL_DECODE_ONLY.", |
| 1928 | self.compilation_config.cudagraph_mode.name, |
| 1929 | ) |
| 1930 | self.compilation_config.cudagraph_mode = ( |
| 1931 | CUDAGraphMode.FULL_DECODE_ONLY |
| 1932 | ) |
| 1933 | |
| 1934 | # Check if KV connector requires PIECEWISE mode for CUDA graphs |
| 1935 | if ( |
| 1936 | self.kv_transfer_config is not None |
| 1937 | and self.kv_transfer_config.is_kv_transfer_instance |
| 1938 | and self.compilation_config.cudagraph_mode.has_full_cudagraphs() |
| 1939 | ): |
| 1940 | # Lazy import to avoid circular dependencies |
| 1941 | from vllm.distributed.kv_transfer.kv_connector.factory import ( |
| 1942 | KVConnectorFactory, |
| 1943 | ) |
| 1944 | |
| 1945 | connector_cls = KVConnectorFactory.get_connector_class( |
| 1946 | self.kv_transfer_config |
| 1947 | ) |
| 1948 | if connector_cls.requires_piecewise_for_cudagraph( |
| 1949 | self.kv_transfer_config.kv_connector_extra_config |
| 1950 | ): |
| 1951 | logger.warning_once( |
| 1952 | "KV connector %s requires PIECEWISE CUDA graph mode " |
| 1953 | "due to layerwise async operations that cannot be " |
| 1954 | "captured in CUDA graphs. " |
| 1955 | "Overriding cudagraph_mode from %s to PIECEWISE.", |
| 1956 | connector_cls.__name__, |
| 1957 | self.compilation_config.cudagraph_mode.name, |
| 1958 | ) |
| 1959 | self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE |
| 1960 | |
| 1961 | # disable cudagraph when enforce eager execution |
| 1962 | if self.model_config is not None and self.model_config.enforce_eager: |
| 1963 | logger.info_once("Cudagraph is disabled under eager mode") |
| 1964 | self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE |
| 1965 | # override related settings when enforce eager |
| 1966 | self.compilation_config.max_cudagraph_capture_size = 0 |
| 1967 | self.compilation_config.cudagraph_capture_sizes = [] |
| 1968 | else: |
| 1969 | self.compilation_config.cudagraph_num_of_warmups = 1 |
| 1970 | |
| 1971 | self._set_cudagraph_sizes() |
| 1972 | |
| 1973 | else: |
| 1974 | self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE |
| 1975 | |
| 1976 | if self.cache_config.kv_sharing_fast_prefill: |
| 1977 | if ( |
| 1978 | self.speculative_config is not None |
| 1979 | and self.speculative_config.use_eagle() |
| 1980 | ): |
| 1981 | raise ValueError( |
| 1982 | "Fast prefill optimization for KV sharing is not " |
| 1983 | "compatible with EAGLE as EAGLE requires correct logits " |
| 1984 | "for all tokens while fast prefill gives incorrect logits " |
| 1985 | "for prompt tokens." |
| 1986 | ) |
| 1987 | |
| 1988 | logger.warning_once( |
| 1989 | "--kv-sharing-fast-prefill requires changes on model side for " |
| 1990 | "correctness and to realize prefill savings." |
| 1991 | ) |
| 1992 | |
| 1993 | if ( |
| 1994 | self.model_config |
| 1995 | and self.model_config.architecture == "WhisperForConditionalGeneration" |
| 1996 | and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD") != "spawn" |
| 1997 | ): |
| 1998 | logger.warning_once( |
| 1999 | "Whisper is known to have issues with " |
| 2000 | "forked workers. If startup is hanging, " |
| 2001 | "try setting 'VLLM_WORKER_MULTIPROC_METHOD' " |
| 2002 | "to 'spawn'." |
| 2003 | ) |
| 2004 | |
| 2005 | if ( |
| 2006 | self.kv_events_config is not None |
| 2007 | and self.kv_events_config.enable_kv_cache_events |
| 2008 | and not self.cache_config.enable_prefix_caching |
| 2009 | ): |
| 2010 | logger.warning_once( |
| 2011 | "KV cache events are on, but prefix caching is not enabled. " |
| 2012 | "Use --enable-prefix-caching to enable." |
| 2013 | ) |
| 2014 | if ( |
| 2015 | self.kv_events_config is not None |
| 2016 | and self.kv_events_config.publisher != "null" |
| 2017 | and not self.kv_events_config.enable_kv_cache_events |
| 2018 | ): |
| 2019 | logger.warning_once( |
| 2020 | "KV cache events are disabled, " |
| 2021 | "but the scheduler is configured to publish them. " |
| 2022 | "Modify KVEventsConfig.enable_kv_cache_events " |
| 2023 | "to True to enable." |
| 2024 | ) |
| 2025 | current_platform.check_and_update_config(self) |
| 2026 | |
| 2027 | # After the platform hook, which has the last word on async scheduling. |
| 2028 | if ( |
| 2029 | self.diffusion_config is not None |
| 2030 | and self.scheduler_config.scheduler_cls is None |
| 2031 | ): |
| 2032 | scheduler_name = ( |
| 2033 | "DiffusionAsyncScheduler" |
| 2034 | if self.scheduler_config.async_scheduling |
| 2035 | else "DiffusionScheduler" |
| 2036 | ) |
| 2037 | self.scheduler_config.scheduler_cls = ( |
| 2038 | f"vllm.v1.core.sched.diffusion_scheduler.{scheduler_name}" |
| 2039 | ) |
| 2040 | |
| 2041 | self._normalize_piecewise_cudagraph_mode( |
| 2042 | breakable_cudagraph_enabled=breakable_cudagraph_enabled |
| 2043 | ) |
| 2044 | |
| 2045 | self._resolve_mm_embedding_inputs() |
| 2046 | self._resolve_mm_processor_device() |
| 2047 | self._resolve_mm_video_decode_device() |
| 2048 | self._validate_mm_processor_device() |
| 2049 | |
| 2050 | if self.use_v2_model_runner: |
| 2051 | self._disable_cudagraphs_for_v2_stock_torch_compile() |
| 2052 | self._validate_v2_model_runner() |
| 2053 | else: |
| 2054 | self._validate_v1_model_runner() |
| 2055 | |
| 2056 | self._validate_profiler_config() |
| 2057 | self._validate_batch_sharded_sampling() |
| 2058 | self._validate_adaptive_verification() |
| 2059 | |
| 2060 | # Re-compute compile ranges after platform-specific config updates |
| 2061 | # (e.g., XPU may lower max_num_batched_tokens when MLA is enabled) |
| 2062 | self._set_compile_ranges() |
| 2063 | |
| 2064 | if self.parallel_config.all2all_backend == "moonep": |
| 2065 | if ( |
| 2066 | self.model_config is not None |
| 2067 | and self.model_config.quantization is not None |
| 2068 | ): |
| 2069 | raise ValueError( |
| 2070 | "The moonep all2all backend currently supports unquantized " |
| 2071 | "BF16 models only; got " |
| 2072 | f"quantization={self.model_config.quantization!r}. Use a " |
| 2073 | "different --all2all-backend for quantized models." |
| 2074 | ) |
| 2075 | if ( |
| 2076 | self.model_config is not None |
| 2077 | and self.model_config.dtype != torch.bfloat16 |
| 2078 | ): |
| 2079 | raise ValueError( |
| 2080 | "The moonep all2all backend currently supports BF16 models " |
| 2081 | f"only; got dtype={self.model_config.dtype}. Use a " |
| 2082 | "different --all2all-backend or --dtype bfloat16." |
| 2083 | ) |
| 2084 | if self.parallel_config.enable_eplb: |
| 2085 | raise ValueError( |
| 2086 | "The moonep all2all backend does not support EPLB yet: " |
| 2087 | "EPLB rearranges expert parameters in a layout MoonEP's " |
| 2088 | "replicated [E+B] weights do not follow. Disable " |
| 2089 | "--enable-eplb or use a different --all2all-backend." |
| 2090 | ) |
| 2091 | if self.parallel_config.expert_placement_strategy != "linear": |
| 2092 | raise ValueError( |
| 2093 | "The moonep all2all backend requires linear expert " |
| 2094 | "placement: its load-time all-gather assumes each rank " |
| 2095 | "holds a contiguous chunk of the global expert range. Got " |
| 2096 | "--expert-placement-strategy " |
| 2097 | f"{self.parallel_config.expert_placement_strategy!r}." |
| 2098 | ) |
| 2099 | # Enforced here rather than in set_splitting_ops_for_v1 so it |
| 2100 | # holds for every compilation mode, and keyed on use_all2all so |
| 2101 | # PCP/SP-only topologies are covered too: MoonEP dispatch/combine |
| 2102 | # are eager-only and must not be captured. |
| 2103 | if ( |
| 2104 | self.parallel_config.use_all2all |
| 2105 | and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE |
| 2106 | ): |
| 2107 | logger.info( |
| 2108 | "MoonEP: Disabling CUDA Graphs since the MoonEP " |
| 2109 | "integration is currently eager-only." |
| 2110 | ) |
| 2111 | self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE |
| 2112 | |
| 2113 | # Do this after all the updates to compilation_config.mode |
| 2114 | effective_dp_size = ( |
| 2115 | self.parallel_config.data_parallel_size |
| 2116 | if self.model_config is None or self.model_config.is_moe |
| 2117 | else 1 |
| 2118 | ) |
| 2119 | self.compilation_config.set_splitting_ops_for_v1( |
| 2120 | all2all_backend=self.parallel_config.all2all_backend, |
| 2121 | data_parallel_size=effective_dp_size, |
| 2122 | ) |
| 2123 | |
| 2124 | if self.compilation_config.pass_config.enable_sp: |
| 2125 | # With pipeline parallelism, native rms norm tracing errors due to |
| 2126 | # incorrect residual shape. |
| 2127 | # Use custom rms norm to unblock. In the future, |
| 2128 | # the pass will operate on higher-level IR to avoid the issue. |
| 2129 | # TODO: https://github.com/vllm-project/vllm/issues/27894 |
| 2130 | if self.compilation_config.mode != CompilationMode.VLLM_COMPILE: |
| 2131 | logger.warning_once( |
| 2132 | "Sequence parallelism is enabled, but running in wrong " |
| 2133 | "vllm compile mode: %s.", |
| 2134 | self.compilation_config.mode, |
| 2135 | ) |
| 2136 | |
| 2137 | if self.parallel_config.pipeline_parallel_size > 1: |
| 2138 | if "-rms_norm" not in self.compilation_config.custom_ops: |
| 2139 | self.compilation_config.custom_ops.append("+rms_norm") |
| 2140 | else: |
| 2141 | logger.warning_once( |
| 2142 | "Sequence parallelism not supported with " |
| 2143 | "native rms_norm when using %s, " |
| 2144 | "this will likely lead to an error.", |
| 2145 | "pipeline parallelism", |
| 2146 | ) |
| 2147 | |
| 2148 | # final check of cudagraph mode after all possible updates |
| 2149 | if current_platform.is_cuda_alike(): |
| 2150 | if ( |
| 2151 | self.compilation_config.cudagraph_mode.has_full_cudagraphs() |
| 2152 | and self.model_config is not None |
| 2153 | and not self.model_config.disable_cascade_attn |
| 2154 | and not self.compilation_config.cudagraph_mode.has_piecewise_cudagraphs() # noqa: E501 |
| 2155 | ): |
| 2156 | logger.warning_once( |
| 2157 | "No piecewise cudagraph for executing cascade attention. " |
| 2158 | "Will fall back to eager execution if a batch runs into " |
| 2159 | "cascade attentions." |
| 2160 | ) |
| 2161 | |
| 2162 | if self.compilation_config.cudagraph_mode.requires_piecewise_compilation(): |
| 2163 | assert ( |
| 2164 | self.compilation_config.mode == CompilationMode.VLLM_COMPILE |
| 2165 | or envs.VLLM_USE_BREAKABLE_CUDAGRAPH |
| 2166 | ), ( |
| 2167 | "Compilation mode should be CompilationMode.VLLM_COMPILE " |
| 2168 | "when cudagraph_mode piecewise cudagraphs is used, " |
| 2169 | f"cudagraph_mode={self.compilation_config.cudagraph_mode}" |
| 2170 | ) |
| 2171 | if ( |
| 2172 | self.model_config |
| 2173 | and envs.VLLM_BATCH_INVARIANT |
| 2174 | and not self.model_config.disable_cascade_attn |
| 2175 | ): |
| 2176 | self.model_config.disable_cascade_attn = True |
| 2177 | logger.warning_once( |
| 2178 | "Disabling cascade attention when VLLM_BATCH_INVARIANT is enabled.", |
| 2179 | ) |
| 2180 | |
| 2181 | if self.parallel_config.use_ubatching: |
| 2182 | a2a_backend = self.parallel_config.all2all_backend |
| 2183 | assert a2a_backend in [ |
| 2184 | "deepep_low_latency", |
| 2185 | "deepep_high_throughput", |
| 2186 | "nixl_ep", |
| 2187 | ], ( |
| 2188 | "Microbatching currently only supports the deepep_low_latency, " |
| 2189 | "deepep_high_throughput, and nixl_ep all2all backends. " |
| 2190 | f"{a2a_backend} is not supported. To fix use " |
| 2191 | "--all2all-backend=deepep_low_latency, " |
| 2192 | "--all2all-backend=deepep_high_throughput, or " |
| 2193 | "--all2all-backend=nixl_ep and install the matching kernels." |
| 2194 | ) |
| 2195 | |
| 2196 | if not self.model_config.disable_cascade_attn: |
| 2197 | self.model_config.disable_cascade_attn = True |
| 2198 | logger.warning_once("Disabling cascade attention when DBO is enabled.") |
| 2199 | |
| 2200 | if not self.instance_id: |
| 2201 | self.instance_id = random_uuid()[:5] |
| 2202 | |
| 2203 | if self.reasoning_config is not None and self.model_config is not None: |
| 2204 | self.reasoning_config.initialize_token_ids(self.model_config) |
| 2205 | if not self.reasoning_config.enabled: |
| 2206 | logger.warning_once( |
| 2207 | "Auto-initialization of reasoning token IDs failed. " |
| 2208 | "Please check whether your reasoning parser has implemented " |
| 2209 | "the `reasoning_start_str` and `reasoning_end_str`." |
| 2210 | ) |
| 2211 | |
| 2212 | # Resolve kv_offloading-derived connector name into kv_transfer_config |
| 2213 | # before the HMA check below, which inspects the connector class. |
| 2214 | self._post_init_kv_transfer_config() |
| 2215 | self._verify_aux_output_compatibility() |
| 2216 | |
| 2217 | # Hybrid KV cache manager (HMA) runtime rules: |
| 2218 | # - Explicit enable (--no-disable-kv-cache-manager): error if runtime |
| 2219 | # disables it |
| 2220 | # - No preference: auto-disable for unsupported features or connector configs |
| 2221 | # - Explicit disable (--disable-kv-cache-manager): always respect it |
| 2222 | need_disable_hybrid_kv_cache_manager = False |
| 2223 | # logger should only print warning message for hybrid models. As we |
| 2224 | # can't know whether the model is hybrid or not now, so we don't log |
| 2225 | # warning message here and will log it later. |
| 2226 | if not current_platform.support_hybrid_kv_cache(): |
| 2227 | # Hybrid KV cache manager is not supported on non-GPU platforms. |
| 2228 | need_disable_hybrid_kv_cache_manager = True |
| 2229 | if ( |
| 2230 | self.model_config is not None |
| 2231 | and self.model_config.attention_chunk_size is not None |
| 2232 | ): |
| 2233 | if ( |
| 2234 | self.speculative_config is not None |
| 2235 | and self.speculative_config.use_eagle() |
| 2236 | ): |
| 2237 | # Hybrid KV cache manager is not yet supported with chunked |
| 2238 | # local attention + eagle. |
| 2239 | need_disable_hybrid_kv_cache_manager = True |
| 2240 | elif not envs.VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: |
| 2241 | logger.warning( |
| 2242 | "There is a latency regression when using chunked local" |
| 2243 | " attention with the hybrid KV cache manager. Disabling" |
| 2244 | " it, by default. To enable it, set the environment " |
| 2245 | "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE=1." |
| 2246 | ) |
| 2247 | # Hybrid KV cache manager is not yet supported with chunked |
| 2248 | # local attention. |
| 2249 | need_disable_hybrid_kv_cache_manager = True |
| 2250 | |
| 2251 | if self.scheduler_config.disable_hybrid_kv_cache_manager is None: |
| 2252 | # Auto-disable HMA only when the connector config does not support it. |
| 2253 | if self.kv_transfer_config is not None: |
| 2254 | from vllm.distributed.kv_transfer.kv_connector.factory import ( |
| 2255 | KVConnectorFactory, |
| 2256 | ) |
| 2257 | |
| 2258 | if not KVConnectorFactory.supports_hma_config(self.kv_transfer_config): |
| 2259 | need_disable_hybrid_kv_cache_manager = True |
| 2260 | logger.warning( |
| 2261 | "Turning off hybrid kv cache manager because " |
| 2262 | "`--kv-transfer-config` selects a KV connector that " |
| 2263 | "does not support it. Impact: hybrid SSM models " |
| 2264 | "(e.g. Jamba, Bamba) require HMA and will fail at " |
| 2265 | "startup without it; models with sliding window " |
| 2266 | "attention will run with reduced performance. " |
| 2267 | "To add HMA support to a KV connector, subclass " |
| 2268 | "`SupportsHMA` defined in kv_connector/v1/base.py " |
| 2269 | "(for MultiConnector, all child connectors must " |
| 2270 | "support HMA)." |
| 2271 | ) |
| 2272 | self.scheduler_config.disable_hybrid_kv_cache_manager = ( |
| 2273 | need_disable_hybrid_kv_cache_manager |
| 2274 | ) |
| 2275 | elif ( |
| 2276 | self.scheduler_config.disable_hybrid_kv_cache_manager is False |
| 2277 | and need_disable_hybrid_kv_cache_manager |
| 2278 | ): |
| 2279 | raise ValueError( |
| 2280 | "Hybrid KV cache manager was explicitly enabled but is not " |
| 2281 | "supported in this configuration. Consider omitting the " |
| 2282 | "--no-disable-hybrid-kv-cache-manager flag to let vLLM decide" |
| 2283 | " automatically." |
| 2284 | ) |
| 2285 | |
| 2286 | if self.scheduler_config.disable_hybrid_kv_cache_manager is None: |
| 2287 | # Default to enable HMA if not explicitly disabled by user or logic above. |
| 2288 | self.scheduler_config.disable_hybrid_kv_cache_manager = False |
| 2289 | |
| 2290 | if ( |
| 2291 | self.attention_config.hisparse_config is not None |
| 2292 | and self.scheduler_config.disable_hybrid_kv_cache_manager |
| 2293 | ): |
| 2294 | raise ValueError( |
| 2295 | "HiSparse requires the hybrid KV cache manager; remove " |
| 2296 | "--disable-hybrid-kv-cache-manager or use connectors that " |
| 2297 | "support HMA." |
| 2298 | ) |
| 2299 | |
| 2300 | if self.compilation_config.debug_dump_path: |
| 2301 | self.compilation_config.debug_dump_path = ( |
| 2302 | self.compilation_config.debug_dump_path.absolute().expanduser() |
| 2303 | ) |
| 2304 | if envs.VLLM_DEBUG_DUMP_PATH is not None: |
| 2305 | env_path = Path(envs.VLLM_DEBUG_DUMP_PATH).absolute().expanduser() |
| 2306 | if self.compilation_config.debug_dump_path: |
| 2307 | logger.warning( |
| 2308 | "Config-specified debug dump path is overridden" |
| 2309 | " by VLLM_DEBUG_DUMP_PATH to %s", |
| 2310 | env_path, |
| 2311 | ) |
| 2312 | self.compilation_config.debug_dump_path = env_path |
| 2313 | |
| 2314 | # Enable quant_fp8 CUDA ops (TODO disable in follow up) |
| 2315 | # On H100 the CUDA kernel is faster than |
| 2316 | # native implementation |
| 2317 | # https://github.com/vllm-project/vllm/issues/25094 |
| 2318 | if has_blocked_weights(): |
| 2319 | custom_ops = self.compilation_config.custom_ops |
| 2320 | if "-quant_fp8" not in custom_ops: |
| 2321 | custom_ops.append("+quant_fp8") |
| 2322 | |
| 2323 | self._verify_kv_transfer_compat() |
| 2324 | # Log the custom passes that are enabled |
| 2325 | self.compilation_config.pass_config.log_enabled_passes() |
| 2326 | |
| 2327 | def update_sizes_for_sequence_parallelism(self, possible_sizes: list) -> list: |
| 2328 | # remove the sizes that not multiple of tp_size when |
| 2329 | # enable sequence parallelism |
| 2330 | removed_sizes = [ |
| 2331 | size |
| 2332 | for size in possible_sizes |
| 2333 | if size % self.parallel_config.tensor_parallel_size != 0 |
| 2334 | ] |
| 2335 | if removed_sizes: |
| 2336 | logger.warning( |
| 2337 | "Batch sizes %s are removed because they are not " |
| 2338 | "multiple of tp_size %d when " |
| 2339 | "sequence parallelism is enabled", |
| 2340 | removed_sizes, |
| 2341 | self.parallel_config.tensor_parallel_size, |
| 2342 | ) |
| 2343 | |
| 2344 | return [ |
| 2345 | size |
| 2346 | for size in possible_sizes |
| 2347 | if size % self.parallel_config.tensor_parallel_size == 0 |
| 2348 | ] |
| 2349 | |
| 2350 | def _set_max_num_scheduled_tokens(self): |
| 2351 | """In most cases, the scheduler may schedule a batch with as many tokens as the |
| 2352 | worker is configured to handle. |
| 2353 | """ |
| 2354 | if self.speculative_config is not None: |
| 2355 | scheduled_token_delta = ( |
| 2356 | self.speculative_config.max_num_new_slots_for_drafting |
| 2357 | ) |
| 2358 | max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens |
| 2359 | if self.scheduler_config.max_num_scheduled_tokens is None: |
| 2360 | self.scheduler_config.max_num_scheduled_tokens = max_num_batched_tokens |
| 2361 | |
| 2362 | if self.scheduler_config.max_num_scheduled_tokens <= 0: |
| 2363 | raise ValueError( |
| 2364 | "max_num_scheduled_tokens is set to" |
| 2365 | f" {self.scheduler_config.max_num_scheduled_tokens} based on" |
| 2366 | " the speculative decoding settings, which does not allow" |
| 2367 | " any tokens to be scheduled. Increase max_num_batched_tokens" |
| 2368 | " to accommodate the additional draft token slots, or decrease" |
| 2369 | " num_speculative_tokens." |
| 2370 | ) |
| 2371 | if self.scheduler_config.max_num_scheduled_tokens < 8192: |
| 2372 | logger.warning_once( |
| 2373 | "max_num_scheduled_tokens is set to" |
| 2374 | f" {self.scheduler_config.max_num_scheduled_tokens} based on" |
| 2375 | " the speculative decoding settings. This may lead to suboptimal" |
| 2376 | " performance. Consider increasing max_num_batched_tokens to" |
| 2377 | " accommodate the additional draft token slots, or decrease" |
| 2378 | " num_speculative_tokens.", |
| 2379 | ) |
| 2380 | |
| 2381 | if max_num_batched_tokens <= scheduled_token_delta: |
| 2382 | raise ValueError( |
| 2383 | "VllmConfig does not have enough slots to schedule a token and" |
| 2384 | " support the speculative decoding settings." |
| 2385 | f" Got {max_num_batched_tokens=} and {scheduled_token_delta=}." |
| 2386 | ) |
| 2387 | |
| 2388 | def _set_cudagraph_sizes(self): |
| 2389 | """VLLM defines the default candidate list of batch sizes for CUDA graph |
| 2390 | capture as: |
| 2391 | |
| 2392 | ```python |
| 2393 | default_max_graph_size = 1024 if is_data_center_blackwell else 512 |
| 2394 | decode_query_len = self.uniform_decode_query_len |
| 2395 | max_graph_size = min( |
| 2396 | max_num_seqs * decode_query_len * 2, default_max_graph_size |
| 2397 | ) |
| 2398 | # 1, 2, 4, then multiples of 8 up to 256 and then multiples of 16 |
| 2399 | # up to max_graph_size |
| 2400 | cudagraph_capture_sizes = [1, 2, 4] + list(range(8, 256, 8)) + list( |
| 2401 | range(256, max_graph_size + 1, 16)) |
| 2402 | |
| 2403 | `max_num_batched_tokens` is also appended to the list if it fits |
| 2404 | within `max_cudagraph_capture_size`, so the max batch size is captured |
| 2405 | even when off-stride. Uniform decode sizes are appended when they fit |
| 2406 | within the platform's default capture ceiling, since they need not land |
| 2407 | on an 8- or 16-token stride. |
| 2408 | |
| 2409 | In the end, `vllm_config.compilation_config.cudagraph_capture_sizes` |
| 2410 | will be the final sizes to capture cudagraph (in ascending order). |
| 2411 | |
| 2412 | These sizes are used to capture and reuse CUDA graphs for |
| 2413 | performance-critical paths (e.g., decoding). Capturing enables |
| 2414 | significantly faster kernel dispatch by avoiding Python overhead. The |
| 2415 | list is then filtered based on `max_num_batched_tokens` (e.g., 8192 on |
| 2416 | most GPUs), which controls the total allowed number of tokens in a |
| 2417 | batch. Since each sequence may have a variable number of tokens, the |
| 2418 | maximum usable batch size will depend on actual sequence lengths. |
| 2419 | |
| 2420 | Example: |
| 2421 | With `max_num_batched_tokens = 8192`, and typical sequences |
| 2422 | averaging ~32 tokens, most practical batch sizes fall below 256. |
| 2423 | However, the system will still allow capture sizes up to the |
| 2424 | platform default if shape and memory permit. |
| 2425 | |
| 2426 | Note: |
| 2427 | If users explicitly specify cudagraph capture sizes in the |
| 2428 | compilation config, those will override this default logic. |
| 2429 | At runtime: |
| 2430 | |
| 2431 | - If batch size <= one of the `cudagraph_capture_sizes`, the closest |
| 2432 | padded CUDA graph will be used. |
| 2433 | - If batch size > largest `cudagraph_capture_sizes`, cudagraph will |
| 2434 | not be used. |
| 2435 | |
| 2436 | """ |
| 2437 | if ( |
| 2438 | self.model_config is not None |
| 2439 | and not self.model_config.enforce_eager |
| 2440 | and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE |
| 2441 | ): |
| 2442 | # determine the initial max_cudagraph_capture_size |
| 2443 | max_cudagraph_capture_size = ( |
| 2444 | self.compilation_config.max_cudagraph_capture_size |
| 2445 | ) |
| 2446 | # Decode sizes to cover, in tokens. Populated only when the default |
| 2447 | # is computed here, so an explicit capture range is left exactly as |
| 2448 | # configured. |
| 2449 | uniform_decode_sizes: list[int] = [] |
| 2450 | if max_cudagraph_capture_size is None: |
| 2451 | from vllm.platforms import current_platform |
| 2452 | |
| 2453 | default_max_graph_size = ( |
| 2454 | 1024 if current_platform.is_device_capability_family(100) else 512 |
| 2455 | ) |
| 2456 | decode_query_len = self.uniform_decode_query_len |
| 2457 | max_num_seqs = self.scheduler_config.max_num_seqs |
| 2458 | max_cudagraph_capture_size = min( |
| 2459 | max_num_seqs * decode_query_len * 2, default_max_graph_size |
| 2460 | ) |
| 2461 | if decode_query_len > 1: |
| 2462 | # A uniform decode batch is decode_query_len tokens per |
| 2463 | # request, so the widest one is far outside this ceiling. |
| 2464 | # Coverage comes from appending the decode sizes rather than |
| 2465 | # extending the token-strided grid. Extending that grid to |
| 2466 | # the widest decode size produces 581 sizes at |
| 2467 | # max_num_seqs=512 and 16 draft tokens, versus 100 with the |
| 2468 | # request-count grid. |
| 2469 | # |
| 2470 | # The grid would not buy decode coverage anyway. Dispatch |
| 2471 | # requires an exact multiple of decode_query_len, so a |
| 2472 | # token-strided entry is only usable when it happens to be |
| 2473 | # one; at query length 17 a captured 560 rounds to 561 and |
| 2474 | # is rejected. Scaling a request-count grid keeps every |
| 2475 | # entry usable and the count comparable to the non- |
| 2476 | # speculative case. |
| 2477 | def request_counts(max_reqs: int) -> list[int]: |
| 2478 | # At most the platform default number of requests, |
| 2479 | # mirroring the one-token-per-request decode ceiling. |
| 2480 | max_reqs = min(max_reqs, default_max_graph_size) |
| 2481 | counts = [n for n in (1, 2, 4) if n <= max_reqs] |
| 2482 | counts += list(range(8, min(max_reqs + 1, 256), 8)) |
| 2483 | counts += list(range(256, max_reqs + 1, 16)) |
| 2484 | return sorted(set(counts + [max_reqs])) |
| 2485 | |
| 2486 | # Dynamic speculative decoding picks the draft width from |
| 2487 | # the batch size, so a decode step is only uniform within a |
| 2488 | # tier and each tier needs its own sizes. Scaling by the |
| 2489 | # widest one alone leaves the narrower tiers short: the |
| 2490 | # manager rounds a capture size up to a multiple of the |
| 2491 | # tier's query length and drops it once the implied request |
| 2492 | # count exceeds max_num_seqs, so at query length 3 sizes |
| 2493 | # built from 17 stop covering at 227 of 256 requests. |
| 2494 | decode_tiers = [(decode_query_len, max_num_seqs)] |
| 2495 | speculative_config = self.speculative_config |
| 2496 | if ( |
| 2497 | speculative_config is not None |
| 2498 | and speculative_config.uses_dynamic_speculative_decoding() |
| 2499 | ): |
| 2500 | from vllm.v1.spec_decode.dynamic.utils import ( |
| 2501 | build_dynamic_sd_schedule_lookup, |
| 2502 | ) |
| 2503 | |
| 2504 | schedule = ( |
| 2505 | speculative_config.num_speculative_tokens_per_batch_size |
| 2506 | ) |
| 2507 | assert schedule is not None |
| 2508 | # Read the tiers off the dense lookup the scheduler |
| 2509 | # runs on, so the clamp against num_speculative_tokens |
| 2510 | # and the carry-forward through gaps and the tail |
| 2511 | # cannot drift from it. Validation lives elsewhere; an |
| 2512 | # invalid schedule keeps the single-tier default. |
| 2513 | try: |
| 2514 | dense_schedule = build_dynamic_sd_schedule_lookup( |
| 2515 | schedule, |
| 2516 | vllm_max_batch_size=max_num_seqs, |
| 2517 | vllm_num_speculative_tokens=self.num_speculative_tokens, |
| 2518 | ) |
| 2519 | except ValueError: |
| 2520 | pass |
| 2521 | else: |
| 2522 | # Ascending batch size, so the last write per |
| 2523 | # query length is the widest batch running at it. |
| 2524 | widest_batch: dict[int, int] = {} |
| 2525 | for batch_size, num_spec in enumerate( |
| 2526 | dense_schedule[1:], start=1 |
| 2527 | ): |
| 2528 | widest_batch[num_spec + 1] = batch_size |
| 2529 | decode_tiers = list(widest_batch.items()) |
| 2530 | |
| 2531 | uniform_decode_sizes = sorted( |
| 2532 | { |
| 2533 | n * query_len |
| 2534 | for query_len, tier_max_reqs in decode_tiers |
| 2535 | for n in request_counts(tier_max_reqs) |
| 2536 | if n * query_len <= max_cudagraph_capture_size |
| 2537 | } |
| 2538 | ) |
| 2539 | elif max_num_seqs <= max_cudagraph_capture_size: |
| 2540 | uniform_decode_sizes = [max_num_seqs] |
| 2541 | max_num_tokens = self.scheduler_config.max_num_batched_tokens |
| 2542 | max_cudagraph_capture_size = min(max_num_tokens, max_cudagraph_capture_size) |
| 2543 | |
| 2544 | assert max_cudagraph_capture_size >= 1, ( |
| 2545 | "Maximum cudagraph size should be greater than or equal to 1 " |
| 2546 | "when using cuda graph." |
| 2547 | ) |
| 2548 | |
| 2549 | # determine the cudagraph_capture_sizes |
| 2550 | if self.compilation_config.cudagraph_capture_sizes is not None: |
| 2551 | assert len(self.compilation_config.cudagraph_capture_sizes) > 0, ( |
| 2552 | "cudagraph_capture_sizes should contain at least one element " |
| 2553 | "when using cuda graph." |
| 2554 | ) |
| 2555 | # de-duplicate the sizes provided by the config |
| 2556 | dedup_sizes = list(set(self.compilation_config.cudagraph_capture_sizes)) |
| 2557 | cudagraph_capture_sizes = [ |
| 2558 | i for i in dedup_sizes if i <= max_num_tokens |
| 2559 | ] |
| 2560 | # sort to make sure the sizes are in ascending order |
| 2561 | cudagraph_capture_sizes.sort() |
| 2562 | else: |
| 2563 | if self.performance_mode == "interactivity": |
| 2564 | # Fine-grained CUDA graphs at small batch sizes |
| 2565 | # for minimal padding overhead |
| 2566 | interactivity_max = min(max_cudagraph_capture_size, 32) |
| 2567 | cudagraph_capture_sizes = list(range(1, interactivity_max + 1)) |
| 2568 | else: |
| 2569 | cudagraph_capture_sizes = [ |
| 2570 | i for i in [1, 2, 4] if i <= max_cudagraph_capture_size |
| 2571 | ] |
| 2572 | if max_cudagraph_capture_size >= 8: |
| 2573 | # Step size 8 for small batch sizes, up to 256(not included) |
| 2574 | cudagraph_capture_sizes += list( |
| 2575 | range(8, min(max_cudagraph_capture_size + 1, 256), 8) |
| 2576 | ) |
| 2577 | if max_cudagraph_capture_size >= 256: |
| 2578 | # Step size 16 for larger batch sizes |
| 2579 | cudagraph_capture_sizes += list( |
| 2580 | range(256, max_cudagraph_capture_size + 1, 16) |
| 2581 | ) |
| 2582 | # ensure max_num_tokens is captured if within max capture size |
| 2583 | if ( |
| 2584 | max_num_tokens <= max_cudagraph_capture_size |
| 2585 | and max_num_tokens not in cudagraph_capture_sizes |
| 2586 | ): |
| 2587 | cudagraph_capture_sizes.append(max_num_tokens) |
| 2588 | # Preserve the platform's default capture ceiling. Larger |
| 2589 | # uniform decode batches fall back to eager execution unless |
| 2590 | # users explicitly configure wider capture sizes. |
| 2591 | cudagraph_capture_sizes += [ |
| 2592 | size for size in uniform_decode_sizes if size <= max_num_tokens |
| 2593 | ] |
| 2594 | # de-duplicate and sort the sizes |
| 2595 | cudagraph_capture_sizes = sorted(set(cudagraph_capture_sizes)) |
| 2596 | |
| 2597 | if ( |
| 2598 | self.parallel_config.tensor_parallel_size > 1 |
| 2599 | and self.compilation_config.pass_config.enable_sp |
| 2600 | ): |
| 2601 | # Sequence parallelism only captures TP-divisible sizes, so a |
| 2602 | # wider non-divisible decode batch cannot be captured under SP. |
| 2603 | cudagraph_capture_sizes = self.update_sizes_for_sequence_parallelism( |
| 2604 | cudagraph_capture_sizes |
| 2605 | ) |
| 2606 | |
| 2607 | # user-specific compilation_config.max_cudagraph_capture_size get |
| 2608 | # truncated to valid_max_size when they are inconsistent. |
| 2609 | valid_max_size = ( |
| 2610 | cudagraph_capture_sizes[-1] if cudagraph_capture_sizes else 0 |
| 2611 | ) |
| 2612 | if ( |
| 2613 | self.compilation_config.max_cudagraph_capture_size is not None |
| 2614 | and self.compilation_config.max_cudagraph_capture_size != valid_max_size |
| 2615 | ): |
| 2616 | # raise error only when both two flags are user-specified |
| 2617 | # and they are inconsistent with each other |
| 2618 | if self.compilation_config.cudagraph_capture_sizes is not None: |
| 2619 | raise ValueError( |
| 2620 | "customized max_cudagraph_capture_size" |
| 2621 | f"(={self.compilation_config.max_cudagraph_capture_size}) " |
| 2622 | "should be consistent with the max value of " |
| 2623 | f"cudagraph_capture_sizes(={valid_max_size})" |
| 2624 | ) |
| 2625 | |
| 2626 | logger.warning( |
| 2627 | "Truncating max_cudagraph_capture_size to %d", |
| 2628 | valid_max_size, |
| 2629 | ) |
| 2630 | # always set the final max_cudagraph_capture_size |
| 2631 | self.compilation_config.max_cudagraph_capture_size = valid_max_size |
| 2632 | |
| 2633 | if self.compilation_config.cudagraph_capture_sizes is not None and len( |
| 2634 | cudagraph_capture_sizes |
| 2635 | ) < len(self.compilation_config.cudagraph_capture_sizes): |
| 2636 | # If users have specified capture sizes, we only need to |
| 2637 | # compare the lens before and after modification since the modified |
| 2638 | # list is only the subset of the original list. |
| 2639 | logger.warning( |
| 2640 | ( |
| 2641 | "cudagraph_capture_sizes specified in compilation_config" |
| 2642 | " %s is overridden by config %s" |
| 2643 | ), |
| 2644 | self.compilation_config.cudagraph_capture_sizes, |
| 2645 | cudagraph_capture_sizes, |
| 2646 | ) |
| 2647 | # always write back the final sizes |
| 2648 | self.compilation_config.cudagraph_capture_sizes = cudagraph_capture_sizes |
| 2649 | |
| 2650 | else: |
| 2651 | # no cudagraph in use |
| 2652 | self.compilation_config.max_cudagraph_capture_size = 0 |
| 2653 | self.compilation_config.cudagraph_capture_sizes = [] |
| 2654 | |
| 2655 | # complete the remaining process. |
| 2656 | self.compilation_config.post_init_cudagraph_sizes() |
| 2657 | |
| 2658 | def _set_compile_ranges(self): |
| 2659 | """Set the compile ranges for the compilation config.""" |
| 2660 | compilation_config = self.compilation_config |
| 2661 | computed_compile_ranges_endpoints = [] |
| 2662 | |
| 2663 | # The upper bound of the compile ranges is the max_num_batched_tokens. |
| 2664 | compile_range_end = self.scheduler_config.max_num_batched_tokens |
| 2665 | if compile_range_end is not None: |
| 2666 | computed_compile_ranges_endpoints.append(compile_range_end) |
| 2667 | |
| 2668 | # Add the compile ranges for flashinfer/aiter. |
| 2669 | if compilation_config.pass_config.fuse_allreduce_rms: |
| 2670 | tp_size = self.parallel_config.tensor_parallel_size |
| 2671 | from vllm._aiter_ops import rocm_aiter_ops |
| 2672 | |
| 2673 | max_size: int | None = None |
| 2674 | if rocm_aiter_ops.is_custom_all_reduce_enabled(): |
| 2675 | from vllm.distributed.device_communicators.aiter_custom_all_reduce import ( # noqa: E501 |
| 2676 | AiterCustomAllreduce, |
| 2677 | ) |
| 2678 | |
| 2679 | max_size = AiterCustomAllreduce.effective_max_size() |
| 2680 | else: |
| 2681 | max_size = compilation_config.pass_config.flashinfer_max_size(tp_size) |
| 2682 | if max_size is not None and self.model_config is not None: |
| 2683 | assert isinstance(self.model_config.dtype, torch.dtype) |
| 2684 | max_token_num = max_size // ( |
| 2685 | self.model_config.get_hidden_size() |
| 2686 | * self.model_config.dtype.itemsize |
| 2687 | ) |
| 2688 | if compile_range_end is not None and max_token_num < compile_range_end: |
| 2689 | computed_compile_ranges_endpoints.append(max_token_num) |
| 2690 | else: |
| 2691 | logger.debug( |
| 2692 | "Max num batched tokens below allreduce-rms fusion threshold, " |
| 2693 | "allreduce-rms fusion will be enabled for all num_tokens." |
| 2694 | ) |
| 2695 | |
| 2696 | # Add the compile ranges for sequence parallelism |
| 2697 | if compilation_config.pass_config.enable_sp: |
| 2698 | pass_config = compilation_config.pass_config |
| 2699 | |
| 2700 | # Calculate min_token_num if not explicitly provided |
| 2701 | # User override works regardless of hidden_size |
| 2702 | if pass_config.sp_min_token_num is None: |
| 2703 | from vllm.compilation.passes.fusion.sequence_parallelism import ( |
| 2704 | get_sequence_parallelism_threshold, |
| 2705 | ) |
| 2706 | |
| 2707 | tp_size = self.parallel_config.tensor_parallel_size |
| 2708 | hidden_size = self.model_config.get_hidden_size() |
| 2709 | assert isinstance(self.model_config.dtype, torch.dtype) |
| 2710 | element_size = self.model_config.dtype.itemsize |
| 2711 | pass_config.sp_min_token_num = get_sequence_parallelism_threshold( |
| 2712 | hidden_size, tp_size, element_size |
| 2713 | ) |
| 2714 | |
| 2715 | min_token_num = pass_config.sp_min_token_num |
| 2716 | max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens |
| 2717 | if min_token_num is not None and ( |
| 2718 | max_num_batched_tokens is not None |
| 2719 | and min_token_num < max_num_batched_tokens |
| 2720 | and min_token_num > 1 |
| 2721 | ): |
| 2722 | # Add endpoint at min_token_num - 1 to ensure SP applies |
| 2723 | # starting from min_token_num |
| 2724 | # This creates ranges: [1, min-1] (no SP), [min, max] (SP applies) |
| 2725 | computed_compile_ranges_endpoints.append(min_token_num - 1) |
| 2726 | |
| 2727 | if compilation_config.pass_config.fuse_rope_kvcache: |
| 2728 | max_token_num = ( |
| 2729 | compilation_config.pass_config.rope_kvcache_fusion_max_token_num |
| 2730 | ) |
| 2731 | if max_token_num is not None: |
| 2732 | if compile_range_end is not None and max_token_num < compile_range_end: |
| 2733 | computed_compile_ranges_endpoints.append(max_token_num) |
| 2734 | else: |
| 2735 | logger.debug( |
| 2736 | "Max num batched tokens below rope+kvcache fusion threshold, " |
| 2737 | "rope+kvcache fusion enabled for num_tokens <= %d.", |
| 2738 | compile_range_end, |
| 2739 | ) |
| 2740 | |
| 2741 | if compilation_config.pass_config.fuse_qk_norm_rope_kvcache: |
| 2742 | max_token_num = ( |
| 2743 | compilation_config.pass_config.rope_kvcache_fusion_max_token_num |
| 2744 | ) |
| 2745 | if max_token_num is not None: |
| 2746 | if compile_range_end is not None and max_token_num < compile_range_end: |
| 2747 | computed_compile_ranges_endpoints.append(max_token_num) |
| 2748 | else: |
| 2749 | logger.debug( |
| 2750 | "Max num batched tokens below qk_norm+rope+kvcache " |
| 2751 | "fusion threshold, fusion enabled for " |
| 2752 | "num_tokens <= %d.", |
| 2753 | compile_range_end, |
| 2754 | ) |
| 2755 | |
| 2756 | if compilation_config.compile_ranges_endpoints is not None: |
| 2757 | for x in compilation_config.compile_ranges_endpoints: |
| 2758 | assert isinstance(x, int) |
| 2759 | assert x > 0, f"Invalid compile range endpoint: {x}" |
| 2760 | if compile_range_end is not None and x < compile_range_end and x > 1: |
| 2761 | computed_compile_ranges_endpoints.append(x) |
| 2762 | compilation_config.compile_ranges_endpoints = sorted( |
| 2763 | computed_compile_ranges_endpoints |
| 2764 | ) |
| 2765 | |
| 2766 | def try_verify_and_update_config(self): |
| 2767 | if self.model_config is None: |
| 2768 | return |
| 2769 | |
| 2770 | # Avoid running try_verify_and_update_config multiple times |
| 2771 | if getattr(self.model_config, "config_updated", False): |
| 2772 | return |
| 2773 | self.model_config.config_updated = True |
| 2774 | |
| 2775 | architecture = self.model_config.architecture |
| 2776 | if architecture is None: |
| 2777 | return |
| 2778 | |
| 2779 | from vllm.model_executor.models import ModelRegistry |
| 2780 | from vllm.model_executor.models.config import ( |
| 2781 | MODELS_CONFIG_MAP, |
| 2782 | HybridAttentionMambaModelConfig, |
| 2783 | ) |
| 2784 | |
| 2785 | cls = MODELS_CONFIG_MAP.get(architecture, None) |
| 2786 | if cls is None: |
| 2787 | # `architecture` may be an HF base-model name (e.g. "Mamba2Model" |
| 2788 | # when `architectures` is omitted); normalize to the resolved arch |
| 2789 | # so per-arch config hooks are not skipped. |
| 2790 | architecture = ModelRegistry._normalize_arch( |
| 2791 | architecture, self.model_config |
| 2792 | ) |
| 2793 | cls = MODELS_CONFIG_MAP.get(architecture, None) |
| 2794 | if cls is not None: |
| 2795 | cls.verify_and_update_config(self) |
| 2796 | |
| 2797 | if self.model_config.is_hybrid: |
| 2798 | HybridAttentionMambaModelConfig.verify_and_update_config(self) |
| 2799 | |
| 2800 | if self.model_config.convert_type == "classify": |
| 2801 | # Maybe convert ForCausalLM into ForSequenceClassification model. |
| 2802 | from vllm.model_executor.models.adapters import SequenceClassificationConfig |
| 2803 | |
| 2804 | SequenceClassificationConfig.verify_and_update_config(self) |
| 2805 | |
| 2806 | if hasattr(self.model_config, "model_weights") and is_runai_obj_uri( |
| 2807 | self.model_config.model_weights |
| 2808 | ): |
| 2809 | if self.load_config.load_format == "auto": |
| 2810 | logger.info( |
| 2811 | "Detected Run:ai model config. " |
| 2812 | "Overriding `load_format` to 'runai_streamer'" |
| 2813 | ) |
| 2814 | self.load_config.load_format = "runai_streamer" |
| 2815 | elif self.load_config.load_format not in ( |
| 2816 | "modelexpress", |
| 2817 | "runai_streamer", |
| 2818 | "runai_streamer_sharded", |
| 2819 | ): |
| 2820 | raise ValueError( |
| 2821 | f"To load a model from object storage (S3/GCS/Azure), " |
| 2822 | f"'load_format' must be 'modelexpress', 'runai_streamer' or " |
| 2823 | f"'runai_streamer_sharded', " |
| 2824 | f"but got '{self.load_config.load_format}'. " |
| 2825 | f"Model: {self.model_config.model}" |
| 2826 | ) |
| 2827 | |
| 2828 | def compile_debug_dump_path(self) -> Path | None: |
| 2829 | """Returns a rank-aware path for dumping |
| 2830 | torch.compile debug information. |
| 2831 | """ |
| 2832 | if self.compilation_config.debug_dump_path is None: |
| 2833 | return None |
| 2834 | tp_rank = self.parallel_config.rank |
| 2835 | dp_rank = self.parallel_config.data_parallel_index |
| 2836 | append_path = f"rank_{tp_rank}_dp_{dp_rank}" |
| 2837 | path = self.compilation_config.debug_dump_path / append_path |
| 2838 | return path |
| 2839 | |
| 2840 | def __str__(self): |
| 2841 | return ( |
| 2842 | f"model={self.model_config.model!r}, " |
| 2843 | f"speculative_config={self.speculative_config!r}, " |
| 2844 | f"tokenizer={self.model_config.tokenizer!r}, " |
| 2845 | f"skip_tokenizer_init={self.model_config.skip_tokenizer_init}, " |
| 2846 | f"tokenizer_mode={self.model_config.tokenizer_mode}, " |
| 2847 | f"revision={self.model_config.revision}, " |
| 2848 | f"tokenizer_revision={self.model_config.tokenizer_revision}, " |
| 2849 | f"trust_remote_code={self.model_config.trust_remote_code}, " |
| 2850 | f"dtype={self.model_config.dtype}, " |
| 2851 | f"max_seq_len={self.model_config.max_model_len}, " |
| 2852 | f"download_dir={self.load_config.download_dir!r}, " |
| 2853 | f"load_format={self.load_config.load_format}, " |
| 2854 | f"tensor_parallel_size={self.parallel_config.tensor_parallel_size}, " # noqa |
| 2855 | f"pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, " # noqa |
| 2856 | f"data_parallel_size={self.parallel_config.data_parallel_size}, " # noqa |
| 2857 | f"decode_context_parallel_size={self.parallel_config.decode_context_parallel_size}, " # noqa |
| 2858 | f"dcp_comm_backend={self.parallel_config.dcp_comm_backend}, " # noqa |
| 2859 | f"disable_custom_all_reduce={self.parallel_config.disable_custom_all_reduce}, " # noqa |
| 2860 | f"quantization={self.model_config.quantization}, " |
| 2861 | f"quantization_config={self.model_config.quantization_config}, " # noqa |
| 2862 | f"enforce_eager={self.model_config.enforce_eager}, " |
| 2863 | f"aux_output_config={self.aux_output_config!r}, " |
| 2864 | f"kv_cache_dtype={self.cache_config.cache_dtype}, " |
| 2865 | f"device_config={self.device_config.device}, " |
| 2866 | f"structured_outputs_config={self.structured_outputs_config!r}, " |
| 2867 | f"observability_config={self.observability_config!r}, " |
| 2868 | f"seed={self.model_config.seed}, " |
| 2869 | f"served_model_name={self.model_config.served_model_name}, " |
| 2870 | f"enable_prefix_caching={self.cache_config.enable_prefix_caching}, " |
| 2871 | f"enable_chunked_prefill={self.scheduler_config.enable_chunked_prefill}, " # noqa |
| 2872 | f"pooler_config={self.model_config.pooler_config!r}, " |
| 2873 | f"compilation_config={self.compilation_config!r}, " |
| 2874 | f"kernel_config={self.kernel_config!r}" |
| 2875 | ) |
| 2876 | |
| 2877 | def _resolve_mm_embedding_inputs(self) -> None: |
| 2878 | """Accept embedding inputs, tensor optional, on disaggregated consumers. |
| 2879 | |
| 2880 | An EC consumer loads embeddings from its connector. A KV consumer |
| 2881 | receives the prompt KV produced from those embeddings, so it does not |
| 2882 | need the tensors either. On every other deployment a missing tensor is |
| 2883 | a client error and must keep failing fast in the frontend. |
| 2884 | """ |
| 2885 | model_config = self.model_config |
| 2886 | if model_config is None: |
| 2887 | return |
| 2888 | mm_config = model_config.multimodal_config |
| 2889 | if mm_config is None: |
| 2890 | return |
| 2891 | |
| 2892 | ec_config = self.ec_transfer_config |
| 2893 | kv_config = self.kv_transfer_config |
| 2894 | # Derived, so overwrite unconditionally rather than honouring a value |
| 2895 | # that was set by hand. |
| 2896 | mm_config.allow_missing_mm_embeddings = ( |
| 2897 | ec_config is not None and ec_config.is_ec_consumer |
| 2898 | ) or (kv_config is not None and kv_config.is_kv_consumer) |
| 2899 | if not mm_config.allow_missing_mm_embeddings: |
| 2900 | return |
| 2901 | |
| 2902 | if not mm_config.enable_mm_embeds: |
| 2903 | # Allowing missing tensors still requires enabling embedding inputs |
| 2904 | # for the frontend to accept metadata-only requests. |
| 2905 | mm_config.enable_mm_embeds = True |
| 2906 | logger.info_once( |
| 2907 | "EC/KV consumer: accepting pre-computed-embedding inputs, " |
| 2908 | "which this role is sent by definition." |
| 2909 | ) |
| 2910 | logger.info_once( |
| 2911 | "EC/KV consumer: pre-computed-embedding inputs may " |
| 2912 | "omit the embedding tensor." |
| 2913 | ) |
| 2914 | |
| 2915 | def _resolve_mm_encoder_only(self) -> None: |
| 2916 | """Enable encoder-only mode for a dedicated EC producer.""" |
| 2917 | ec_config = self.ec_transfer_config |
| 2918 | if ec_config is None or not ec_config.is_encode_only: |
| 2919 | return |
| 2920 | |
| 2921 | model_config = self.model_config |
| 2922 | mm_config = model_config.multimodal_config if model_config is not None else None |
| 2923 | if mm_config is None: |
| 2924 | raise ValueError( |
| 2925 | "An EC producer-only instance requires a multimodal model." |
| 2926 | ) |
| 2927 | mm_config.mm_encoder_only = True |
| 2928 | |
| 2929 | def _resolve_mm_processor_device(self) -> None: |
| 2930 | """Settle `--mm-processor-device=auto` now that the EC role is known. |
| 2931 | |
| 2932 | "auto" means "the accelerator, but only where the processor has it to |
| 2933 | itself and its output can be handed over without a copy back to host": |
| 2934 | an encode-only instance whose tensor transport carries device tensors. |
| 2935 | Every other deployment keeps the processor on CPU. |
| 2936 | |
| 2937 | An explicit device -- from `--mm-processor-device` or straight from |
| 2938 | `mm_processor_kwargs` -- is already folded in by `MultiModalConfig`, so |
| 2939 | it is left alone here and validated by `_validate_mm_processor_device`. |
| 2940 | """ |
| 2941 | model_config = self.model_config |
| 2942 | if model_config is None: |
| 2943 | return |
| 2944 | mm_config = model_config.multimodal_config |
| 2945 | if mm_config is None: |
| 2946 | return |
| 2947 | if mm_config.get_mm_processor_device_type() is not None: |
| 2948 | return |
| 2949 | |
| 2950 | from vllm.platforms import current_platform |
| 2951 | |
| 2952 | device_type = current_platform.device_type |
| 2953 | if device_type in ("", "cpu"): |
| 2954 | return |
| 2955 | |
| 2956 | ec_config = self.ec_transfer_config |
| 2957 | # An EC producer that is not also a consumer runs no forward pass and |
| 2958 | # allocates no KV cache, so frontend accelerator work has the device to |
| 2959 | # itself. |
| 2960 | if ec_config is None or not ec_config.is_encode_only: |
| 2961 | return |
| 2962 | |
| 2963 | if mm_config.mm_tensor_ipc != "torch_shm": |
| 2964 | # Any other transport serializes host bytes, so the output would be |
| 2965 | # copied back, and that copy costs more than running the transform |
| 2966 | # on device saves. |
| 2967 | logger.info_once( |
| 2968 | "EPD encoder instance: keeping the multi-modal processor on CPU " |
| 2969 | "because mm_tensor_ipc=%s cannot carry device tensors. Add " |
| 2970 | "--mm-tensor-ipc=torch_shm to run it on the accelerator.", |
| 2971 | mm_config.mm_tensor_ipc, |
| 2972 | ) |
| 2973 | return |
| 2974 | |
| 2975 | mm_config.mm_processor_kwargs = { |
| 2976 | **(mm_config.mm_processor_kwargs or {}), |
| 2977 | "device": device_type, |
| 2978 | } |
| 2979 | logger.info_once( |
| 2980 | "EPD encoder instance: running the multi-modal processor on %s. " |
| 2981 | "Override with --mm-processor-device=cpu.", |
| 2982 | device_type, |
| 2983 | ) |
| 2984 | |
| 2985 | def _resolve_mm_video_decode_device(self) -> None: |
| 2986 | """Default video decoding to torchcodec GPU backend for EPD encoder-only |
| 2987 | instance if the mm processor runs on CUDA. |
| 2988 | |
| 2989 | The processor consumes the decoded frames on-device in that case, so |
| 2990 | keeping the frames on the GPU skips the host round-trip through the |
| 2991 | CPU media path. An explicit codec/backend choice in |
| 2992 | `--media-io-kwargs` is left alone, and the default is skipped where |
| 2993 | torchcodec (or its FFmpeg runtime) is unavailable. |
| 2994 | """ |
| 2995 | if self.model_config is None or self.model_config.multimodal_config is None: |
| 2996 | return |
| 2997 | mm_config = self.model_config.multimodal_config |
| 2998 | |
| 2999 | ec_config = self.ec_transfer_config |
| 3000 | # An EC producer that is not also a consumer runs no forward pass and |
| 3001 | # allocates no KV cache, so frontend accelerator work has the device to |
| 3002 | # itself. |
| 3003 | if ec_config is None or not ec_config.is_encode_only: |
| 3004 | return |
| 3005 | |
| 3006 | from vllm.platforms import current_platform |
| 3007 | |
| 3008 | device_type = current_platform.device_type |
| 3009 | if ( |
| 3010 | device_type != "cuda" |
| 3011 | or mm_config.get_mm_processor_device_type() != device_type |
| 3012 | ): |
| 3013 | return |
| 3014 | |
| 3015 | # User set video backend or device explicitly |
| 3016 | video_kwargs = mm_config.media_io_kwargs.setdefault("video", {}) |
| 3017 | if "backend" in video_kwargs or "device" in video_kwargs: |
| 3018 | return |
| 3019 | |
| 3020 | from vllm.utils.import_utils import check_torchcodec_available |
| 3021 | |
| 3022 | try: |
| 3023 | check_torchcodec_available() |
| 3024 | except (ImportError, RuntimeError): |
| 3025 | # torchcodec is not installed, or is installed without a usable |
| 3026 | # FFmpeg runtime (it raises rather than returning False). |
| 3027 | logger.info_once( |
| 3028 | "EPD encoder instance: keeping CPU video decoding because " |
| 3029 | "torchcodec is not available (needs a CUDA build with FFmpeg)." |
| 3030 | ) |
| 3031 | return |
| 3032 | |
| 3033 | video_kwargs["backend"] = "torchcodec" |
| 3034 | video_kwargs["device"] = device_type |
| 3035 | logger.info_once( |
| 3036 | "EPD encoder instance: decoding video with NVDEC (torchcodec device=%s).", |
| 3037 | device_type, |
| 3038 | ) |
| 3039 | |
| 3040 | def _validate_mm_processor_device(self) -> None: |
| 3041 | """Hand the EC config to `MultiModalConfig`, which owns the rule.""" |
| 3042 | model_config = self.model_config |
| 3043 | if model_config is None: |
| 3044 | return |
| 3045 | mm_config = model_config.multimodal_config |
| 3046 | if mm_config is None: |
| 3047 | return |
| 3048 | |
| 3049 | mm_config.validate_mm_processor_device(self.ec_transfer_config) |
| 3050 | |
| 3051 | def _get_v2_model_runner_unsupported_features(self) -> list[str]: |
| 3052 | """Collect features not yet supported by the V2 model runner.""" |
| 3053 | unsupported: list[str] = [] |
| 3054 | speculative_config = self.speculative_config |
| 3055 | |
| 3056 | if ( |
| 3057 | self.compilation_config.pass_config.enable_sp |
| 3058 | and self.parallel_config.tensor_parallel_size > 1 |
| 3059 | ): |
| 3060 | unsupported.append("sequence parallelism") |
| 3061 | |
| 3062 | # V2 does not implement the external_launcher (torchrun) PP-output |
| 3063 | # broadcast that V1 uses to keep all ranks in sync (broadcast_pp_output). |
| 3064 | if ( |
| 3065 | self.parallel_config.distributed_executor_backend == "external_launcher" |
| 3066 | and self.parallel_config.pipeline_parallel_size > 1 |
| 3067 | ): |
| 3068 | unsupported.append("pipeline parallelism with external_launcher") |
| 3069 | |
| 3070 | if speculative_config is not None: |
| 3071 | if speculative_config.method in ( |
| 3072 | # https://github.com/vllm-project/vllm/pull/40704 |
| 3073 | "ngram", |
| 3074 | "ngram_gpu", |
| 3075 | "suffix", |
| 3076 | "medusa", |
| 3077 | "mlp_speculator", |
| 3078 | "custom_class", |
| 3079 | ): |
| 3080 | unsupported.append(f"speculative method '{speculative_config.method}'") |
| 3081 | |
| 3082 | # V2 EagleSpeculator does not support parallel_drafting (for P-Eagle). |
| 3083 | # DFlash and DSpark use parallel drafting natively in V2 via their |
| 3084 | # own speculators. |
| 3085 | if ( |
| 3086 | speculative_config.parallel_drafting |
| 3087 | and speculative_config.method not in ("dflash", "dspark") |
| 3088 | ): |
| 3089 | unsupported.append("parallel drafting for EAGLE speculative decoding") |
| 3090 | |
| 3091 | if self.parallel_config.use_ubatching: |
| 3092 | unsupported.extend(self._get_dbo_unsupported_features()) |
| 3093 | |
| 3094 | return unsupported |
| 3095 | |
| 3096 | def _get_v1_model_runner_unsupported_features(self) -> list[str]: |
| 3097 | unsupported: list[str] = [] |
| 3098 | |
| 3099 | # PCP runtime support is implemented only by the V2 model runner. |
| 3100 | if self.parallel_config.prefill_context_parallel_size > 1: |
| 3101 | unsupported.append("prefill context parallel") |
| 3102 | |
| 3103 | # Note(arpera): |
| 3104 | # MRV1 + PP>1 + async sched + structured output |
| 3105 | # does not work in vLLM. For more info see: |
| 3106 | # https://github.com/vllm-project/vllm/issues/45014 |
| 3107 | # Since recently MRV1 has been deprecated then |
| 3108 | # there was decided not to fix the issue but instead |
| 3109 | # to disallow such configuration. |
| 3110 | # At the same time MRV2 works fine in this case. |
| 3111 | if ( |
| 3112 | self.parallel_config.pipeline_parallel_size > 1 |
| 3113 | and self.scheduler_config.async_scheduling |
| 3114 | ): |
| 3115 | unsupported.append("pipeline parallelism with async scheduling") |
| 3116 | |
| 3117 | # DSpark is implemented only by the V2 GPU model runner. |
| 3118 | if self.speculative_config: |
| 3119 | if self.speculative_config.method == "dspark": |
| 3120 | unsupported.append("dspark speculative decoding") |
| 3121 | if self.speculative_config.enable_adaptive_verification: |
| 3122 | unsupported.append("adaptive draft verification") |
| 3123 | |
| 3124 | # Mixed sliding/full DFlash drafts need multiple KV groups (V2 only). |
| 3125 | if self._dflash_needs_multi_kv_group(): |
| 3126 | unsupported.append("mixed sliding/full dflash drafts") |
| 3127 | |
| 3128 | # DFlash candidate heads exist only in the V2 speculator. On |
| 3129 | # V1 the same checkpoint drafts through DFlashProposer, which never |
| 3130 | # calls it, so the draft would degrade to DFlash1 silently. |
| 3131 | if self._is_dflash_candidate_draft(): |
| 3132 | unsupported.append("DFlash candidate-head drafts") |
| 3133 | |
| 3134 | if self.model_config is not None and self.model_config.is_diffusion: |
| 3135 | unsupported.append("diffusion models") |
| 3136 | |
| 3137 | if self.parallel_config.enable_batch_sharded_sampling: |
| 3138 | unsupported.append("batch-sharded sampling") |
| 3139 | |
| 3140 | return unsupported |
| 3141 | |
| 3142 | def _validate_adaptive_verification(self) -> None: |
| 3143 | spec_config = self.speculative_config |
| 3144 | if not spec_config or not spec_config.enable_adaptive_verification: |
| 3145 | return |
| 3146 | |
| 3147 | if self.lora_config is not None: |
| 3148 | # The per-token LoRA mapping is built from CPU placeholder boundaries, |
| 3149 | # while the trimmed batch's true boundaries are decided on the GPU. |
| 3150 | raise ValueError( |
| 3151 | "Adaptive verification is not currently compatible with LoRA" |
| 3152 | ) |
| 3153 | |
| 3154 | if not self.compilation_config.cudagraph_mode.has_full_cudagraphs(): |
| 3155 | raise ValueError( |
| 3156 | "Adaptive verification requires full CUDA graphs. Use cudagraph " |
| 3157 | "mode FULL, FULL_DECODE_ONLY, or FULL_AND_PIECEWISE, or disable " |
| 3158 | "adaptive verification." |
| 3159 | ) |
| 3160 | |
| 3161 | if self.parallel_config.pipeline_parallel_size > 1: |
| 3162 | # Cost curves and confidences currently only exist on the last PP rank; |
| 3163 | # earlier ranks would diverge on the trimmed batch shape. |
| 3164 | # TODO: we should be able to support adaptive verification with PP by |
| 3165 | # broadcasting the cost curves and confidences to all ranks. |
| 3166 | raise ValueError( |
| 3167 | "Adaptive verification is not currently compatible " |
| 3168 | "with pipeline parallelism" |
| 3169 | ) |
| 3170 | |
| 3171 | def _validate_batch_sharded_sampling(self) -> None: |
| 3172 | """Validate `enable_batch_sharded_sampling` against the rest of the config.""" |
| 3173 | if not self.parallel_config.enable_batch_sharded_sampling: |
| 3174 | # Default to False if not set. |
| 3175 | self.parallel_config.enable_batch_sharded_sampling = False |
| 3176 | return |
| 3177 | |
| 3178 | blockers: list[str] = [] |
| 3179 | tp_size = self.parallel_config.tensor_parallel_size |
| 3180 | |
| 3181 | if tp_size <= 1: |
| 3182 | blockers.append("tensor_parallel_size is 1, so there is nothing to shard") |
| 3183 | elif self.scheduler_config.max_num_seqs < tp_size: |
| 3184 | # Requests are assigned to ranks whole, so fewer slots than ranks |
| 3185 | # leaves some ranks without work in every step. |
| 3186 | blockers.append( |
| 3187 | f"max_num_seqs ({self.scheduler_config.max_num_seqs}) is below " |
| 3188 | f"tensor_parallel_size ({tp_size})" |
| 3189 | ) |
| 3190 | |
| 3191 | if self.model_config is not None and self.model_config.max_logprobs < 0: |
| 3192 | # max_logprobs == -1 allows vocab-size logprob requests, which the |
| 3193 | # fixed-width logprobs gather cannot reasonably size for. |
| 3194 | blockers.append("max_logprobs is -1, allowing vocab-size logprob requests") |
| 3195 | |
| 3196 | if self.model_config is not None and self.model_config.return_sampling_mask: |
| 3197 | # gather_sampler_output() drops SamplingMaskTensors: masks come back None. |
| 3198 | blockers.append( |
| 3199 | "return_sampling_mask is set and the batch-sharded gather does " |
| 3200 | "not forward sampling masks" |
| 3201 | ) |
| 3202 | |
| 3203 | if ( |
| 3204 | self.speculative_config is not None |
| 3205 | and self.speculative_config.enable_adaptive_verification |
| 3206 | ): |
| 3207 | # Adaptive verification picks the per-request draft split on the GPU, |
| 3208 | # so cu_num_logits_np is only an upper bound, while the shard plan is |
| 3209 | # built from that CPU array. The two disagree once the budget binds. |
| 3210 | # TODO(TheEpicDolphin): Support adaptive verification with batch-sharded |
| 3211 | # sampling. |
| 3212 | blockers.append( |
| 3213 | "it does not yet work with adaptive verification, which decides " |
| 3214 | "the per-request logits counts on the GPU, where the CPU-side " |
| 3215 | "shard plan cannot see them" |
| 3216 | ) |
| 3217 | |
| 3218 | if blockers: |
| 3219 | raise ValueError( |
| 3220 | "Batch-sharded sampling was explicitly enabled via " |
| 3221 | "the --enable-batch-sharded-sampling flag, but is not supported " |
| 3222 | "in this configuration for the following reason(s): " |
| 3223 | f"{'; '.join(blockers)}." |
| 3224 | ) |
| 3225 | |
| 3226 | def _get_dbo_unsupported_features(self) -> list[str]: |
| 3227 | """Collect what the V2 model runner cannot combine with DBO. |
| 3228 | |
| 3229 | The V2 runner microbatches a plain decoder forward pass. Anything that |
| 3230 | slices or replays the batch differently (drafting, adapters, pipeline |
| 3231 | stages, context parallelism, encoders) is not handled yet. |
| 3232 | """ |
| 3233 | # TODO: DBO with model runner V2 is under development. |
| 3234 | # It should be enabled with explicit VLLM_USE_V2_MODEL_RUNNER environ. |
| 3235 | # Remove it when stable. |
| 3236 | if envs.VLLM_USE_V2_MODEL_RUNNER is None: |
| 3237 | return ["dual batch overlap"] |
| 3238 | |
| 3239 | unsupported: list[str] = [] |
| 3240 | model_config = self.model_config |
| 3241 | parallel_config = self.parallel_config |
| 3242 | |
| 3243 | if self.lora_config is not None: |
| 3244 | unsupported.append("dual batch overlap with LoRA") |
| 3245 | if self.speculative_config is not None: |
| 3246 | unsupported.append("dual batch overlap with speculative decoding") |
| 3247 | if parallel_config.pipeline_parallel_size > 1: |
| 3248 | unsupported.append("dual batch overlap with pipeline parallelism") |
| 3249 | if ( |
| 3250 | parallel_config.decode_context_parallel_size > 1 |
| 3251 | or parallel_config.prefill_context_parallel_size > 1 |
| 3252 | ): |
| 3253 | unsupported.append("dual batch overlap with context parallelism") |
| 3254 | if model_config is not None and ( |
| 3255 | model_config.is_multimodal_model or model_config.is_encoder_decoder |
| 3256 | ): |
| 3257 | unsupported.append("dual batch overlap with multimodal models") |
| 3258 | if model_config is not None and model_config.is_hybrid: |
| 3259 | unsupported.append("dual batch overlap with hybrid models") |
| 3260 | if self.compilation_config.cudagraph_mode == CUDAGraphMode.PIECEWISE: |
| 3261 | # DBO captures FULL graphs only. |
| 3262 | unsupported.append("dual batch overlap with PIECEWISE CUDA graphs") |
| 3263 | if self.is_mm_encoder_only: |
| 3264 | unsupported.append("dual batch overlap with encoder only models") |
| 3265 | |
| 3266 | return unsupported |
| 3267 | |
| 3268 | def _validate_profiler_config(self) -> None: |
| 3269 | if self.profiler_config.profiler != "proton": |
| 3270 | return |
| 3271 | |
| 3272 | from vllm.platforms import current_platform |
| 3273 | |
| 3274 | if not current_platform.is_cuda(): |
| 3275 | raise ValueError("The Proton profiler currently supports NVIDIA CUDA only") |
| 3276 | if ( |
| 3277 | self.profiler_config.proton_graph_attribution |
| 3278 | and not self.use_v2_model_runner |
| 3279 | ): |
| 3280 | raise ValueError( |
| 3281 | "Proton CUDA graph attribution requires the V2 model runner." |
| 3282 | ) |
| 3283 | |
| 3284 | has_cuda_graphs = ( |
| 3285 | self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE |
| 3286 | or ( |
| 3287 | self.compilation_config.cudagraph_mm_encoder |
| 3288 | and not (self.model_config and self.model_config.enforce_eager) |
| 3289 | ) |
| 3290 | ) |
| 3291 | if not has_cuda_graphs: |
| 3292 | return |
| 3293 | mode = self.profiler_config.proton_mode |
| 3294 | if mode and mode.split(":", 1)[0].lower() == "pcsampling": |
| 3295 | raise ValueError( |
| 3296 | "Proton PC sampling requires CUDA graphs to be disabled. " |
| 3297 | "Use --enforce-eager." |
| 3298 | ) |
| 3299 | if not self.profiler_config.proton_graph_attribution: |
| 3300 | raise ValueError( |
| 3301 | "Proton profiling with CUDA graphs requires " |
| 3302 | "proton_graph_attribution=True to capture replayed kernels. " |
| 3303 | "Enable attribution or use --enforce-eager." |
| 3304 | ) |
| 3305 | |
| 3306 | def _disable_cudagraphs_for_v2_stock_torch_compile(self) -> None: |
| 3307 | """Run stock torch.compile without CUDA graphs in Model Runner V2. |
| 3308 | |
| 3309 | V1 never wraps a stock-compiled model in CUDAGraphWrapper, so it runs |
| 3310 | without CUDA graphs. V2's CUDA graph manager would otherwise capture |
| 3311 | FULL graphs around the stock-compiled model, so disable them to match. |
| 3312 | """ |
| 3313 | compilation_config = self.compilation_config |
| 3314 | if ( |
| 3315 | compilation_config.mode != CompilationMode.STOCK_TORCH_COMPILE |
| 3316 | or compilation_config.cudagraph_mode == CUDAGraphMode.NONE |
| 3317 | ): |
| 3318 | return |
| 3319 | logger.info_once( |
| 3320 | "CUDA graphs are not supported with stock torch.compile in Model " |
| 3321 | "Runner V2. Overriding cudagraph_mode %s to NONE.", |
| 3322 | compilation_config.cudagraph_mode.name, |
| 3323 | ) |
| 3324 | compilation_config.cudagraph_mode = CUDAGraphMode.NONE |
| 3325 | compilation_config.max_cudagraph_capture_size = 0 |
| 3326 | compilation_config.cudagraph_capture_sizes = [] |
| 3327 | |
| 3328 | def _validate_v2_model_runner(self) -> None: |
| 3329 | """Check for features not yet supported by the V2 model runner.""" |
| 3330 | if not HAS_TRITON: |
| 3331 | raise ValueError("Model Runner V2 requires Triton.") |
| 3332 | |
| 3333 | unsupported = self._get_v2_model_runner_unsupported_features() |
| 3334 | if unsupported: |
| 3335 | raise ValueError( |
| 3336 | f"Model Runner V2 does not yet support: {', '.join(unsupported)}" |
| 3337 | ) |
| 3338 | |
| 3339 | def _validate_v1_model_runner(self) -> None: |
| 3340 | unsupported = self._get_v1_model_runner_unsupported_features() |
| 3341 | if unsupported: |
| 3342 | raise ValueError( |
| 3343 | f"Model Runner V1 does not support: {', '.join(unsupported)}" |
| 3344 | ) |
| 3345 | |
| 3346 | def adjust_dcp_kv_cache_interleave_size( |
| 3347 | self, kv_cache_config: "KVCacheConfig" |
| 3348 | ) -> None: |
| 3349 | """Normalize DCP interleave size against block_size for NIXL P/D. |
| 3350 | |
| 3351 | Called by each worker (via ensure_kv_transfer_initialized), once it knows its |
| 3352 | own final block_size via kv_cache_config. |
| 3353 | """ |
| 3354 | dcp_size = self.parallel_config.decode_context_parallel_size |
| 3355 | if dcp_size <= 1: |
| 3356 | return |
| 3357 | if self.parallel_config.dcp_kv_cache_interleave_size > 1 and ( |
| 3358 | self.parallel_config.cp_kv_cache_interleave_size |
| 3359 | != self.parallel_config.dcp_kv_cache_interleave_size |
| 3360 | ): |
| 3361 | self.parallel_config.cp_kv_cache_interleave_size = ( |
| 3362 | self.parallel_config.dcp_kv_cache_interleave_size |
| 3363 | ) |
| 3364 | logger.warning_once( |
| 3365 | "cp_kv_cache_interleave_size is overridden by dcp_kv_cache" |
| 3366 | "_interleave_size. And dcp-kv-cache-interleave-size will be " |
| 3367 | "deprecated when PCP is fully supported." |
| 3368 | ) |
| 3369 | |
| 3370 | if self.kv_transfer_config is None or not self.kv_transfer_config.has_connector( |
| 3371 | "NixlConnector" |
| 3372 | ): |
| 3373 | return |
| 3374 | if not self.parallel_config._allow_auto_resolve_cp_interleave_size: |
| 3375 | return |
| 3376 | |
| 3377 | # Get the kernel block_size, but don't use resolve_kv_cache_block_size to avoid |
| 3378 | # scaling by dcp_size (we need the local block_size here). |
| 3379 | local_block_size = min( |
| 3380 | g.kv_cache_spec.block_size for g in kv_cache_config.kv_cache_groups |
| 3381 | ) |
| 3382 | if self.parallel_config.cp_kv_cache_interleave_size != local_block_size: |
| 3383 | interleave = self.parallel_config.cp_kv_cache_interleave_size |
| 3384 | self.parallel_config.cp_kv_cache_interleave_size = local_block_size |
| 3385 | logger.info_once( |
| 3386 | "When using PD disaggregation with DCP " |
| 3387 | "(decode_context_parallel_size=%d), " |
| 3388 | "cp_kv_cache_interleave_size is automatically adjusted " |
| 3389 | "from %d to block_size %d for block-level alignment.", |
| 3390 | dcp_size, |
| 3391 | interleave, |
| 3392 | local_block_size, |
| 3393 | ) |
| 3394 | |
| 3395 | def validate_block_size(self) -> None: |
| 3396 | """Validate block_size against DCP and mamba constraints. |
| 3397 | |
| 3398 | Called after Platform.update_block_size_for_backend() has |
| 3399 | finalised block_size. |
| 3400 | """ |
| 3401 | block_size = self.cache_config.block_size |
| 3402 | |
| 3403 | # Skip DCP interleave-size compatibility for NIXL P/D: the interleave |
| 3404 | # size is pinned to block_size by each worker. |
| 3405 | nixl_pd_active = ( |
| 3406 | self.kv_transfer_config is not None |
| 3407 | and self.kv_transfer_config.has_connector("NixlConnector") |
| 3408 | ) |
| 3409 | if self.parallel_config.decode_context_parallel_size > 1 and not nixl_pd_active: |
| 3410 | assert ( |
| 3411 | self.parallel_config.cp_kv_cache_interleave_size <= block_size |
| 3412 | and block_size % self.parallel_config.cp_kv_cache_interleave_size == 0 |
| 3413 | ), ( |
| 3414 | f"Block_size({block_size}) should be greater " |
| 3415 | "than or equal to and divisible by cp_kv_cache_interleave_size " |
| 3416 | f"({self.parallel_config.cp_kv_cache_interleave_size})." |
| 3417 | ) |
| 3418 | # Mamba cache align-mode constraints |
| 3419 | if self.cache_config.mamba_cache_mode == "align": |
| 3420 | assert not self.scheduler_config.disable_chunked_mm_input, ( |
| 3421 | "Chunked MM input is required because we need the flexibility " |
| 3422 | "to schedule a multiple of block_size tokens even if they are " |
| 3423 | "in the middle of a mm input" |
| 3424 | ) |
| 3425 | |
| 3426 | @model_validator(mode="after") |
| 3427 | def validate_nvfp4_kv_cache_with_mla(self) -> "VllmConfig": |
| 3428 | if self.model_config is None: |
| 3429 | return self |
| 3430 | # The ds_mla layouts are MLA-only by construction; the plain nvfp4 |
| 3431 | # layout (head_size//2 + head_size//16) does not apply to MLA. |
| 3432 | if ( |
| 3433 | self.cache_config.cache_dtype.startswith("nvfp4") |
| 3434 | and not self.cache_config.cache_dtype.endswith("_ds_mla") |
| 3435 | and self.model_config.use_mla |
| 3436 | ): |
| 3437 | raise ValueError( |
| 3438 | "nvfp4 KV cache is not supported with MLA (Multi-head Latent " |
| 3439 | "Attention) backends. Please use a different --kv-cache-dtype " |
| 3440 | "(e.g., 'fp8', 'auto', or 'nvfp4_ds_mla' with a sparse MLA " |
| 3441 | "backend) for MLA models such as DeepSeek." |
| 3442 | ) |
| 3443 | return self |
| 3444 | |
| 3445 | @model_validator(mode="after") |
| 3446 | def validate_mamba_block_size(self) -> "VllmConfig": |
| 3447 | if self.model_config is None: |
| 3448 | return self |
| 3449 | mamba_block_size_is_set = ( |
| 3450 | self.cache_config.mamba_block_size is not None |
| 3451 | and self.cache_config.mamba_block_size != self.model_config.max_model_len |
| 3452 | ) |
| 3453 | if mamba_block_size_is_set and not self.cache_config.enable_prefix_caching: |
| 3454 | raise ValueError( |
| 3455 | "--mamba-block-size can only be set with --enable-prefix-caching" |
| 3456 | ) |
| 3457 | return self |
| 3458 | |
| 3459 | @model_validator(mode="after") |
| 3460 | def validate_mamba_cached_kernel(self) -> "VllmConfig": |
| 3461 | if not self.cache_config.use_replayssm: |
| 3462 | self.cache_config.use_kda_recoverssm = False |
| 3463 | return self |
| 3464 | |
| 3465 | kda_architectures = ( |
| 3466 | "KimiLinearForCausalLM", |
| 3467 | "KimiK3ForConditionalGeneration", |
| 3468 | ) |
| 3469 | is_kda_model = ( |
| 3470 | self.model_config is not None |
| 3471 | and self.model_config.architecture in kda_architectures |
| 3472 | ) |
| 3473 | self.cache_config.use_kda_recoverssm = ( |
| 3474 | self.num_speculative_tokens > 0 and is_kda_model |
| 3475 | ) |
| 3476 | use_mamba_replayssm_spec = ( |
| 3477 | self.num_speculative_tokens > 0 and not self.cache_config.use_kda_recoverssm |
| 3478 | ) |
| 3479 | |
| 3480 | if self.model_config is not None and not self.model_config.supports_replayssm: |
| 3481 | raise ValueError( |
| 3482 | "--use-replayssm is not supported for architecture " |
| 3483 | f"{self.model_config.architecture!r}" |
| 3484 | ) |
| 3485 | if ( |
| 3486 | self.mamba_config.backend == MambaBackendEnum.FLASHINFER |
| 3487 | and self.cache_config.replayssm_buffer_len > 16 |
| 3488 | ): |
| 3489 | raise ValueError( |
| 3490 | "FlashInfer ReplaySSM requires --replayssm-buffer-len <= 16" |
| 3491 | ) |
| 3492 | if self.cache_config.use_kda_recoverssm: |
| 3493 | if self.mamba_config.enable_stochastic_rounding: |
| 3494 | raise ValueError( |
| 3495 | "RecoverSSM supports bfloat16/float32 " |
| 3496 | "SSM state caches, not --enable-mamba-cache-stochastic-" |
| 3497 | "rounding, which requires an explicit float16 cache" |
| 3498 | ) |
| 3499 | if ( |
| 3500 | self.cache_config.mamba_cache_mode == "align" |
| 3501 | and not self.use_v2_model_runner |
| 3502 | ): |
| 3503 | raise ValueError( |
| 3504 | "RecoverSSM with align mode requires VLLM_USE_V2_MODEL_RUNNER=1" |
| 3505 | ) |
| 3506 | if self.parallel_config.pipeline_parallel_size > 1: |
| 3507 | raise ValueError( |
| 3508 | "RecoverSSM currently requires pipeline_parallel_size=1" |
| 3509 | ) |
| 3510 | if self.mamba_config.backend != MambaBackendEnum.TRITON: |
| 3511 | raise ValueError("RecoverSSM requires --mamba-backend triton") |
| 3512 | elif use_mamba_replayssm_spec: |
| 3513 | if self.cache_config.mamba_cache_mode != "none": |
| 3514 | raise ValueError( |
| 3515 | "FlashInfer ReplaySSM speculative decoding requires " |
| 3516 | "--mamba-cache-mode none" |
| 3517 | ) |
| 3518 | query_len = 1 + self.num_speculative_tokens |
| 3519 | if self.cache_config.replayssm_buffer_len < query_len: |
| 3520 | raise ValueError( |
| 3521 | "FlashInfer ReplaySSM speculative decoding requires " |
| 3522 | "--replayssm-buffer-len >= 1 + num_speculative_tokens " |
| 3523 | f"({query_len}); got " |
| 3524 | f"{self.cache_config.replayssm_buffer_len}" |
| 3525 | ) |
| 3526 | if self.mamba_config.backend != MambaBackendEnum.FLASHINFER: |
| 3527 | raise ValueError( |
| 3528 | "Mamba2 ReplaySSM speculative decoding requires " |
| 3529 | "--mamba-backend flashinfer" |
| 3530 | ) |
| 3531 | elif self.mamba_config.backend == MambaBackendEnum.FLASHINFER: |
| 3532 | if self.cache_config.mamba_cache_mode == "align": |
| 3533 | raise ValueError( |
| 3534 | "FlashInfer ReplaySSM does not support " |
| 3535 | "--mamba-cache-mode align yet; use none" |
| 3536 | ) |
| 3537 | elif self.mamba_config.backend != MambaBackendEnum.TRITON: |
| 3538 | raise ValueError( |
| 3539 | "--use-replayssm requires --mamba-backend triton or flashinfer" |
| 3540 | ) |
| 3541 | elif self.use_v2_model_runner: |
| 3542 | raise ValueError( |
| 3543 | "Triton ReplaySSM requires Model Runner V1; use " |
| 3544 | "--mamba-backend flashinfer or Model Runner V1" |
| 3545 | ) |
| 3546 | if ( |
| 3547 | self.kv_transfer_config is not None |
| 3548 | and self.kv_transfer_config.is_kv_transfer_instance |
| 3549 | ): |
| 3550 | raise ValueError( |
| 3551 | "--use-replayssm is incompatible with KV connectors " |
| 3552 | "(P/D disaggregation, KV cache offload)" |
| 3553 | ) |
| 3554 | return self |
| 3555 | |
| 3556 | |
| 3557 | _current_vllm_config: VllmConfig | None = None |
| 3558 | _current_prefix: str | None = None |
| 3559 | |
| 3560 | |
| 3561 | @contextmanager |
| 3562 | def set_current_vllm_config( |
| 3563 | vllm_config: VllmConfig, check_compile=False, prefix: str | None = None |
| 3564 | ): |
| 3565 | """Temporarily set the current vLLM config. |
| 3566 | Used during model initialization. |
| 3567 | We save the current vLLM config in a global variable, |
| 3568 | so that all modules can access it, e.g. custom ops |
| 3569 | can access the vLLM config to determine how to dispatch. |
| 3570 | """ |
| 3571 | global _current_vllm_config, _current_prefix |
| 3572 | old_vllm_config = _current_vllm_config |
| 3573 | old_prefix = _current_prefix |
| 3574 | from vllm.compilation.counter import compilation_counter |
| 3575 | |
| 3576 | num_models_seen = compilation_counter.num_models_seen |
| 3577 | try: |
| 3578 | # Clear the compilation config cache when context changes. |
| 3579 | # This is needed since the old config may have been accessed |
| 3580 | # and cached before the new config is set. |
| 3581 | get_cached_compilation_config.cache_clear() |
| 3582 | |
| 3583 | _current_vllm_config = vllm_config |
| 3584 | _current_prefix = prefix |
| 3585 | yield |
| 3586 | except Exception: |
| 3587 | raise |
| 3588 | else: |
| 3589 | if check_compile: |
| 3590 | vllm_config.compilation_config.custom_op_log_check() |
| 3591 | |
| 3592 | if ( |
| 3593 | check_compile |
| 3594 | and vllm_config.compilation_config.mode == CompilationMode.VLLM_COMPILE |
| 3595 | and compilation_counter.num_models_seen == num_models_seen |
| 3596 | ): |
| 3597 | # If the model supports compilation, |
| 3598 | # compilation_counter.num_models_seen should be increased |
| 3599 | # by at least 1. |
| 3600 | # If it is not increased, it means the model does not support |
| 3601 | # compilation (does not have @support_torch_compile decorator). |
| 3602 | logger.warning( |
| 3603 | "`torch.compile` is turned on, but the model %s" |
| 3604 | " does not support it. Please open an issue on GitHub" |
| 3605 | " if you want it to be supported.", |
| 3606 | vllm_config.model_config.model, |
| 3607 | ) |
| 3608 | finally: |
| 3609 | _current_vllm_config = old_vllm_config |
| 3610 | _current_prefix = old_prefix |
| 3611 | # Clear the compilation config cache when context changes |
| 3612 | get_cached_compilation_config.cache_clear() |
| 3613 | |
| 3614 | |
| 3615 | @lru_cache(maxsize=1) |
| 3616 | def get_cached_compilation_config(): |
| 3617 | """Cache config to avoid repeated calls to get_current_vllm_config()""" |
| 3618 | return get_current_vllm_config().compilation_config |
| 3619 | |
| 3620 | |
| 3621 | def get_current_vllm_config() -> VllmConfig: |
| 3622 | if _current_vllm_config is None: |
| 3623 | raise AssertionError( |
| 3624 | "Current vLLM config is not set. This typically means " |
| 3625 | "get_current_vllm_config() was called outside of a " |
| 3626 | "set_current_vllm_config() context, or a CustomOp was instantiated " |
| 3627 | "at module import time or model forward time when config is not set. " |
| 3628 | "For tests that directly test custom ops/modules, use the " |
| 3629 | "'default_vllm_config' pytest fixture from tests/conftest.py." |
| 3630 | ) |
| 3631 | return _current_vllm_config |
| 3632 | |
| 3633 | |
| 3634 | def get_current_vllm_config_or_none() -> VllmConfig | None: |
| 3635 | return _current_vllm_config |
| 3636 | |
| 3637 | |
| 3638 | T = TypeVar("T") |
| 3639 | |
| 3640 | |
| 3641 | def get_layers_from_vllm_config( |
| 3642 | vllm_config: VllmConfig, |
| 3643 | layer_type: type[T], |
| 3644 | layer_names: Iterable[str] | None = None, |
| 3645 | ) -> dict[str, T]: |
| 3646 | """Get layers from the vLLM config. |
| 3647 | |
| 3648 | Args: |
| 3649 | vllm_config: The vLLM config. |
| 3650 | layer_type: The type of the layer to get. |
| 3651 | layer_names: The names of the layers to get. If None, return all layers. |
| 3652 | |
| 3653 | """ |
| 3654 | forward_context = vllm_config.compilation_config.static_forward_context |
| 3655 | if layer_names is None: |
| 3656 | layer_names = forward_context.keys() |
| 3657 | |
| 3658 | return { |
| 3659 | layer_name: layer |
| 3660 | for layer_name in layer_names |
| 3661 | if isinstance(layer := forward_context.get(layer_name), layer_type) |
| 3662 | } |