"vllm/entrypoints/openai/speech_to_text.py" did not exist on "8359f4c8d840d409fe698d9bf428ecd2f7e85b75"
backends.py 48.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import ast
5
import dataclasses
6
import hashlib
7
import json
8
import operator
9
10
import os
import pprint
11
import time
12
from collections import defaultdict
13
from collections.abc import Callable, Generator, Sequence
14
from contextlib import contextmanager
15
from copy import deepcopy
16
from functools import partial
17
from typing import Any
18
19
20

import torch
import torch.fx as fx
21
from torch._dynamo.utils import dynamo_timed
22
from torch._logging._internal import trace_structured
23
from torch.fx._lazy_graph_module import _use_lazy_graph_module
24

25
import vllm.envs as envs
26
from vllm.config import CompilationConfig, CUDAGraphMode, VllmConfig
27
from vllm.config.compilation import DynamicShapesType
28
from vllm.config.utils import Range, hash_factors
29
from vllm.logger import init_logger
30
from vllm.logging_utils import lazy
31
from vllm.platforms import current_platform
32
from vllm.tracing import instrument, instrument_manual
33
from vllm.utils.import_utils import resolve_obj_by_qualname
34

35
36
37
38
39
from .compiler_interface import (
    CompilerInterface,
    EagerAdaptor,
    InductorAdaptor,
    InductorStandaloneAdaptor,
40
    is_compile_cache_enabled,
41
)
42
from .counter import compilation_counter
43
44
45
46
47
48
from .partition_rules import (
    inductor_partition_rule_context,
    should_split,
)
from .passes.inductor_pass import InductorPass, pass_context
from .passes.pass_manager import PostGradPassManager
49
50
51

logger = init_logger(__name__)

52

53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
def make_copy_and_call(
    sym_tensor_indices: list[int],
    input_buffers: list[torch.Tensor | None],
    callable_fn: Callable[..., Any],
) -> Callable[..., Any]:
    """Create a wrapper that copies inputs to static buffers before calling.

    This is used for cudagraph input copying where we need to copy dynamic
    tensors to static buffers before invoking the compiled graph.

    Args:
        sym_tensor_indices: Indices of tensors with symbolic shapes
        input_buffers: List of static buffers (can contain None for lazy init)
        callable_fn: The compiled function to call

    Returns:
        A wrapper function that copies inputs and calls the compiled function
    """

72
    def copy_and_call(*args: Any) -> Any:
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
        list_args = list(args)
        for i, index in enumerate(sym_tensor_indices):
            runtime_tensor = list_args[index]
            runtime_shape = runtime_tensor.shape[0]

            # lazy initialization of buffer on first call
            if input_buffers[i] is None:
                input_buffers[i] = runtime_tensor.clone()

            static_tensor = input_buffers[i][:runtime_shape]  # type: ignore[index]
            static_tensor.copy_(runtime_tensor)
            list_args[index] = static_tensor
        return callable_fn(*list_args)

    return copy_and_call


90
def make_compiler(compilation_config: CompilationConfig) -> CompilerInterface:
91
92
93
94
    assert not envs.VLLM_USE_MEGA_AOT_ARTIFACT or envs.VLLM_USE_STANDALONE_COMPILE, (
        "VLLM_USE_MEGA_AOT_ARTIFACT=1 requires VLLM_USE_STANDALONE_COMPILE=1"
    )

95
    if compilation_config.backend == "inductor":
96
97
        # Use standalone compile only if requested, version is new enough,
        # and the symbol actually exists in this PyTorch build.
98
99
        if envs.VLLM_USE_STANDALONE_COMPILE and hasattr(
            torch._inductor, "standalone_compile"
100
        ):
101
            logger.debug("Using InductorStandaloneAdaptor")
102
103
104
            return InductorStandaloneAdaptor(
                compilation_config.compile_cache_save_format
            )
105
        else:
106
            logger.debug("Using InductorAdaptor")
107
            return InductorAdaptor()
108
    elif compilation_config.backend == "eager":
109
        logger.debug("Using EagerAdaptor")
110
        return EagerAdaptor()
111
112
113
114
115
    else:
        logger.debug("Using custom backend: %s", compilation_config.backend)
        compiler = resolve_obj_by_qualname(current_platform.get_compile_backend())()
        assert isinstance(compiler, CompilerInterface)
        return compiler
116
117


118
119
120
121
122
class CompilerManager:
    """
    A manager to manage the compilation process, including
    caching the compiled graph, loading the compiled graph,
    and compiling the graph.
123

124
    The cache is a dict mapping
125
    `(runtime_shape, graph_index, backend_name)`
126
    to `any_data` returned from the compiler.
127

128
129
130
    When serializing the cache, we save it to a Python file
    for readability. We don't use json here because json doesn't
    support int as key.
131
132
    """

133
    def __init__(self, compilation_config: CompilationConfig) -> None:
134
        self.cache: dict[tuple[Range, int, str], Any] = dict()
135
        self.is_cache_updated = False
136
137
        self.compilation_config = compilation_config
        self.compiler = make_compiler(compilation_config)
138
        self.loaded_artifacts: dict[str, Any] = {}
139

140
141
    def compute_hash(self, vllm_config: VllmConfig) -> str:
        return self.compiler.compute_hash(vllm_config)
142

143
    @contextmanager
144
    def compile_context(self, compile_range: Range) -> Generator[None, None, None]:
145
146
147
        """Provide compilation context for the duration of compilation to set
        any torch global properties we want to scope to a single Inductor
        compilation (e.g. partition rules, pass context)."""
148
        with pass_context(compile_range):
149
            if self.compilation_config.use_inductor_graph_partition:
150
                with inductor_partition_rule_context(
151
                    self.compilation_config.splitting_ops
152
                ):
153
154
155
156
                    yield
            else:
                yield

157
158
    def initialize_cache(
        self, cache_dir: str, disable_cache: bool = False, prefix: str = ""
159
    ) -> None:
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
        """
        Initialize the cache directory for the compiler.

        The organization of the cache directory is as follows:
        cache_dir=/path/to/hash_str/rank_i_j/prefix/
        inside cache_dir, there will be:
        - vllm_compile_cache.py
        - computation_graph.py
        - transformed_code.py

        for multiple prefixes, they can share the same
        base cache dir of /path/to/hash_str/rank_i_j/ ,
        to store some common compilation artifacts.
        """

175
        self.disable_cache = disable_cache
176
        self.cache_dir = cache_dir
177
178
179
180
        self.cache_file_path = os.path.join(cache_dir, "vllm_compile_cache.py")

        if not disable_cache and os.path.exists(self.cache_file_path):
            # load the cache from the file
181
            with open(self.cache_file_path) as f:
182
183
184
                # we use ast.literal_eval to parse the data
                # because it is a safe way to parse Python literals.
                # do not use eval(), it is unsafe.
185
186
                cache = ast.literal_eval(f.read())

187
            def check_type(value: Any, ty: type) -> None:
188
189
190
                if not isinstance(value, ty):
                    raise TypeError(f"Expected {ty} but got {type(value)} for {value}")

191
192
193
            def parse_key(key: Any) -> tuple[Range, int, str]:
                range_tuple, graph_index, compiler_name = key
                check_type(graph_index, int)
194
195
196
197
198
199
200
                check_type(compiler_name, str)
                if isinstance(range_tuple, tuple):
                    start, end = range_tuple
                    check_type(start, int)
                    check_type(end, int)
                    range_tuple = Range(start=start, end=end)
                check_type(range_tuple, Range)
201
                return range_tuple, graph_index, compiler_name
202
203

            self.cache = {parse_key(key): value for key, value in cache.items()}
204

205
206
207
        self.compiler.initialize_cache(
            cache_dir=cache_dir, disable_cache=disable_cache, prefix=prefix
        )
208

209
    def save_to_file(self) -> None:
210
        if self.disable_cache or not self.is_cache_updated:
211
            return
212
213
        printer = pprint.PrettyPrinter(indent=4)
        data = printer.pformat(self.cache)
214
        with open(self.cache_file_path, "w") as f:
215
216
            f.write(data)

217
218
219
220
    def load(
        self,
        graph: fx.GraphModule,
        example_inputs: list[Any],
221
        graph_index: int,
222
        compile_range: Range,
223
    ) -> Callable[..., Any] | None:
224
        if (compile_range, graph_index, self.compiler.name) not in self.cache:
225
            return None
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243

        def parse_value(value: Any) -> tuple[tuple[str, str], str]:
            assert isinstance(value, dict)
            handle = value["graph_handle"]
            assert isinstance(handle[0], str)
            assert isinstance(handle[1], str)
            cache_key = value["cache_key"]
            return handle, cache_key

        try:
            handle, cache_key = parse_value(
                self.cache[(compile_range, graph_index, self.compiler.name)]
            )
        except Exception:
            # When the cache is outdated, we should ignore the existing file.
            # This should cause the correct cache to be generated again.
            return None

244
        compiled_graph = self.compiler.load(
245
            handle, graph, example_inputs, graph_index, compile_range
246
        )
247
        self.loaded_artifacts[cache_key] = compiled_graph
248
        logger.debug(
249
250
            "Directly load the %s-th graph for compile range %sfrom %s via handle %s",
            graph_index,
251
252
253
            str(compile_range),
            self.compiler.name,
            handle,
254
        )
255
256
        return compiled_graph

257
    @instrument(span_name="Compile graph")
258
259
260
    def compile(
        self,
        graph: fx.GraphModule,
261
        example_inputs: list[Any],
262
        additional_inductor_config: dict[str, Any],
263
        compilation_config: CompilationConfig,
264
        compile_range: Range,
265
266
267
        graph_index: int = 0,
        num_graphs: int = 1,
    ) -> Any:
268
        if graph_index == 0:
269
270
            # before compiling the first graph, record the start time
            global compilation_start_time
271
            compilation_start_time = time.perf_counter()
272
273
274
275
276
277

        compilation_counter.num_backend_compilations += 1

        compiled_graph = None

        # try to load from the cache
278
        compiled_graph = self.load(graph, example_inputs, graph_index, compile_range)
279
        if compiled_graph is not None:
280
281
282
            if graph_index == num_graphs - 1:
                # after loading the last graph for this shape, record the time.
                # there can be multiple graphs due to piecewise compilation.
283
                elapsed = time.perf_counter() - compilation_start_time
284
                compilation_config.compilation_time += elapsed
285
                logger.info_once(
286
287
288
289
                    "Directly load the compiled graph(s) for compile range %s "
                    "from the cache, took %.3f s",
                    str(compile_range),
                    elapsed,
290
                    scope="local",
291
                )
292
293
294
295
            return compiled_graph

        # no compiler cached the graph, or the cache is disabled,
        # we need to compile it
296
297
298
299
        if isinstance(self.compiler, InductorAdaptor):
            # Let compile_fx generate a key for us
            maybe_key = None
        else:
300
301
302
303
            maybe_key = "artifact_compile_range_"
            maybe_key += f"{compile_range.start}_{compile_range.end}"
            maybe_key += f"_subgraph_{graph_index}"
        with self.compile_context(compile_range):
304
305
            # There is a compilation time optimization here.
            #
306
            # If the (input metadata, graph, compiler config) are the same, then
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
            # we want to avoid compiling the same artifact again. If we didn't
            # do this optimization, the backend compilation (InductorAdaptor or
            # InductorStandaloneAdaptor)
            # is able to cache hit and produce an artifact faster if it was
            # already created, but it is still a duplicate artifact that
            # requires unnecessary things e.g. disk IO.
            #
            # The optimization is: If the backend compilation cache hits,
            # then do an early return from the backend compilation and look up
            # which of the previous in-memory artifacts we created to reuse.
            #
            # We implemented this by monkey-patching torch (torch does not
            # easily expose the cache_key function), but in the future torch
            # should expose the cache_key function that we can just call
            # directly before invoking backend compilation.
            cache_key = None
            orig = torch._functorch._aot_autograd.autograd_cache.autograd_cache_key

            def autograd_cache_key(*args, **kwargs):
                result = orig(*args, **kwargs)
                if result is None:
                    return None
                nonlocal cache_key
                cache_key = result[0]
                if cache_key in self.loaded_artifacts:
                    raise StopCompiling()
                return result

            from unittest.mock import patch

            with (
                # Graphs that are isometric (different node names but same
                # structure) should be treated as the same.
                torch._functorch.config.patch(autograd_cache_normalize_inputs=True),
                patch(
                    "torch._functorch._aot_autograd.autograd_cache.autograd_cache_key",
                    autograd_cache_key,
                ),
            ):
                try:
                    compiled_graph, handle = self.compiler.compile(
                        graph,
                        example_inputs,
                        additional_inductor_config,
                        compile_range,
                        maybe_key,
                    )
                except StopCompiling:
                    assert cache_key is not None
                    return self.loaded_artifacts[cache_key]
            if cache_key is not None and compiled_graph is not None:
                self.loaded_artifacts[cache_key] = compiled_graph
359
360
361
362

        assert compiled_graph is not None, "Failed to compile the graph"

        # store the artifact in the cache
363
        if is_compile_cache_enabled(additional_inductor_config) and handle is not None:
364
365
366
367
            self.cache[(compile_range, graph_index, self.compiler.name)] = {
                "graph_handle": handle,
                "cache_key": cache_key,
            }
368
            compilation_counter.num_cache_entries_updated += 1
369
            self.is_cache_updated = True
370
371
            if graph_index == 0:
                # adds some info logging for the first graph
372
373
374
                logger.info_once(
                    "Cache the graph of compile range %s for later use",
                    str(compile_range),
375
                    scope="local",
376
                )
377
            logger.debug_once(
378
379
380
381
382
                "Store the %s-th graph for compile range%s from %s via handle %s",
                graph_index,
                str(compile_range),
                self.compiler.name,
                handle,
383
                scope="local",
384
            )
385
386
387

        # after compiling the last graph, record the end time
        if graph_index == num_graphs - 1:
388
            elapsed = time.perf_counter() - compilation_start_time
389
            compilation_config.compilation_time += elapsed
390
391
392
393
394
395
            logger.info_once(
                "Compiling a graph for compile range %s takes %.2f s",
                str(compile_range),
                elapsed,
                scope="local",
            )
396

397
        return compiled_graph
398
399


400
401
402
403
class StopCompiling(BaseException):
    pass


404
405
406
@dataclasses.dataclass
class SplitItem:
    submod_name: str
407
    graph_id: int
408
409
410
411
    is_splitting_graph: bool
    graph: fx.GraphModule


412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
def _is_empty_allocation_node(node: fx.Node) -> bool:
    if node.op == "call_method":
        return node.target == "new_empty"

    if node.op != "call_function":
        return False

    target = node.target
    if target in (torch.empty, torch.empty_like, torch.empty_strided):
        return True

    if isinstance(target, torch._ops.OpOverloadPacket):
        packet_name = target._qualified_op_name
    elif isinstance(target, torch._ops.OpOverload):
        packet_name = target.name()
    else:
        return False

    return packet_name.startswith("aten::empty") or packet_name.startswith(
        "aten::new_empty"
    )


def _merge_empty_only_subgraphs(
    node_to_subgraph_id: dict[fx.Node, int],
437
    split_op_graphs: list[int],
438
439
440
441
442
443
444
445
446
447
448
) -> None:
    """
    Merge a partition that only contains an empty allocation op into the
    previous partition. This avoids generating standalone empty submodules,
    which can lead to empty cudagraph captures.
    """

    nodes_by_subgraph_id: dict[int, list[fx.Node]] = defaultdict(list)
    for node, subgraph_id in node_to_subgraph_id.items():
        nodes_by_subgraph_id[subgraph_id].append(node)

449
450
451
452
453
454
455
    splitting_subgraphs = set(split_op_graphs)
    prev_non_splitting_subgraph_id: int | None = None

    max_subgraph_id = max(node_to_subgraph_id.values(), default=-1)
    for subgraph_id in range(max_subgraph_id + 1):
        nodes = nodes_by_subgraph_id.get(subgraph_id, [])
        if not nodes:
456
            continue
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474

        is_non_splitting_subgraph = subgraph_id not in splitting_subgraphs
        is_empty_only_subgraph = len(nodes) == 1 and _is_empty_allocation_node(nodes[0])
        merged = False

        if is_empty_only_subgraph and prev_non_splitting_subgraph_id is not None:
            # Safety check: don't move allocation before any input producer.
            empty_node = nodes[0]
            if all(
                input_node.op == "placeholder"
                or node_to_subgraph_id[input_node] <= prev_non_splitting_subgraph_id
                for input_node in empty_node.all_input_nodes
            ):
                node_to_subgraph_id[empty_node] = prev_non_splitting_subgraph_id
                merged = True

        if not merged and is_non_splitting_subgraph:
            prev_non_splitting_subgraph_id = subgraph_id
475
476


477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
def _decompose_size_nodes(graph: fx.GraphModule) -> None:
    """Decompose x.size() into per-dim sym_size.int calls.

    torch.Size objects cannot cross split boundaries because aot_autograd
    cannot handle them as submodule outputs. This replaces each size() call
    with individual sym_size.int(x, dim) nodes:
      - Dynamic dims (SymInt) → new sym_size.int node
      - Static dims (plain int) → inlined as literal constant
    """
    # Dynamo captures x.size()/x.shape as call_method target="size".
    size_nodes = list(graph.graph.find_nodes(op="call_method", target="size"))

    for node in size_nodes:
        tensor_node = node.args[0]
        ev = tensor_node.meta.get("example_value")
        assert ev is not None, (
            f"Tensor node '{tensor_node.name}' has no example_value metadata. "
            f"Cannot decompose size node '{node.name}'."
        )

        # Build per-dim replacements: sym_size.int node or literal int.
        dims: list[fx.Node | int] = []
        with graph.graph.inserting_after(tensor_node):
            for i in range(ev.dim()):
                dim_val = ev.shape[i]
                if isinstance(dim_val, torch.SymInt):
                    dn = graph.graph.call_function(
                        torch.ops.aten.sym_size.int, args=(tensor_node, i)
                    )
                    dn.meta["example_value"] = dim_val
                    dims.append(dn)
                elif isinstance(dim_val, int):
                    dims.append(dim_val)
                else:
                    raise AssertionError(
                        f"dim_val is either torch.SymInt or int, "
                        f"got {type(dim_val)} for dim {i} of "
                        f"'{node.name}'"
                    )

        # Replace size node in each user's args.
        # Dynamo always passes size as a direct arg: view(clone, size)
        # → view(clone, d0, d1, ...)
        for user in list(node.users):
            new_args = []
            for arg in user.args:
                if arg is node:
                    new_args.extend(dims)
                else:
                    new_args.append(arg)
            user.args = tuple(new_args)
        graph.graph.erase_node(node)


531
def split_graph(
532
    graph: fx.GraphModule, splitting_ops: list[str]
533
) -> tuple[fx.GraphModule, list[SplitItem]]:
534
535
    _decompose_size_nodes(graph)

536
537
    # split graph by ops
    subgraph_id = 0
538
539
    node_to_subgraph_id: dict[fx.Node, int] = {}
    split_op_graphs: list[int] = []
540
541
542
    for node in graph.graph.nodes:
        if node.op in ("output", "placeholder"):
            continue
543

544
545
546
547
548
549
550
551
552
553
554
555
        # Check if this is a getitem operation on a node from an earlier subgraph.
        # If so, assign it to the same subgraph as its input to avoid passing entire
        # tuple as input to submodules, which is against standalone_compile and
        # AoTAutograd input requirement.
        if node.op == "call_function" and node.target == operator.getitem:
            # Assign this getitem to the same subgraph as its input
            input_node = node.args[0]
            if input_node.op != "placeholder":
                assert input_node in node_to_subgraph_id
                node_to_subgraph_id[node] = node_to_subgraph_id[input_node]
                continue

556
        if should_split(node, splitting_ops):
557
558
559
            subgraph_id += 1
            node_to_subgraph_id[node] = subgraph_id
            split_op_graphs.append(subgraph_id)
560
561
562
563
564
565
566
567

            # keep consecutive splitting ops together
            # (we know node.next exists because node isn't the last (output) node)
            if should_split(node.next, splitting_ops):
                # this will get incremented by the next node
                subgraph_id -= 1
            else:
                subgraph_id += 1
568
569
570
        else:
            node_to_subgraph_id[node] = subgraph_id

571
    _merge_empty_only_subgraphs(node_to_subgraph_id, split_op_graphs)
572

573
574
575
576
    # `keep_original_order` is important!
    # otherwise pytorch might reorder the nodes and
    # the semantics of the graph will change when we
    # have mutations in the graph
577
578
579
580
581
582
583
    with _use_lazy_graph_module(True):
        split_gm = torch.fx.passes.split_module.split_module(
            graph,
            None,
            lambda node: node_to_subgraph_id[node],
            keep_original_order=True,
        )
584

585
    outputs = []
586

587
    names = [name for (name, module) in split_gm.named_modules()]
588

589
590
591
592
    for name in names:
        if "." in name or name == "":
            # recursive child module or the root module
            continue
593

594
        module = getattr(split_gm, name)
595

596
        graph_id = int(name.replace("submod_", ""))
597
        outputs.append(SplitItem(name, graph_id, (graph_id in split_op_graphs), module))
598

599
    # sort by integer graph_id, rather than string name
600
    outputs.sort(key=lambda x: x.graph_id)
601

602
    return split_gm, outputs
603
604


605
606
compilation_start_time = 0.0

607

608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
def wrap_with_cudagraph_if_needed(
    piecewise_backend: Any,
    vllm_config: VllmConfig,
    compilation_config: CompilationConfig,
    is_first_graph: bool,
    is_last_graph: bool,
) -> Any:
    """
    Wrap a piecewise backend with CUDA graph wrapper if needed.
    This function is shared between VllmBackend and
    construct_serializable_fn_from_inductor_cache.

    Args:
        piecewise_backend: The backend to wrap
        vllm_config: The vLLM configuration
        compilation_config: The compilation configuration
        is_first_graph: Whether this is the first graph in the sequence
        is_last_graph: Whether this is the last graph in the sequence

    Returns:
        The wrapped backend if CUDA graphs are enabled, otherwise the original backend
    """
    if (
        not compilation_config.cudagraph_mode.has_piecewise_cudagraphs()
        or compilation_config.use_inductor_graph_partition
    ):
        return piecewise_backend

    # We're using Dynamo-based piecewise splitting, so we wrap
    # the whole subgraph with a static graph wrapper.
    from .cuda_graph import CUDAGraphOptions

    # resolve the static graph wrapper class (e.g. CUDAGraphWrapper
    # class) as platform dependent.
    static_graph_wrapper_class = resolve_obj_by_qualname(
        current_platform.get_static_graph_wrapper_cls()
    )

    # Always assign PIECEWISE runtime mode to the
    # CUDAGraphWrapper for piecewise_backend, to distinguish
    # it from the FULL cudagraph runtime mode, no matter it
    # is wrapped on a full or piecewise fx graph.
    return static_graph_wrapper_class(
        runnable=piecewise_backend,
        vllm_config=vllm_config,
        runtime_mode=CUDAGraphMode.PIECEWISE,
        cudagraph_options=CUDAGraphOptions(
            debug_log_enable=is_first_graph,
            gc_disable=not is_first_graph,
            weak_ref_output=is_last_graph,
        ),
    )


662
class PiecewiseCompileInterpreter(torch.fx.Interpreter):  # type: ignore[misc]
663
    """Code adapted from `torch.fx.passes.shape_prop.ShapeProp`.
664
665
666
    It runs the given split graph interpreter, and for each submodule in
    `compile_submod_names`, creates a PiecewiseBackend and compiles all
    ranges up front.
667
668
669
670
671

    NOTE: the order in `compile_submod_names` matters, because
    it will be used to determine the order of the compiled piecewise
    graphs. The first graph will handle logging, and the last graph
    has some special cudagraph output handling.
672
673
674
675
676
677
678
679
680
681
682
683

    Note: This class shares similar logic with
    reconstruct_serializable_fn_from_mega_artifact in caching.py.
    Both create PiecewiseBackend instances and wrap them with cudagraph.
    The key difference is:
    - reconstruct_serializable_fn_from_mega_artifact: PiecewiseBackend receives
      pre-compiled runnables (compiled_runnables is set, graph is None)
    - this class: PiecewiseBackend receives the FX graph to compile
      (graph is set, compiled_runnables is None)


    If modifying the backend creation/wrapping logic, consider updating both.
684
685
    """

686
687
688
689
690
691
    def __init__(
        self,
        module: torch.fx.GraphModule,
        compile_submod_names: list[str],
        vllm_config: VllmConfig,
        vllm_backend: "VllmBackend",
692
    ) -> None:
693
694
        super().__init__(module)
        self.compile_submod_names = compile_submod_names
695
696
        self.compilation_config = vllm_config.compilation_config
        self.vllm_config = vllm_config
697
        self.vllm_backend = vllm_backend
698
699
        # When True, it annoyingly dumps the torch.fx.Graph on errors.
        self.extra_traceback = False
700

701
    @instrument(span_name="Inductor compilation")
702
    def run(self, *args: Any) -> Any:
703
        return super().run(*args)
704

705
706
707
708
709
710
    def call_module(
        self,
        target: torch.fx.node.Target,
        args: tuple[torch.fx.node.Argument, ...],
        kwargs: dict[str, Any],
    ) -> Any:
711
        assert isinstance(target, str)
712

713
714
715
        gm = getattr(self.module, target)
        outputs = gm.graph.output_node().args[0]
        output = fx.map_arg(outputs, lambda node: node.meta["example_value"])
716
717

        if target in self.compile_submod_names:
718
            index = self.compile_submod_names.index(target)
719
            submod = self.fetch_attr(target)
720

721
722
723
            sym_shape_indices = [
                i for i, x in enumerate(args) if isinstance(x, torch.SymInt)
            ]
724

725
            # Lazy import here to avoid circular import
726
727
            from torch._inductor.compile_fx import graph_returns_tuple

728
            from .piecewise_backend import PiecewiseBackend
729

730
            piecewise_backend = PiecewiseBackend(
731
732
733
734
735
736
                submod,
                self.vllm_config,
                index,
                len(self.compile_submod_names),
                sym_shape_indices,
                self.vllm_backend,
737
                graph_returns_tuple(submod),
738
                submod_name=target,
739
            )
740

741
742
743
744
745
746
747
            self.module.__dict__[target] = wrap_with_cudagraph_if_needed(
                piecewise_backend,
                self.vllm_config,
                self.compilation_config,
                piecewise_backend.is_first_graph,
                piecewise_backend.is_last_graph,
            )
748

749
750
751
752
753
            compilation_counter.num_piecewise_capturable_graphs_seen += 1

        return output


754
755
756
# the tag for the part of model being compiled,
# e.g. backbone/eagle_head
model_tag: str = "backbone"
757
model_is_encoder: bool = False
758
759
760


@contextmanager
761
def set_model_tag(tag: str, is_encoder: bool = False) -> Generator[None, None, None]:
762
763
    """Context manager to set the model tag."""
    global model_tag
764
    global model_is_encoder
765
    assert tag != model_tag, (
766
        f"Model tag {tag} is the same as the current tag {model_tag}."
767
    )
768
    old_tag = model_tag
769
770
    old_is_encoder = model_is_encoder

771
    model_tag = tag
772
    model_is_encoder = is_encoder
773
774
775
776
    try:
        yield
    finally:
        model_tag = old_tag
777
        model_is_encoder = old_is_encoder
778
779


780
class VllmBackend:
781
    """The compilation backend for `torch.compile` with vLLM.
782
    It is used for compilation mode of `CompilationMode.VLLM_COMPILE`,
783
    where we customize the compilation.
784

785
786
    The major work of this backend is to split the graph into
    piecewise graphs, and pass them to the piecewise backend.
787

788
789
    This backend also adds the PostGradPassManager to Inductor config,
    which handles the post-grad passes.
790
    """
791

792
793
    vllm_config: VllmConfig
    compilation_config: CompilationConfig
794
795
796
797
798
    _called: bool = False
    # the graph we compiled
    graph: fx.GraphModule
    # the stiching graph module for all the piecewise graphs
    split_gm: fx.GraphModule
799
    piecewise_graphs: list[SplitItem]
800
    returned_callable: Callable[..., Any]
801
    # Inductor passes to run on the graph pre-defunctionalization
802
    post_grad_passes: Sequence[Callable[..., Any]]
803
    compiler_manager: CompilerManager
804
805
806
    # Copy of CompilationConfig.inductor_compile_config +
    # an entry for PostGradPassManager
    inductor_config: dict[str, Any]
807

808
809
    def __init__(
        self,
810
        vllm_config: VllmConfig,
811
        prefix: str = "",
812
        is_encoder: bool = False,
813
    ) -> None:
814
815
        # if the model is initialized with a non-empty prefix,
        # then usually it's enough to use that prefix,
816
        # e.g. language_model, vision_model, etc.
817
818
819
820
821
        # when multiple parts are initialized as independent
        # models, we need to use the model_tag to distinguish
        # them, e.g. backbone (default), eagle_head, etc.
        self.prefix = prefix or model_tag

822
        # Mark compilation for encoder.
823
        self.is_encoder = is_encoder or model_is_encoder
824

825
        # Passes to run on the graph post-grad.
826
827
828
829
        self.pass_manager = resolve_obj_by_qualname(
            current_platform.get_pass_manager_cls()
        )()
        self.pass_key = current_platform.pass_key
830

831
832
        self.vllm_config = vllm_config
        self.compilation_config = vllm_config.compilation_config
833

834
        self.compiler_manager: CompilerManager = CompilerManager(
835
836
            self.compilation_config
        )
837

838
839
840
841
842
843
        # Deepcopy the inductor config to detach the post-grad custom pass
        # from CompilationConfig.
        # We want to avoid PostGradPassManager in CompilationConfig because
        # in future we need PostGradPassManager.uuid() to be executed
        # only at compile time.
        self.inductor_config = deepcopy(self.compilation_config.inductor_compile_config)
844
845
        # `torch.compile` is JIT compiled, so we don't need to
        # do anything here
846

847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
    def collect_standalone_compile_artifacts(
        self,
    ) -> tuple[Any, dict[str, list[int]] | None, dict[str, bool] | None]:
        """Collect inductor cache artifacts from all piecewise backends.

        Returns:
            tuple: (standalone_compile_artifacts, sym_shape_indices_map,
                    returns_tuple_map)
                - standalone_compile_artifacts: StandaloneCompiledArtifacts
                  with compiled artifacts
                - sym_shape_indices_map: dict mapping submod_name to
                  sym_shape_indices
                - returns_tuple_map: dict mapping submod_name to
                  returns_tuple
        """

        if not envs.VLLM_USE_MEGA_AOT_ARTIFACT:
            return None, None, None

        from .caching import StandaloneCompiledArtifacts
        from .piecewise_backend import PiecewiseBackend

        standalone_compile_artifacts = StandaloneCompiledArtifacts()
        sym_shape_indices_map = {}
        returns_tuple_map = {}

        for name, _ in self.split_gm.named_children():
            # get the actual attribute (shadowed by PiecewiseBackend in __dict__)
            child = getattr(self.split_gm, name)
            # unwrap the static graph wrapper class if applicable
            piecewise_backend = child.runnable if hasattr(child, "runnable") else child

            if not isinstance(piecewise_backend, PiecewiseBackend):
                continue

            submod_name = name
            sym_shape_indices_map[submod_name] = piecewise_backend.sym_shape_indices
            returns_tuple_map[submod_name] = piecewise_backend.returns_tuple

            for shape_str, bytes_data in piecewise_backend.to_bytes().items():
                standalone_compile_artifacts.insert(submod_name, shape_str, bytes_data)
                logger.debug(
                    "collected artifact for %s shape %s (%d bytes)",
                    submod_name,
                    shape_str,
                    len(bytes_data),
                )

        logger.info(
            "collected artifacts: %d entries, %d artifacts, %d bytes total",
            standalone_compile_artifacts.num_entries(),
            standalone_compile_artifacts.num_artifacts(),
            standalone_compile_artifacts.size_bytes(),
        )

        logger.debug(
            "standalone compile artifact keys: %s",
            list(standalone_compile_artifacts.submodule_bytes.keys()),
        )

        return standalone_compile_artifacts, sym_shape_indices_map, returns_tuple_map

909
    def configure_post_pass(self) -> None:
910
        self.pass_manager.configure(self.vllm_config)
911

912
913
        # Post-grad custom passes are run using the post_grad_custom_post_pass
        # hook. If a pass for that hook exists, add it to the pass manager.
914
915
916
917
918
        if self.pass_key in self.inductor_config:
            if isinstance(self.inductor_config[self.pass_key], PostGradPassManager):
                raise ValueError(
                    "PostGradPassManager can not be kept in CompilationConfig."
                )
919
            else:
920
                # Config should automatically wrap all inductor passes
921
922
923
924
925
926
927
                assert isinstance(
                    self.compilation_config.inductor_compile_config[self.pass_key],
                    InductorPass,
                )
                self.pass_manager.add(
                    self.compilation_config.inductor_compile_config[self.pass_key]
                )
928
        self.inductor_config[self.pass_key] = self.pass_manager
929

930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
    def _log_compilation_config(self):
        """Log vLLM compilation config for TORCH_TRACE/tlparse."""
        cc = self.compilation_config
        pass_cfg = cc.pass_config

        # Helper to convert lists to comma-separated strings for tlparse display
        def list_to_str(lst: list | None) -> str:
            if lst is None:
                return ""
            return ", ".join(str(x) for x in lst)

        # Get enabled passes by introspecting dataclass fields
        enabled_passes = [
            f.name
            for f in dataclasses.fields(pass_cfg)
            if isinstance(getattr(pass_cfg, f.name), bool) and getattr(pass_cfg, f.name)
        ]

        trace_structured(
            "artifact",
            metadata_fn=lambda: {
                "name": "vllm_compilation_config",
                "encoding": "json",
            },
            payload_fn=lambda: json.dumps(
                {
                    "model": self.vllm_config.model_config.model,
                    "prefix": self.prefix,
                    "mode": str(cc.mode),
                    "backend": cc.backend,
                    "custom_ops": list_to_str(cc.custom_ops),
                    "splitting_ops": list_to_str(cc.splitting_ops),
                    "cudagraph_mode": str(cc.cudagraph_mode),
                    "compile_sizes": list_to_str(cc.compile_sizes),
964
965
                    "compile_ranges_endpoints": list_to_str(
                        cc.compile_ranges_endpoints
966
967
968
969
970
971
972
973
974
975
                    ),
                    "use_inductor_graph_partition": cc.use_inductor_graph_partition,
                    "inductor_passes": list_to_str(list(cc.inductor_passes.keys())),
                    "enabled_passes": list_to_str(enabled_passes),
                    "dynamic_shapes_type": str(cc.dynamic_shapes_config.type),
                    "dynamic_shapes_evaluate_guards": cc.dynamic_shapes_config.evaluate_guards,  # noqa: E501
                }
            ),
        )

976
    @dynamo_timed("vllm_backend")
977
978
979
980
981
    def __call__(self, graph: fx.GraphModule, example_inputs: Sequence[Any]) -> Any:
        from .caching import (
            VllmSerializableFunction,
        )

982
        vllm_config = self.vllm_config
983
984
985

        self._log_compilation_config()

986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
        # Minimal hashing here with existing utilities, reused below.

        env_factors = envs.compile_factors()
        env_hash = hash_factors(env_factors)
        # Compute config/compiler/code hashes once and reuse
        config_hash = vllm_config.compute_hash()
        compiler_hash = self.compiler_manager.compute_hash(vllm_config)
        forward_code_files = list(sorted(self.compilation_config.traced_files))

        logger.debug(
            "Traced files (to be considered for compilation cache):\n%s",
            lazy(lambda: "\n".join(forward_code_files)),
        )
        hash_content = []
        for filepath in forward_code_files:
            hash_content.append(filepath)
            if filepath == "<string>":
                # This means the function was dynamically generated, with
                # e.g. exec(). We can't actually check these.
                continue
            try:
                with open(filepath) as f:
                    hash_content.append(f.read())
1009
            except (OSError, UnicodeDecodeError):
1010
1011
1012
1013
1014
                logger.warning("Failed to read file %s", filepath)
                continue
        code_hash = hashlib.sha256("\n".join(hash_content).encode()).hexdigest()
        # Clear after consumption
        self.compilation_config.traced_files.clear()
1015
1016
1017
1018
1019
        if not self.compilation_config.cache_dir:
            # no provided cache dir, generate one based on the known factors
            # that affects the compilation. if none of the factors change,
            # the cache dir will be the same so that we can reuse the compiled
            # graph.
1020
1021
1022
1023
            factors = [env_hash, config_hash, code_hash, compiler_hash]
            # Use SHA-256 for cache key hashing to be consistent across
            # compute_hash functions. Truncate for a short cache dir name.
            hash_key = hashlib.sha256(str(factors).encode()).hexdigest()[:10]
1024
            cache_dir = os.path.join(
1025
                envs.VLLM_CACHE_ROOT, "torch_compile_cache", hash_key
1026
1027
1028
            )
            self.compilation_config.cache_dir = cache_dir

1029
        cache_dir = self.compilation_config.cache_dir
1030
        os.makedirs(cache_dir, exist_ok=True)
1031
        self.compilation_config.cache_dir = cache_dir
1032
        rank = vllm_config.parallel_config.rank
1033
        dp_rank = vllm_config.parallel_config.data_parallel_index
1034
        local_cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}", self.prefix)
1035
        os.makedirs(local_cache_dir, exist_ok=True)
1036
        self.compilation_config.local_cache_dir = local_cache_dir
1037

1038
        # Honors opt-outs such as CompilationMode.NONE or VLLM_DISABLE_COMPILE_CACHE.
1039
        disable_cache = not is_compile_cache_enabled(self.inductor_config)
1040

1041
1042
1043
1044
1045
1046
1047
        # TODO(patchy): ngram gpu kernel will cause vllm torch compile cache errors.
        is_ngram_gpu_enabled = (
            vllm_config.speculative_config is not None
            and vllm_config.speculative_config.use_ngram_gpu()
        )
        disable_cache = disable_cache or is_ngram_gpu_enabled

1048
        if disable_cache:
1049
            logger.info_once("vLLM's torch.compile cache is disabled.", scope="local")
1050
        else:
1051
1052
1053
1054
            logger.info_once(
                "Using cache directory: %s for vLLM's torch.compile",
                local_cache_dir,
                scope="local",
1055
            )
1056

1057
1058
1059
        self.compiler_manager.initialize_cache(
            local_cache_dir, disable_cache, self.prefix
        )
1060

1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
        # Reuses existing cache key

        logger.debug(
            "torch.compile cache factors: env=%s cfg=%s comp=%s code=%s dir=%s",
            env_hash,
            config_hash,
            compiler_hash,
            code_hash,
            local_cache_dir,
        )

        # Persist and log only hash-relevant factors together.
        try:
            logger.debug(
                "Compile env factors (raw):\n%s\nVllm config hash: %s",
                lazy(partial(pprint.pformat, env_factors, width=120)),
                config_hash,
            )
            meta_path = os.path.join(local_cache_dir, "cache_key_factors.json")
            if not os.path.exists(meta_path):
                with open(meta_path, "w") as f:
                    json.dump(
                        {
                            "env": env_factors,  # raw factors used for env_hash
                            "config_hash": config_hash,
                            "code_hash": code_hash,
                            "compiler_hash": compiler_hash,
                        },
                        f,
                        indent=2,
                        sort_keys=True,
                    )
        except Exception:
            # Best-effort only; metadata write failures are non-fatal.
            logger.warning(
                (
                    "Could not write compile cache metadata at %s; continuing without "
                    "metadata. Compiled cache remains valid; diagnostics may be "
                    "limited."
                ),
                local_cache_dir,
                exc_info=True,
            )

1105
1106
        # when dynamo calls the backend, it means the bytecode
        # transform and analysis are done
1107
        compilation_counter.num_graphs_seen += 1
1108
        from .monitor import torch_compile_start_time
1109

1110
        dynamo_time = time.perf_counter() - torch_compile_start_time
1111
1112
1113
        logger.info_once(
            "Dynamo bytecode transform time: %.2f s", dynamo_time, scope="local"
        )
1114
        self.compilation_config.compilation_time += dynamo_time
1115

1116
1117
1118
1119
1120
        # Record Dynamo time in tracing if available
        start_time = int(torch_compile_start_time * 1e9)
        attributes = {"dynamo.time_seconds": dynamo_time}
        instrument_manual("Dynamo bytecode transform", start_time, None, attributes)

1121
1122
1123
1124
1125
        # we control the compilation process, each instance can only be
        # called once
        assert not self._called, "VllmBackend can only be called once"

        self.graph = graph
1126
        self.configure_post_pass()
1127

1128
1129
1130
1131
1132
1133
        if self.compilation_config.use_inductor_graph_partition:
            # Let Inductor decide partitioning; avoid FX-level pre-splitting.
            fx_split_ops: list[str] = []
        else:
            fx_split_ops = self.compilation_config.splitting_ops or []

1134
        self.split_gm, self.piecewise_graphs = split_graph(graph, fx_split_ops)
1135

1136
1137
1138
1139
1140
1141
        # keep a split_gm copy from BEFORE the interpreter replaces
        # submodules with PiecewiseBackend -- used for serialization
        original_split_gm = None
        if envs.VLLM_USE_MEGA_AOT_ARTIFACT:
            original_split_gm = deepcopy(self.split_gm)

1142
        from torch._dynamo.utils import lazy_format_graph_code
1143
1144
1145
1146
1147

        # depyf will hook lazy_format_graph_code and dump the graph
        # for debugging, no need to print the graph here
        lazy_format_graph_code("before split", self.graph)
        lazy_format_graph_code("after split", self.split_gm)
1148

1149
1150
1151
1152
1153
1154
1155
        # Log the piecewise split graph for TORCH_TRACE/tlparse
        trace_structured(
            "graph_dump",
            metadata_fn=lambda: {"name": "vllm_piecewise_split_graph"},
            payload_fn=lambda: self.split_gm.print_readable(print_output=False),
        )

1156
        compilation_counter.num_piecewise_graphs_seen += len(self.piecewise_graphs)
1157
        submod_names_to_compile = [
1158
1159
            item.submod_name
            for item in self.piecewise_graphs
1160
1161
1162
            if not item.is_splitting_graph
        ]

1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
        # Extract fake values from the graph to use them when needed.
        all_fake_values = []
        for i in graph.graph.find_nodes(op="placeholder"):
            all_fake_values.append(i.meta["example_value"])

        fake_args = [
            all_fake_values[i] if isinstance(t, torch.Tensor) else t
            for i, t in enumerate(example_inputs)
        ]

1173
        # propagate the split graph to the piecewise backend,
1174
1175
1176
        # compile submodules with symbolic shapes, and compile all ranges
        # up front so that compilation is complete before the callable
        # is returned.
1177
1178
        PiecewiseCompileInterpreter(
            self.split_gm, submod_names_to_compile, self.vllm_config, self
1179
        ).run(*fake_args)
1180

1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
        # All compilation is done. Save the cache.
        time_before_saving = time.perf_counter()
        self.compiler_manager.save_to_file()
        elapsed = time.perf_counter() - time_before_saving
        if elapsed > 1:
            logger.info_once(
                "Saved compiler manager cache in %.2f seconds.",
                elapsed,
                scope="local",
            )

1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
        from torch._guards import detect_fake_mode

        fake_mode = detect_fake_mode()

        if (
            self.compilation_config.dynamic_shapes_config.evaluate_guards
            and self.compilation_config.dynamic_shapes_config.type
            == DynamicShapesType.BACKED
        ):
            from torch.utils._sympy.value_ranges import ValueRanges

            # Drop counter-0/1 specializations guards; for backed dynamic shapes,
            # torch.compile will specialize for 0/1 inputs or otherwise guards that
            # shape is >= 2. This is because it's really hard not to hit a check
            # against 0/1. When we evaluate shape guards, we exclude checking those
            # guards (We would fail always otherwise).

            # We avoid that by updating the ranges of backed sizes when the min is
            # 2 for any, we assume it's 0.
            for s, r in fake_mode.shape_env.var_to_range.items():
                if r.lower == 2:
                    fake_mode.shape_env.var_to_range[s] = ValueRanges(0, r.upper)

1215
1216
        graph_path = os.path.join(local_cache_dir, "computation_graph.py")
        if not os.path.exists(graph_path):
1217
1218
            # code adapted from
            # https://github.com/thuml/depyf/blob/dab831108a752d1facc00acdd6d4243891845c37/depyf/explain/patched_lazy_format_graph_code.py#L30
1219
            # use `print_readable` because it can include submodules
1220
1221
1222
1223
            src = (
                "from __future__ import annotations\nimport torch\n"
                + self.split_gm.print_readable(print_output=False)
            )
1224
1225
1226
1227
            src = src.replace("<lambda>", "GraphModule")
            with open(graph_path, "w") as f:
                f.write(src)

1228
1229
1230
            logger.debug_once(
                "Computation graph saved to %s", graph_path, scope="local"
            )
1231

1232
        self._called = True
1233
1234
1235
        graph_to_serialize = (
            original_split_gm if envs.VLLM_USE_MEGA_AOT_ARTIFACT else self.graph
        )
1236

1237
1238
1239
1240
        if (
            self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE
            or not self.compilation_config.cudagraph_copy_inputs
        ):
1241
            return VllmSerializableFunction(
1242
1243
1244
1245
1246
1247
                graph_to_serialize,
                example_inputs,
                self.prefix,
                self.split_gm,
                is_encoder=self.is_encoder,
                vllm_backend=self,
1248
            )
1249
1250

        # index of tensors that have symbolic shapes (batch size)
1251
1252
1253
        # for weights and static buffers, they will have concrete shapes.
        # symbolic shape only happens for input tensors.
        from torch.fx.experimental.symbolic_shapes import is_symbolic
1254

1255
        sym_tensor_indices = [
1256
1257
1258
1259
            i
            for i, x in enumerate(fake_args)
            if isinstance(x, torch._subclasses.fake_tensor.FakeTensor)
            and any(is_symbolic(d) for d in x.size())
1260
1261
1262
1263
1264
        ]

        # compiler managed cudagraph input buffers
        # we assume the first run with symbolic shapes
        # has the maximum size among all the tensors
1265
1266
1267
1268
1269
        copy_and_call = make_copy_and_call(
            sym_tensor_indices,
            [example_inputs[x].clone() for x in sym_tensor_indices],
            self.split_gm,
        )
1270

1271
        return VllmSerializableFunction(
1272
1273
1274
1275
1276
1277
1278
            graph_to_serialize,
            example_inputs,
            self.prefix,
            copy_and_call,
            is_encoder=self.is_encoder,
            vllm_backend=self,
            sym_tensor_indices=sym_tensor_indices,
1279
        )