compilation.py 47.5 KB
Newer Older
1
2
3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

4
import enum
5
from collections import Counter
6
from collections.abc import Callable
7
from dataclasses import field
8
from pathlib import Path
9
from typing import TYPE_CHECKING, Any, ClassVar, Literal
10

11
from pydantic import ConfigDict, Field, TypeAdapter, field_validator
12
13
from pydantic.dataclasses import dataclass

14
import vllm.envs as envs
15
from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
16
17
18
19
20
21
from vllm.config.utils import (
    Range,
    config,
    get_hash_factors,
    hash_factors,
)
22
from vllm.logger import init_logger
23
from vllm.platforms import current_platform
24
from vllm.utils.import_utils import resolve_obj_by_qualname
25
from vllm.utils.math_utils import round_up
26
from vllm.utils.torch_utils import is_torch_equal_or_newer
27
from vllm import envs
28
29

if TYPE_CHECKING:
30
    from vllm.config import VllmConfig
31
32
33
34
35
else:
    VllmConfig = object

logger = init_logger(__name__)

36
37
38
# Explicitly exports Range
__all__ = ["Range"]

39

40
class CompilationMode(enum.IntEnum):
41
42
43
44
45
46
47
48
49
50
51
52
53
    """The compilation approach used for torch.compile-based compilation of the
    model."""

    NONE = 0
    """No torch.compile compilation is applied, model runs in fully eager pytorch mode.
    The model runs as-is."""
    STOCK_TORCH_COMPILE = 1
    """The standard `torch.compile` compilation pipeline."""
    DYNAMO_TRACE_ONCE = 2
    """Single Dynamo trace through the model, avoiding recompilation."""
    VLLM_COMPILE = 3
    """Custom vLLM Inductor-based backend with caching, piecewise compilation,
    shape specialization, and custom passes."""
54
55


56
class CUDAGraphMode(enum.Enum):
57
    """Constants for the cudagraph mode in CompilationConfig.
58
59
60
    Meanwhile, the subset enum `NONE`, `PIECEWISE` and `FULL` are also
    treated as concrete runtime mode for cudagraph runtime dispatching.
    """
61

62
63
64
65
66
67
    NONE = 0
    PIECEWISE = 1
    FULL = 2
    FULL_DECODE_ONLY = (FULL, NONE)
    FULL_AND_PIECEWISE = (FULL, PIECEWISE)

68
69
    def decode_mode(self) -> "CUDAGraphMode":
        return CUDAGraphMode(self.value[0]) if self.separate_routine() else self
70

71
72
    def mixed_mode(self) -> "CUDAGraphMode":
        return CUDAGraphMode(self.value[1]) if self.separate_routine() else self
73

74
75
76
77
78
    def has_mode(self, mode: "CUDAGraphMode") -> bool:
        assert not mode.separate_routine()
        if self.separate_routine():
            return mode.value in self.value
        return self == mode
79
80

    def requires_piecewise_compilation(self) -> bool:
81
        return self.has_mode(CUDAGraphMode.PIECEWISE)
82

83
    def max_cudagraph_mode(self) -> "CUDAGraphMode":
84
        return CUDAGraphMode(max(self.value)) if self.separate_routine() else self
85
86
87
88

    def has_full_cudagraphs(self) -> bool:
        return self.max_cudagraph_mode() == CUDAGraphMode.FULL

89
90
91
    def has_piecewise_cudagraphs(self) -> bool:
        return self.requires_piecewise_compilation()

92
93
94
    def separate_routine(self) -> bool:
        return isinstance(self.value, tuple)

95
    def valid_runtime_modes(self) -> bool:
96
        return self in [CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL]
97

98
99
100
    def __str__(self) -> str:
        return self.name

101

102
@config
103
@dataclass(config=ConfigDict(extra="forbid"))
104
105
106
107
108
class PassConfig:
    """Configuration for custom Inductor passes.

    This is separate from general `CompilationConfig` so that inductor passes
    don't all have access to full configuration - that would create a cycle as
109
    the `PassManager` is set as a property of config.
110

111
112
113
114
115
116
    You must pass PassConfig to VLLMConfig constructor via the CompilationConfig
    constructor. VLLMConfig's post_init does further initialization.
    If used outside of the VLLMConfig, some fields may be left in an
    improper state.
    """

117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
    # New flags
    fuse_norm_quant: bool = Field(default=None)
    """Fuse the custom RMSNorm + quant ops."""
    fuse_act_quant: bool = Field(default=None)
    """Fuse the custom SiluMul + quant ops."""
    fuse_attn_quant: bool = Field(default=None)
    """Fuse the custom attention + quant ops."""
    eliminate_noops: bool = Field(default=None)
    """Eliminate no-op ops."""
    enable_sp: bool = Field(default=None)
    """Enable sequence parallelism."""
    fuse_gemm_comms: bool = Field(default=None)
    """Enable async TP."""
    fuse_allreduce_rms: bool = Field(default=None)
    """Enable flashinfer allreduce fusion."""

133
134
135
136
    fi_allreduce_fusion_max_size_mb: float | None = None
    """The threshold of the communicated tensor sizes under which
    vllm should use flashinfer fused allreduce. Specified as a
    float in MB.
137
    Unspecified will fallback to default values
138
139
140
141
142
143
144
145
146
147
148
149
150
    which are compute capability and world size dependent.
        FI_ALLREDUCE_FUSION_MAX_SIZE_MB = {
            90: {
                2: 64,  # 64MB
                4: 2,  # 2MB
                8: 1,  # 1MB
            },
            100: {
                2: 64,  # 64MB
                4: 32,  # 32MB
                8: 1,  # 1MB
            },
        }, where key is the device capability"""
151
    enable_qk_norm_rope_fusion: bool = False
152
    """Enable fused Q/K RMSNorm + RoPE pass."""
153
154
155

    # TODO(luka) better pass enabling system.

156
157
158
159
160
161
162
163
    def flashinfer_max_size(self, world_size: int) -> int | None:
        """
        Returns the max communication size in bytes for flashinfer
        allreduce fusion for the given world size. Returns None if world size
        is not supported by configs as it's not supported by flashinfer.
        """

        MiB = 1024 * 1024
164
165
166
        FI_SUPPORTED_WORLD_SIZES = [2, 4, 8]
        if world_size not in FI_SUPPORTED_WORLD_SIZES:
            return None
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
        max_size_mb = self.fi_allreduce_fusion_max_size_mb
        if max_size_mb is None:
            max_size_mb = self.default_fi_allreduce_fusion_max_size_mb().get(world_size)

        return int(max_size_mb * MiB) if max_size_mb is not None else None

    @staticmethod
    def default_fi_allreduce_fusion_max_size_mb() -> dict[int, float]:
        from vllm.compilation.collective_fusion import FI_ALLREDUCE_FUSION_MAX_SIZE_MB
        from vllm.platforms import current_platform

        if not current_platform.is_cuda():
            return {}
        return FI_ALLREDUCE_FUSION_MAX_SIZE_MB.get(
            current_platform.get_device_capability().to_int(), {}
        )

184
    def compute_hash(self) -> str:
185
186
187
188
189
        """
        Produces a hash unique to the pass configuration.
        Any new fields that affect compilation should be added to the hash.
        Any future fields that don't affect compilation should be excluded.
        """
190

191
        return hash_factors(get_hash_factors(self, set()))
192

193
    @field_validator(
194
195
196
197
198
199
200
        "fuse_norm_quant",
        "fuse_act_quant",
        "fuse_attn_quant",
        "eliminate_noops",
        "enable_sp",
        "fuse_gemm_comms",
        "fuse_allreduce_rms",
201
202
203
204
205
206
207
208
209
        mode="wrap",
    )
    @classmethod
    def _skip_none_validation(cls, value: Any, handler: Callable) -> Any:
        """Skip validation if the value is `None` when initialisation is delayed."""
        if value is None:
            return value
        return handler(value)

210
    def __post_init__(self) -> None:
211
212
213
214
        # Handle deprecation and defaults

        if not self.eliminate_noops:
            if self.fuse_norm_quant or self.fuse_act_quant:
215
216
                logger.warning_once(
                    "Fusion enabled but reshape elimination disabled. "
217
218
                    "RMSNorm/SiluMul + quant (fp8) fusion might not work"
                )
219
            if self.fuse_attn_quant:
220
221
                logger.warning_once(
                    "Fusion enabled but reshape elimination disabled. "
222
223
                    "Attention + quant (fp8) fusion might not work"
                )
224
            if self.fuse_allreduce_rms:
225
226
227
228
                logger.warning_once(
                    "Fusion enabled but reshape elimination disabled. "
                    "Allreduce + rms norm + quant (fp8) fusion might not work"
                )
229
        if self.enable_qk_norm_rope_fusion and not current_platform.is_cuda_alike():
230
231
            logger.warning_once(
                "QK Norm + RoPE fusion enabled but the current platform is not "
232
                "CUDA or ROCm. The fusion will be disabled."
233
234
            )
            self.enable_qk_norm_rope_fusion = False
235
236


237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
class DynamicShapesType(str, enum.Enum):
    """Types of dynamic shapes handling in torch.compile().
    see  Dynamic shapes and vllm guard dropping in torch_compile.md
    for more details."""

    BACKED = "backed"
    """Use backed dynamic shapes. torch.compile() guards on backed dynamic
    shapes and may add guards. Symbols are specialized to 0, 1, or >=2 even
    without encountering branching on those ranges."""

    UNBACKED = "unbacked"
    """Use unbacked dynamic shapes. Guaranteed not to be guarded on and not
    0/1 specialized, but may throw data dependent errors when branches require
    their value without explicit unbacked handling."""

    BACKED_SIZE_OBLIVIOUS = "backed_size_oblivious"
    """Experimental flag that treats backed symbols as unbacked when explicit
    unbacked handling is defined."""


@config
258
@dataclass(config=ConfigDict(extra="forbid"))
259
260
261
262
263
264
265
266
267
268
269
270
271
class DynamicShapesConfig:
    """Configuration to control/debug torch compile dynamic shapes."""

    type: DynamicShapesType = DynamicShapesType.BACKED
    """Controls the type of dynamic shapes handling to use with torch.compile().

    - BACKED: Default PyTorch behavior with potential guards ignored.
    - UNBACKED: No guards guaranteed (most sound) but may throw
      data dependent errors.
    - BACKED_SIZE_OBLIVIOUS: Experimental safer alternative to
      backed/unbacked.
    """

272
273
274
275
276
277
278
279
280
281
    evaluate_guards: bool = False
    """
    A debug mode to detect and fail if Dynamo ever specializes a dynamic shape by
    guarding on it. When True, dynamic shape guards are not dropped from dynamo.
    And a failure will be triggered if a recompilation ever happens due to that.
    This mode requires VLLM_USE_BYTECODE_HOOK to be 0.
    Enabling this allow observing the dynamic shapes guards in the tlparse
    artifacts also.
    When type is backed, aot_compile must be disabled for this mode to work.
    until this change picked up https://github.com/pytorch/pytorch/pull/169239.
282
    """
283

284
285
286
    assume_32_bit_indexing: bool = True
    """
    whether all tensor sizes can use 32 bit indexing.
287
    """
288
289
290
291
292
293
294
295
296
297

    def compute_hash(self) -> str:
        """
        Provide a hash for DynamicShapesConfig
        """

        from vllm.config.utils import get_hash_factors, hash_factors

        factors = get_hash_factors(self, {})
        return hash_factors(factors)
298
299
300


@config
301
@dataclass(config=ConfigDict(extra="forbid"))
302
class CompilationConfig:
303
304
305
306
307
308
309
    """Configuration for compilation.

    You must pass CompilationConfig to VLLMConfig constructor.
    VLLMConfig's post_init does further initialization. If used outside of the
    VLLMConfig, some fields will be left in an improper state.

    It has three parts:
310
311

    - Top-level Compilation control:
312
        - [`mode`][vllm.config.CompilationConfig.mode]
313
314
315
316
317
        - [`debug_dump_path`][vllm.config.CompilationConfig.debug_dump_path]
        - [`cache_dir`][vllm.config.CompilationConfig.cache_dir]
        - [`backend`][vllm.config.CompilationConfig.backend]
        - [`custom_ops`][vllm.config.CompilationConfig.custom_ops]
        - [`splitting_ops`][vllm.config.CompilationConfig.splitting_ops]
318
        - [`compile_mm_encoder`][vllm.config.CompilationConfig.compile_mm_encoder]
319
    - CudaGraph capture:
320
        - [`cudagraph_mode`][vllm.config.CompilationConfig.cudagraph_mode]
321
322
        - [`cudagraph_capture_sizes`]
        [vllm.config.CompilationConfig.cudagraph_capture_sizes]
323
324
        - [`max_cudagraph_capture_size`]
        [vllm.config.CompilationConfig.max_cudagraph_capture_size]
325
326
327
328
329
330
        - [`cudagraph_num_of_warmups`]
        [vllm.config.CompilationConfig.cudagraph_num_of_warmups]
        - [`cudagraph_copy_inputs`]
        [vllm.config.CompilationConfig.cudagraph_copy_inputs]
    - Inductor compilation:
        - [`compile_sizes`][vllm.config.CompilationConfig.compile_sizes]
331
332
        - [`compile_ranges_split_points`]
            [vllm.config.CompilationConfig.compile_ranges_split_points]
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
        - [`inductor_compile_config`]
        [vllm.config.CompilationConfig.inductor_compile_config]
        - [`inductor_passes`][vllm.config.CompilationConfig.inductor_passes]
        - custom inductor passes

    Why we have different sizes for cudagraph and inductor:
    - cudagraph: a cudagraph captured for a specific size can only be used
        for the same size. We need to capture all the sizes we want to use.
    - inductor: a graph compiled by inductor for a general shape can be used
        for different sizes. Inductor can also compile for specific sizes,
        where it can have more information to optimize the graph with fully
        static shapes. However, we find the general shape compilation is
        sufficient for most cases. It might be beneficial to compile for
        certain small batchsizes, where inductor is good at optimizing.
    """
348

349
    # Top-level Compilation control
350
    level: int = Field(default=None)
351
352
353
354
355
356
    """
    Level is deprecated and will be removed in the next release,
    either 0.12.0 or 0.11.2 whichever is soonest.
    Please use mode. Currently all levels are mapped to mode.
    """
    # Top-level Compilation control
357
    mode: CompilationMode = Field(default=None)
358
359
360
361
362
363
364
365
366
367
368
369
370
    """The compilation approach used for torch.compile-based compilation of the
    model.

    - None: If None, we will select the default compilation mode.
      For V1 engine this is 3.
    - 0: NONE: No torch.compile compilation is applied, model runs in fully
         eager pytorch mode. The model runs as-is.
    - 1: STOCK_TORCH_COMPILE: The standard `torch.compile` compilation pipeline.
    - 2: DYNAMO_TRACE_ONCE: Single Dynamo trace through the model, avoiding
         recompilation by removing guards.
         Requires no dynamic-shape-dependent control-flow.
    - 3: VLLM_COMPILE: Custom vLLM Inductor-based backend with caching,
         piecewise compilation, shape specialization, and custom passes."""
371
    debug_dump_path: Path | None = None
372
373
374
375
376
    """The path to dump the debug information."""
    cache_dir: str = ""
    """The directory to store the compiled graph, to accelerate Inductor
    compilation. By default, it will use model-related information to generate
    a cache directory."""
377
378
379
380
381
382
383
384
385
    compile_cache_save_format: Literal["binary", "unpacked"] = field(
        default_factory=lambda: envs.VLLM_COMPILE_CACHE_SAVE_FORMAT
    )
    """Format for saving torch compile cache:\n
    - "binary": saves as binary file (multiprocess safe)\n
    - "unpacked": saves as directory structure for inspection/debugging
    (NOT multiprocess safe)\n
    Defaults to `VLLM_COMPILE_CACHE_SAVE_FORMAT` if not specified.
    """
386
387
388
    backend: str = ""
    """The backend for compilation. It needs to be a string:

389
390
    - "" (empty string): use the default backend ("inductor" on CUDA-alike
    platforms).
391
392
393
394
395
    - "eager"/"openxla"/...: use the specified backend registered in PyTorch.
    - "full.module.name": a qualified name which can be used to import the

    backend function.
    We use string to avoid serialization issues when using compilation in a
396
    distributed setting. When the compilation mode is 1 or 2, the backend is
397
    used for the compilation directly (it sees the whole graph). When the
398
399
    compilation mode is 3, the backend supports both whole graph and piecewise
    compilation, available backends include eager, inductor, and custom backends,
400
    the latter of which can be defined via `get_compile_backend`. Furthermore,
401
    compilation is only piecewise if splitting ops is set accordingly and
402
    use_inductor_graph_partition is off. Note that the default options for
403
404
    splitting ops are sufficient for piecewise compilation.
    """
405
406
407
408
409
410
411
412
413
414
    custom_ops: list[str] = field(default_factory=list)
    """Fine-grained control over which custom ops to enable/disable. Use 'all'
    to enable all, 'none' to disable all. Also specify a list of custom op
    names to enable (prefixed with a '+'), or disable (prefixed with a '-').
    Examples:

    - 'all,-op1' to enable all except op1
    - 'none,+op1,+op2' to enable only op1 and op2

    By default, all custom ops are enabled when running without Inductor and
415
416
    disabled when running with Inductor: mode>CompilationMode.NONE and
    backend="inductor".
417
    Inductor generates (fused) Triton kernels for disabled custom ops."""
418
    splitting_ops: list[str] | None = None
419
    """A list of ops to exclude from cudagraphs, used in piecewise compilation.
420

421
422
423
424
425
426
    The behavior depends on use_inductor_graph_partition:

    - When use_inductor_graph_partition=False (default):
        These ops are used for Dynamo FX-level graph splitting. The graph is
        split at these ops before Inductor compilation, creating separate
        subgraphs for cudagraph capture.
427

428
429
430
431
432
    - When use_inductor_graph_partition=True:
        These ops are used to register Inductor partition rules. The graph
        partitioning happens at Inductor codegen time after all passes and
        fusions are finished, allowing compilation and custom passes to operate
        on the full graph while still excluding these ops from cudagraphs.
433

434
435
    If None, defaults to attention ops for piecewise cudagraphs.
    If empty list [], no ops are excluded (suitable for full cudagraphs)."""
436
    compile_mm_encoder: bool = False
Harry Mellor's avatar
Harry Mellor committed
437
    """Whether or not to compile the multimodal encoder.
438
439
440
    Currently, this only works for `Qwen2_5_vl` and `mLLaMa4` models
    on selected platforms. Disabled by default until more models
    are supported/tested to work."""
441
442

    # Inductor capture
443
    compile_sizes: list[int | str] | None = None
444
445
446
    """Sizes to compile for inductor. In addition
    to integers, it also supports "cudagraph_capture_sizes" to
    specify the sizes for cudagraph capture."""
447

448
449
    compile_ranges_split_points: list[int] | None = None
    """Split points that represent compile ranges for inductor.
450
451
452
    The compile ranges are
    [1, split_points[0]],
    [split_points[0] + 1, split_points[1]], ...,
453
454
455
    [split_points[-1] + 1, max_num_batched_tokens].
    Compile sizes are also used single element ranges,
    the range is represented as [compile_sizes[i], compile_sizes[i]].
456
457

    If a range overlaps with the compile size, graph for compile size
458
459
460
461
462
    will be prioritized, i.e. if we have a range [1, 8] and a compile size 4,
    graph for compile size 4 will be compiled and used instead of the graph
    for range [1, 8].
    """

463
464
465
    inductor_compile_config: dict = field(default_factory=dict)
    """Additional configurations for inductor.
    - None: use default configurations."""
466

467
468
469
470
471
472
473
474
    inductor_passes: dict[str, str] = field(default_factory=dict)
    """Additional passes for inductor. It is a dictionary
    from pass name to pass function qualified name. We use function
    name because the config uses JSON format. If we pass the config
    from Python, functions can also be passed directly via Python object
    constructor, e.g. `CompilationConfig(inductor_passes={"a": func})`."""

    # CudaGraph compilation
475
    cudagraph_mode: CUDAGraphMode = Field(default=None)
476
    """
Harry Mellor's avatar
Harry Mellor committed
477
478
    The mode of the cudagraph:

479
    - NONE, no cudagraph capture.
480
    - PIECEWISE.
481
482
    - FULL.
    - FULL_DECODE_ONLY.
483
    - FULL_AND_PIECEWISE. (v1 default)
484
485

    PIECEWISE mode build piecewise cudagraph only, keeping the cudagraph
co63oc's avatar
co63oc committed
486
    incompatible ops (i.e. some attention ops) outside the cudagraph
487
488
489
490
491
    for general flexibility.

    FULL mode: Capture full cudagraph for all batches. Can be good for small
    models or workloads with small prompts; not supported by many backends.
    Generally for performance FULL_AND_PIECEWISE is better.
492

493
494
495
496
    FULL_DECODE_ONLY mode: Capture full cudagraph for decode batches only.
    Mixed prefill-decode batches are run without cudagraphs. Can be good for
    decode instances in a P/D setup where prefill is not as important so we
    can save some memory.
497

498
499
    FULL_AND_PIECEWISE mode: Capture full cudagraph for decode batches and
    piecewise cudagraph for prefill and mixed prefill-decode batches.
500
    This is the most performant mode for most models and is the default.
501
502

    Currently, the cudagraph mode is only used for the v1 engine.
503
504
    Note that the cudagraph logic is generally orthogonal to the
    compilation logic. While piecewise cudagraphs require piecewise
505
    compilation (mode=VLLM_COMPILE and non-empty splitting_ops), full
506
    cudagraphs are supported with and without compilation.
507
508

    Warning: This flag is new and subject to change in addition
509
510
    more modes may be added.
    """
511
512
513
514
515
    cudagraph_num_of_warmups: int = 0
    """Number of warmup runs for cudagraph.
    It means the first several runs will be treated as warmup runs.
    Only after that, the execution will be recorded, and the recorded
    cudagraph will be used for subsequent runs."""
516
    cudagraph_capture_sizes: list[int] | None = None
517
518
519
520
521
522
523
524
    """Sizes to capture cudagraph.
    - None (default): capture sizes are inferred from vllm config.
    - list[int]: capture sizes are specified as given."""
    cudagraph_copy_inputs: bool = False
    """Whether to copy input tensors for
    cudagraph. If the caller can guarantee that the same input buffers
    are always used, it can set this to False. Otherwise, it should
    set this to True, and the compiler will copy the input to an
525
    internally managed buffer. Default is False.
526
527
    Note that this flag is only effective when cudagraph_mode is PIECEWISE.
    """
528
529
530
531
532
533
534
    cudagraph_specialize_lora: bool = True
    """Whether to create separate cuda graphs for cases with and without active
    LoRA adapters. When set to False, the LoRA-enabled cuda graph will be used
    for all cases, incurring the overhead of running LoRA ops even when no
    adapters are active. Setting this to True will remove this overhead at the
    cost of increased startup time and slightly higher memory usage.
    When `enable_lora` is False, this option has no effect.
535
    """
536

537
    use_inductor_graph_partition: bool = Field(default=None)
538
539
540
541
542
543
544
    """Use inductor graph partition to split the graph at cudagraph_unsafe ops.
    This partition happens at inductor codegen time after all passes and fusions
    are finished. It generates a single `call` function which wraps
    cudagraph-safe ops into partition functions and leave cudagraph-unsafe ops
    outside the partition functions. For a graph with N cudagraph-unsafe ops
    (e.g., Attention), there would be N+1 partitions. To mark an op as
    cudagraph unsafe, we can add `tags=(torch._C.Tag.cudagraph_unsafe)` when
545
    register the custom op.
546
547
548
549
550
551
552
553
554
555
556

    This config supports both full cudagraph and piecewise cudagraph without
    compiling twice. For piecewise cudagraph, it applies vLLM CUDAGraph wrapper
    to each partition. For N+1 partitions, there would be N+1
    CUDAGraph wrapper instances.

    For full CUDAGraph, we always apply a single CUDAGraph wrapper outside the
    inductor `call` function in the model runner. The top-level full cudagraph
    capture ignores all partitioning.
    """

557
558
559
    pass_config: PassConfig = field(default_factory=PassConfig)
    """Custom inductor passes, see PassConfig for more details"""

560
561
    max_cudagraph_capture_size: int | None = field(default=None)
    """The maximum cudagraph capture size.
562
563

    If cudagraph_capture_sizes is specified, this will be set to the largest
564
565
566
567
568
569
570
571
    size in that list (or checked for consistency if specified). If
    cudagraph_capture_sizes is not specified, the list of sizes is generated
    automatically following the pattern:

        [1, 2, 4] + list(range(8, 256, 8)) + list(
        range(256, max_cudagraph_capture_size + 1, 16))

    If not specified, max_cudagraph_capture_size is set to min(max_num_seqs*2,
572
    512) by default. This voids OOM in tight memory scenarios with small
573
574
575
    max_num_seqs, and prevents capture of many large graphs (>512) that would
    greatly increase startup time with limited performance benefit.
    """
576
577
578
579
580
581

    dynamic_shapes_config: DynamicShapesConfig = field(
        default_factory=DynamicShapesConfig
    )
    """Configuration for dynamic shapes options"""

582
583
    local_cache_dir: str = field(default=None, init=False)  # type: ignore
    """local cache dir for each rank"""
584

585
586
    bs_to_padded_graph_size: list[int] = field(
        default=None,  # type: ignore
587
588
        init=False,
    )
589
590
    """optimization:
    Intuitively, bs_to_padded_graph_size should be dict[int, int].
591
    since we know all keys are in a range [0, max_cudagraph_capture_size],
592
593
594
    we can optimize it to list[int] for better lookup performance."""

    # keep track of enabled and disabled custom ops
595
    enabled_custom_ops: Counter[str] = field(default_factory=Counter, init=False)
596
    """custom ops that are enabled"""
597
    disabled_custom_ops: Counter[str] = field(default_factory=Counter, init=False)
598
599
600
601
602
603
    """custom ops that are disabled"""
    traced_files: set[str] = field(default_factory=set, init=False)
    """files that are traced for compilation"""
    compilation_time: float = field(default=0.0, init=False)
    """time taken for compilation"""

604
    static_forward_context: dict[str, Any] = field(default_factory=dict, init=False)
605
606
607
608
    """Per-model forward context
    Map from layer name to layer objects that need to be accessed outside
    model code, e.g., Attention, FusedMOE when dp_size>1."""

609
    # Attention ops; used for piecewise cudagraphs
610
    # Use PyTorch operator format: "namespace::name"
611
    _attention_ops: ClassVar[list[str]] = [
612
613
614
615
616
617
618
619
620
        "vllm::unified_attention",
        "vllm::unified_attention_with_output",
        "vllm::unified_mla_attention",
        "vllm::unified_mla_attention_with_output",
        "vllm::mamba_mixer2",
        "vllm::mamba_mixer",
        "vllm::short_conv",
        "vllm::linear_attention",
        "vllm::plamo2_mamba_mixer",
621
        "vllm::gdn_attention_core",
622
        "vllm::kda_attention",
623
        "vllm::sparse_attn_indexer",
624
625
    ]

626
627
628
629
630
631
632
633
    def compute_hash(self) -> str:
        """
        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
634
635
636
637
638
639
640
641
642
643
644
645
646
647
        # Opt-out: default-include declared fields; keep a tiny exclude set;
        # normalize types; keep SHA-256. For nested opaque configs, include a
        # stable identifier (e.g., pass_config.compute_hash()) instead of object id.

        ignored_factors = {
            # Paths/dirs and runtime/metrics that don’t affect compiled graph
            "debug_dump_path",
            "cache_dir",
            "local_cache_dir",
            "bs_to_padded_graph_size",
            "traced_files",
            "compilation_time",
            "static_forward_context",
            "pass_config",  # handled separately below
648
            "dynamic_shapes_config",  # handled separately below
649
650
651
652
653
        }

        from vllm.config.utils import get_hash_factors, hash_factors

        factors = get_hash_factors(self, ignored_factors)
654

655
        factors["pass_config"] = self.pass_config.compute_hash()
656
        factors["dynamic_shapes_config"] = self.dynamic_shapes_config.compute_hash()
657
        return hash_factors(factors)
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679

    def __repr__(self) -> str:
        exclude = {
            "static_forward_context": True,
            "enabled_custom_ops": True,
            "disabled_custom_ops": True,
            "compilation_time": True,
            "bs_to_padded_graph_size": True,
            "traced_files": True,
            "inductor_compile_config": {
                "post_grad_custom_post_pass": True,
            },
        }

        # exclude default attr in pass_config
        pass_config_exclude = {}
        for attr, default_val in vars(PassConfig()).items():
            if getattr(self.pass_config, attr) == default_val:
                pass_config_exclude[attr] = True
        if pass_config_exclude:
            exclude["pass_config"] = pass_config_exclude

680
681
682
        config = TypeAdapter(CompilationConfig).dump_python(
            self, exclude=exclude, exclude_unset=True
        )
683
684

        return str(config)
685
686
687

    __str__ = __repr__

688
    @field_validator("mode", mode="before")
689
    @classmethod
690
    def validate_mode_before(cls, value: Any) -> Any:
691
        """
692
693
694
        Enable parsing the `mode` field from string mode names.
        Accepts both integers (0-3) and string names, like NONE, STOCK_TORCH_COMPILE,
        DYNAMO_TRACE_ONCE, VLLM_COMPILE.
695
        """
696
697
698
699
700
701
702
703
704
705
706
707
708
        if isinstance(value, str):
            # Convert string mode name to integer value
            mode_name = value.upper()

            if mode_name not in CompilationMode.__members__:
                raise ValueError(
                    f"Invalid compilation mode: {value}. "
                    f"Valid modes are: {', '.join(CompilationMode.__members__.keys())}"
                )

            return CompilationMode[mode_name]
        return value

709
710
711
    @field_validator("cudagraph_mode", mode="before")
    @classmethod
    def validate_cudagraph_mode_before(cls, value: Any) -> Any:
712
        """Enable parsing of the `cudagraph_mode` enum type from string."""
713
714
715
716
        if isinstance(value, str):
            return CUDAGraphMode[value.upper()]
        return value

717
718
719
720
721
722
723
724
    @field_validator("pass_config", mode="before")
    @classmethod
    def validate_pass_config_before(cls, value: Any) -> Any:
        """Enable parsing of the `pass_config` field from a dictionary."""
        if isinstance(value, dict):
            return PassConfig(**value)
        return value

725
726
727
728
729
730
731
732
733
734
    @field_validator("compile_cache_save_format")
    @classmethod
    def validate_compile_cache_save_format(cls, value: str) -> str:
        if value not in ("binary", "unpacked"):
            raise ValueError(
                f"compile_cache_save_format must be 'binary' or 'unpacked', "
                f"got: {value}"
            )
        return value

735
736
737
738
739
740
741
742
743
744
745
746
747
748
    @field_validator(
        "level",
        "mode",
        "cudagraph_mode",
        "use_inductor_graph_partition",
        mode="wrap",
    )
    @classmethod
    def _skip_none_validation(cls, value: Any, handler: Callable) -> Any:
        """Skip validation if the value is `None` when initialisation is delayed."""
        if value is None:
            return value
        return handler(value)

749
    def __post_init__(self) -> None:
750
751
752
753
754
755
756
757
758
759
760
        if self.level is not None:
            logger.warning(
                "Level is deprecated and will be removed in the next release,"
                "either 0.12.0 or 0.11.2 whichever is soonest."
                "Use mode instead."
                "If both level and mode are given,"
                "only mode will be used."
            )
            if self.mode is None:
                self.mode = self.level

761
762
763
764
765
766
767
768
769
770
771
772
773
        count_none = self.custom_ops.count("none")
        count_all = self.custom_ops.count("all")
        assert count_none + count_all <= 1, "Can only specify 'none' or 'all'"

        # TODO(zou3519/luka): There are 2 issues with auto-functionalization V2:
        # 1. A bug in PyTorch, fixed in 2.7:
        #    https://github.com/pytorch/pytorch/issues/147924
        # 2. Custom passes (fusion) rely on auto-functionalization V1 and don't
        #    work with V2. Addressing this will take extra engineering effort
        #    and it is not yet a priority. RFC here:
        #    https://github.com/vllm-project/vllm/issues/14703

        if is_torch_equal_or_newer("2.6"):
774
            KEY = "enable_auto_functionalized_v2"
775
776
777
778
779
            if KEY not in self.inductor_compile_config:
                self.inductor_compile_config[KEY] = False

        for k, v in self.inductor_passes.items():
            if not isinstance(v, str):
780
781
782
783
                assert callable(v), f"pass {k} should be callable or a qualified name"
                self.inductor_compile_config[k] = (
                    v if isinstance(v, InductorPass) else CallableInductorPass(v)
                )
784
785
786
787
788
789
790
                continue

            # resolve function from qualified name
            names = v.split(".")
            module = ".".join(names[:-1])
            func_name = names[-1]
            func = __import__(module).__dict__[func_name]
791
792
793
            self.inductor_compile_config[k] = (
                func if isinstance(func, InductorPass) else CallableInductorPass(func)
            )
794

795
796
797
798
799
        if self.pass_config.enable_qk_norm_rope_fusion:
            # TODO(zhuhaoran): support rope native forward match and remove this.
            # Linked issue: https://github.com/vllm-project/vllm/issues/28042
            self.custom_ops.append("+rotary_embedding")

800
801
802
803
        if (
            is_torch_equal_or_newer("2.9.0.dev")
            and "combo_kernels" not in self.inductor_compile_config
            and "benchmark_combo_kernel" not in self.inductor_compile_config
804
805
            # (fixme @boyuan) combo kernel does not support cpu yet.
            and not current_platform.is_cpu()
806
807
808
809
810
811
        ):
            # use horizontal fusion, which is useful for fusing qk-norm and
            # qk-rope when query and key have different shapes.
            self.inductor_compile_config["combo_kernels"] = True
            self.inductor_compile_config["benchmark_combo_kernel"] = True

812
813
814
815
816
817
818
819
        if self.use_inductor_graph_partition and not is_torch_equal_or_newer(
            "2.9.0.dev"
        ):
            raise ValueError(
                "use_inductor_graph_partition is only "
                "supported with torch>=2.9.0.dev. Set "
                "use_inductor_graph_partition=False instead."
            )
820

821
        for op in self.custom_ops:
822
            if op[0] not in {"+", "-"} and op not in {"all", "none"}:
823
                raise ValueError(
824
825
826
827
                    f"Invalid syntax '{op}' for custom op, "
                    "must be 'all', 'none', '+op' or '-op' "
                    "(where 'op' is the registered op name)"
                )
828

829
830
831
        # Currently only eager and inductor backend are supported.
        # for piecewise compilation. Custom backends are not suppported for
        # piecewise compilation. Update when more backends are supported.
832
        if self.mode == CompilationMode.VLLM_COMPILE and self.backend not in [
833
834
835
836
837
838
839
            "",
            "eager",
            "inductor",
        ]:
            raise ValueError(
                f"Invalid backend for piecewise compilation: {self.backend}"
            )
840

841
        if self.backend == "":
842
            self.backend = current_platform.get_compile_backend()
843

844
    def init_backend(self, vllm_config: "VllmConfig") -> str | Callable:
845
846
847
848
849
850
851
        """
        Initialize the backend for the compilation config from a vllm config.
        Arguments:
            vllm_config: The vllm config to initialize the backend from.
        Returns:
            The backend for the compilation config.
        """
852
        if self.mode is None:
853
            raise ValueError(
854
855
856
                "No compilation mode is set. This method should only be "
                "called via vllm config where the level is set if none is "
                "provided."
857
            )
858
859
        if self.mode == CompilationMode.NONE:
            raise ValueError("No compilation mode is set.")
860
861

        from torch._dynamo.backends.registry import list_backends
862

863
        torch_backends = list_backends(exclude_tags=tuple())
864
865
866
        if self.mode in [
            CompilationMode.STOCK_TORCH_COMPILE,
            CompilationMode.DYNAMO_TRACE_ONCE,
867
868
869
870
871
        ]:
            if self.backend in torch_backends:
                return self.backend
            return resolve_obj_by_qualname(self.backend)

872
        assert self.mode == CompilationMode.VLLM_COMPILE
873
        if self.backend not in ["eager", "inductor"]:
874
            logger.info("Using OOT custom backend for compilation.")
875
876
877

        from vllm.compilation.backends import VllmBackend

878
879
        # TODO[@lucaskabela]: See if we can forward prefix
        # https://github.com/vllm-project/vllm/issues/27045
880
881
        return VllmBackend(vllm_config)

882
883
884
885
886
887
    def post_init_cudagraph_sizes(self) -> None:
        """To complete the initialization after cudagraph related
        configs are set. This includes:
        - initialize compile_sizes
        - pre-compute the mapping bs_to_padded_graph_size
        """
888
889
890
891
892
893
894

        computed_compile_sizes = []
        if self.compile_sizes is not None:
            # de-duplicate the sizes provided by the config
            self.compile_sizes = list(set(self.compile_sizes))
            for x in self.compile_sizes:
                if isinstance(x, str):
895
896
                    assert x == "cudagraph_capture_sizes", (
                        "Unrecognized size type in compile_sizes, "
897
                        f"expect 'cudagraph_capture_sizes', got {x}"
898
                    )
899
900
901
902
903
904
                    computed_compile_sizes.extend(self.cudagraph_capture_sizes)
                else:
                    assert isinstance(x, int)
                    computed_compile_sizes.append(x)
        self.compile_sizes = computed_compile_sizes  # type: ignore

905
906
907
908
        # make sure the sizes are in ascending order
        self.cudagraph_capture_sizes.sort()
        if self.cudagraph_capture_sizes:
            assert self.cudagraph_capture_sizes[-1] == self.max_cudagraph_capture_size
909

910
911
        # May get recomputed in the model runner if adjustment is needed for spec-decode
        self.compute_bs_to_padded_graph_size()
912

913
    def set_splitting_ops_for_v1(
914
        self, all2all_backend: str, data_parallel_size: int = 1
915
    ):
916
917
918
919
920
921
922
        # To compatible with OOT hardware plugin platform (for example vllm-ascend)
        # which currently only supports sequence parallelism in eager mode.
        if self.mode != CompilationMode.VLLM_COMPILE:
            if self.splitting_ops is None:
                self.splitting_ops = []
            return

923
924
925
        # NOTE: this function needs to be called only when mode is
        # CompilationMode.VLLM_COMPILE
        assert self.mode == CompilationMode.VLLM_COMPILE, (
926
            "set_splitting_ops_for_v1 should only be called when "
927
            "mode is CompilationMode.VLLM_COMPILE"
928
        )
929

930
        if self.pass_config.fuse_attn_quant and not self.use_inductor_graph_partition:
931
            self.set_splitting_ops_for_attn_fusion()
932
933
934
935
936
937
938
939
940
941
942
943
        else:
            if self.splitting_ops is None:
                # NOTE: When using full cudagraph, instead of setting an empty
                # list and capture the full cudagraph inside the flattened fx
                # graph, we keep the piecewise fx graph structure but capture
                # the full cudagraph outside the fx graph. This reduces some
                # cpu overhead when the runtime batch_size is not cudagraph
                # captured. see https://github.com/vllm-project/vllm/pull/20059
                # for details. Make a copy to avoid mutating the class-level
                # list via reference.
                self.splitting_ops = list(self._attention_ops)
            elif len(self.splitting_ops) == 0:
944
945
946
947
948
                if (
                    self.cudagraph_mode == CUDAGraphMode.PIECEWISE
                    or self.cudagraph_mode == CUDAGraphMode.FULL_AND_PIECEWISE
                ):
                    logger.warning_once(
949
                        "Using piecewise cudagraph with empty splitting_ops"
950
                    )
951
952
                if self.cudagraph_mode == CUDAGraphMode.PIECEWISE:
                    logger.warning_once(
953
954
                        "Piecewise compilation with empty splitting_ops does not "
                        "contain piecewise cudagraph. Setting cudagraph_"
955
956
957
958
959
960
961
962
                        "mode to NONE. Hint: If you are using attention "
                        "backends that support cudagraph, consider manually "
                        "setting cudagraph_mode to FULL or FULL_DECODE_ONLY "
                        "to enable full cudagraphs."
                    )
                    self.cudagraph_mode = CUDAGraphMode.NONE
                elif self.cudagraph_mode == CUDAGraphMode.FULL_AND_PIECEWISE:
                    logger.warning_once(
963
964
                        "Piecewise compilation with empty splitting_ops does "
                        "not contain piecewise cudagraph. Setting "
965
966
967
968
                        "cudagraph_mode to FULL."
                    )
                    self.cudagraph_mode = CUDAGraphMode.FULL
                self.splitting_ops = []
969

970
971
        # Disable CUDA graphs for DeepEP high-throughput since its not CG compatible
        if (
972
973
            all2all_backend == "deepep_high_throughput"
            and data_parallel_size > 1
974
975
976
977
978
979
980
981
982
983
984
            and self.cudagraph_mode != CUDAGraphMode.NONE
        ):
            # TODO: Piecewise Cuda graph might be enabled
            # if torch compile cache key issue fixed
            # See https://github.com/vllm-project/vllm/pull/25093
            logger.info(
                "DeepEP: Disabling CUDA Graphs since DeepEP high-throughput kernels "
                "are optimized for prefill and are incompatible with CUDA Graphs. "
                "In order to use CUDA Graphs for decode-optimized workloads, "
                "use --all2all-backend with another option, such as "
                "deepep_low_latency, pplx, or allgather_reducescatter."
985
            )
986
            self.cudagraph_mode = CUDAGraphMode.NONE
987
988

    def set_splitting_ops_for_attn_fusion(self):
989
        assert self.pass_config.fuse_attn_quant
990
        if self.splitting_ops is None:
991
            self.splitting_ops = []
992
993
            if self.cudagraph_mode.has_piecewise_cudagraphs():
                logger.warning_once(
994
                    "fuse_attn_quant is incompatible with piecewise "
995
996
997
998
999
1000
1001
1002
                    "cudagraph when use_inductor_graph_partition is off. "
                    "In this case, splitting_ops will be set to empty "
                    "list, and cudagraph_mode will be set to FULL. "
                    "Please ensure you are using attention backends that "
                    "support cudagraph or set cudagraph_mode to NONE "
                    "explicitly if encountering any problems."
                )
                self.cudagraph_mode = CUDAGraphMode.FULL
1003
1004

        assert not self.splitting_ops_contain_attention(), (
1005
            "attention ops should not be in splitting_ops when fuse_attn_quant is True"
1006
        )
1007
1008
1009

    def splitting_ops_contain_attention(self) -> bool:
        return self.splitting_ops is not None and all(
1010
1011
            op in self.splitting_ops for op in self._attention_ops
        )
1012
1013

    def is_attention_compiled_piecewise(self) -> bool:
1014
1015
        if not self.splitting_ops_contain_attention():
            return False
1016

1017
1018
        if not self.use_inductor_graph_partition:
            # Dynamo-level FX split case
1019
            return self.mode == CompilationMode.VLLM_COMPILE
1020

1021
        # Inductor partition case
1022
        return self.backend == "inductor" and self.mode != CompilationMode.NONE
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038

    def custom_op_log_check(self):
        """
        This method logs the enabled/disabled custom ops and checks that the
        passed custom_ops field only contains relevant ops.
        It is called at the end of set_current_vllm_config,
        after the custom ops have been instantiated.
        """

        if len(self.enabled_custom_ops) + len(self.disabled_custom_ops) == 0:
            logger.debug("No custom ops found in model.")
            return

        logger.debug("enabled custom ops: %s", self.enabled_custom_ops)
        logger.debug("disabled custom ops: %s", self.disabled_custom_ops)

1039
        all_ops_in_model = self.enabled_custom_ops | self.disabled_custom_ops
1040
1041
1042
1043
        for op in self.custom_ops:
            if op in {"all", "none"}:
                continue

1044
1045
1046
            assert op[0] in {"+", "-"}, (
                "Invalid custom op syntax (should be checked during init)"
            )
1047
1048
1049
1050
1051
1052
1053
1054

            # check if op name exists in model
            op_name = op[1:]
            if op_name not in all_ops_in_model:
                from vllm.model_executor.custom_op import CustomOp

                # Does op exist at all or is it just not present in this model?
                # Note: Only imported op classes appear in the registry.
1055
1056
1057
                missing_str = (
                    "doesn't exist (or wasn't imported/registered)"
                    if op_name not in CustomOp.op_registry
1058
                    else "not present in model"
1059
                )
1060

1061
1062
1063
1064
1065
1066
1067
1068
                enable_str = "enabling" if op[0] == "+" else "disabling"
                logger.warning_once(
                    "Op '%s' %s, %s with '%s' has no effect",
                    op_name,
                    missing_str,
                    enable_str,
                    op,
                )
1069

1070
1071
1072
1073
1074
1075
1076
    def is_custom_op_enabled(self, op: str) -> bool:
        if "all" in self.custom_ops:
            return f"-{op}" not in self.custom_ops

        assert "none" in self.custom_ops
        return f"+{op}" in self.custom_ops

1077
1078
1079
1080
    def adjust_cudagraph_sizes_for_spec_decode(
        self, uniform_decode_query_len: int, tensor_parallel_size: int
    ):
        multiple_of = uniform_decode_query_len
1081
        if tensor_parallel_size > 1 and self.pass_config.enable_sp:
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
            multiple_of = max(uniform_decode_query_len, tensor_parallel_size)
            if (
                multiple_of % uniform_decode_query_len != 0
                or multiple_of % tensor_parallel_size != 0
            ):
                raise ValueError(
                    f"Can't determine cudagraph shapes that are both a "
                    f"multiple of {uniform_decode_query_len} "
                    f"(num_speculative_tokens + 1) required by spec-decode "
                    f"and {tensor_parallel_size} (tensor_parallel_size) "
                    f"required by sequence parallelism please adjust "
                    f"num_speculative_tokens or disable sequence parallelism"
                )

        if not self.cudagraph_capture_sizes or multiple_of <= 1:
            return

        assert self.max_cudagraph_capture_size is not None
        rounded_sizes = sorted(
            set(
                round_up(size, multiple_of)
                for size in self.cudagraph_capture_sizes
                if round_up(size, multiple_of) <= self.max_cudagraph_capture_size
            )
        )

1108
1109
1110
1111
        if len(rounded_sizes) == 0 and multiple_of <= self.max_cudagraph_capture_size:
            # if one valid but would be round_down use that
            rounded_sizes = [multiple_of]

1112
        if len(rounded_sizes) == 0:
1113
1114
1115
1116
1117
1118
            raise ValueError(
                f"No valid cudagraph sizes after rounding to multiple of {multiple_of} "
                f"(num_speculative_tokens + 1 or tp if sequence parallelism is enabled)"
                f" please adjust num_speculative_tokens ({uniform_decode_query_len - 1}"
                f") or max_cudagraph_capture_size ({self.max_cudagraph_capture_size})"
                f" or cudagraph_capture_sizes ({self.cudagraph_capture_sizes})"
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
            )

        self.max_cudagraph_capture_size = rounded_sizes[-1]
        self.cudagraph_capture_sizes = rounded_sizes

        # Recompute after adjusting the cudagraph sizes
        self.compute_bs_to_padded_graph_size()

    def compute_bs_to_padded_graph_size(self):
        # pre-compute the mapping from batch size to padded graph size
        self.bs_to_padded_graph_size = [
            0 for i in range(self.max_cudagraph_capture_size + 1)
        ]
        for end, start in zip(
            self.cudagraph_capture_sizes + [self.max_cudagraph_capture_size + 1],
            [0] + self.cudagraph_capture_sizes,
        ):
            for bs in range(start, end):
                if bs == start:
                    self.bs_to_padded_graph_size[bs] = start
                else:
                    self.bs_to_padded_graph_size[bs] = end
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150

    def get_compile_ranges(self) -> list[Range]:
        """Get the compile ranges for the compilation config."""
        if self.compile_ranges_split_points is None:
            return []
        split_points = sorted(set(self.compile_ranges_split_points))
        return [
            Range(start=s + 1, end=e)
            for s, e in zip([0] + split_points[:-1], split_points)
        ]