Skip to content

vllm.compilation.backends

compilation_start_time module-attribute 

compilation_start_time = 0.0

global_graph_pool module-attribute 

global_graph_pool = None

logger module-attribute 

logger = init_logger(__name__)

CompilerManager

A manager to manage the compilation process, including caching the compiled graph, loading the compiled graph, and compiling the graph.

The cache is a dict mapping (runtime_shape, graph_index, backend_name) to any_data returned from the compiler.

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.

Source code in vllm/compilation/backends.py 
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
class CompilerManager:
    """
    A manager to manage the compilation process, including
    caching the compiled graph, loading the compiled graph,
    and compiling the graph.

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

    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.
    """

    def __init__(self, compilation_config: CompilationConfig):
        self.cache: dict[tuple[Optional[int], int, str], Any] = dict()
        self.is_cache_updated = False
        self.compilation_config = compilation_config
        self.compiler = make_compiler(compilation_config)

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

    def initialize_cache(self, cache_dir: str, disable_cache: bool = False):
        self.disable_cache = disable_cache
        self.cache_dir = cache_dir
        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
            with open(self.cache_file_path) as f:
                # 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.
                self.cache = ast.literal_eval(f.read())

        self.compiler.initialize_cache(cache_dir=cache_dir,
                                       disable_cache=disable_cache)

    def save_to_file(self):
        if self.disable_cache or not self.is_cache_updated:
            return
        printer = pprint.PrettyPrinter(indent=4)
        data = printer.pformat(self.cache)
        with open(self.cache_file_path, "w") as f:
            f.write(data)

    def load(self,
             graph: fx.GraphModule,
             example_inputs: list[Any],
             graph_index: int,
             runtime_shape: Optional[int] = None) -> Optional[Callable]:
        if (runtime_shape, graph_index, self.compiler.name) not in self.cache:
            return None
        handle = self.cache[(runtime_shape, graph_index, self.compiler.name)]
        compiled_graph = self.compiler.load(handle, graph, example_inputs,
                                            graph_index, runtime_shape)
        logger.debug(
            "Directly load the %s-th graph for shape %s from %s via "
            "handle %s", graph_index, str(runtime_shape), self.compiler.name,
            handle)
        return compiled_graph

    def compile(self,
                graph: fx.GraphModule,
                example_inputs,
                additional_inductor_config,
                compilation_config: CompilationConfig,
                graph_index: int = 0,
                num_graphs: int = 1,
                runtime_shape: Optional[int] = None) -> Any:
        if graph_index == 0:
            # before compiling the first graph, record the start time
            global compilation_start_time
            compilation_start_time = time.time()

        compilation_counter.num_backend_compilations += 1

        compiled_graph = None

        # try to load from the cache
        compiled_graph = self.load(graph, example_inputs, graph_index,
                                   runtime_shape)
        if compiled_graph is not None:
            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.
                now = time.time()
                elapsed = now - compilation_start_time
                logger.info(
                    "Directly load the compiled graph(s) for shape %s "
                    "from the cache, took %.3f s", str(runtime_shape), elapsed)
            return compiled_graph

        # no compiler cached the graph, or the cache is disabled,
        # we need to compile it
        if isinstance(self.compiler, InductorAdaptor):
            # Let compile_fx generate a key for us
            maybe_key = None
        else:
            maybe_key = \
                f"artifact_shape_{runtime_shape}_subgraph_{graph_index}"
        compiled_graph, handle = self.compiler.compile(
            graph, example_inputs, additional_inductor_config, runtime_shape,
            maybe_key)

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

        # store the artifact in the cache
        if handle is not None:
            self.cache[(runtime_shape, graph_index,
                        self.compiler.name)] = handle
            self.is_cache_updated = True
            if graph_index == 0:
                # adds some info logging for the first graph
                logger.info("Cache the graph of shape %s for later use",
                            str(runtime_shape))
            logger.debug(
                "store the %s-th graph for shape %s from %s via handle %s",
                graph_index, str(runtime_shape), self.compiler.name, handle)

        # after compiling the last graph, record the end time
        if graph_index == num_graphs - 1:
            now = time.time()
            elapsed = now - compilation_start_time
            compilation_config.compilation_time += elapsed
            if runtime_shape is None:
                logger.info("Compiling a graph for general shape takes %.2f s",
                            elapsed)
            else:
                logger.info("Compiling a graph for shape %s takes %.2f s",
                            runtime_shape, elapsed)

        return compiled_graph

cache instance-attribute 

cache: dict[tuple[Optional[int], int, str], Any] = dict()

compilation_config instance-attribute 

compilation_config = compilation_config

compiler instance-attribute 

compiler = make_compiler(compilation_config)

is_cache_updated instance-attribute 

is_cache_updated = False

__init__ 

__init__(compilation_config: CompilationConfig)
Source code in vllm/compilation/backends.py 
57
58
59
60
61
def __init__(self, compilation_config: CompilationConfig):
    self.cache: dict[tuple[Optional[int], int, str], Any] = dict()
    self.is_cache_updated = False
    self.compilation_config = compilation_config
    self.compiler = make_compiler(compilation_config)

compile 

compile(
    graph: GraphModule,
    example_inputs,
    additional_inductor_config,
    compilation_config: CompilationConfig,
    graph_index: int = 0,
    num_graphs: int = 1,
    runtime_shape: Optional[int] = None,
) -> Any
Source code in vllm/compilation/backends.py 
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
def compile(self,
            graph: fx.GraphModule,
            example_inputs,
            additional_inductor_config,
            compilation_config: CompilationConfig,
            graph_index: int = 0,
            num_graphs: int = 1,
            runtime_shape: Optional[int] = None) -> Any:
    if graph_index == 0:
        # before compiling the first graph, record the start time
        global compilation_start_time
        compilation_start_time = time.time()

    compilation_counter.num_backend_compilations += 1

    compiled_graph = None

    # try to load from the cache
    compiled_graph = self.load(graph, example_inputs, graph_index,
                               runtime_shape)
    if compiled_graph is not None:
        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.
            now = time.time()
            elapsed = now - compilation_start_time
            logger.info(
                "Directly load the compiled graph(s) for shape %s "
                "from the cache, took %.3f s", str(runtime_shape), elapsed)
        return compiled_graph

    # no compiler cached the graph, or the cache is disabled,
    # we need to compile it
    if isinstance(self.compiler, InductorAdaptor):
        # Let compile_fx generate a key for us
        maybe_key = None
    else:
        maybe_key = \
            f"artifact_shape_{runtime_shape}_subgraph_{graph_index}"
    compiled_graph, handle = self.compiler.compile(
        graph, example_inputs, additional_inductor_config, runtime_shape,
        maybe_key)

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

    # store the artifact in the cache
    if handle is not None:
        self.cache[(runtime_shape, graph_index,
                    self.compiler.name)] = handle
        self.is_cache_updated = True
        if graph_index == 0:
            # adds some info logging for the first graph
            logger.info("Cache the graph of shape %s for later use",
                        str(runtime_shape))
        logger.debug(
            "store the %s-th graph for shape %s from %s via handle %s",
            graph_index, str(runtime_shape), self.compiler.name, handle)

    # after compiling the last graph, record the end time
    if graph_index == num_graphs - 1:
        now = time.time()
        elapsed = now - compilation_start_time
        compilation_config.compilation_time += elapsed
        if runtime_shape is None:
            logger.info("Compiling a graph for general shape takes %.2f s",
                        elapsed)
        else:
            logger.info("Compiling a graph for shape %s takes %.2f s",
                        runtime_shape, elapsed)

    return compiled_graph

compute_hash 

compute_hash(vllm_config: VllmConfig) -> str
Source code in vllm/compilation/backends.py 
63
64
def compute_hash(self, vllm_config: VllmConfig) -> str:
    return self.compiler.compute_hash(vllm_config)

initialize_cache 

initialize_cache(
    cache_dir: str, disable_cache: bool = False
)
Source code in vllm/compilation/backends.py 
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
def initialize_cache(self, cache_dir: str, disable_cache: bool = False):
    self.disable_cache = disable_cache
    self.cache_dir = cache_dir
    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
        with open(self.cache_file_path) as f:
            # 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.
            self.cache = ast.literal_eval(f.read())

    self.compiler.initialize_cache(cache_dir=cache_dir,
                                   disable_cache=disable_cache)

load 

load(
    graph: GraphModule,
    example_inputs: list[Any],
    graph_index: int,
    runtime_shape: Optional[int] = None,
) -> Optional[Callable]
Source code in vllm/compilation/backends.py 
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
def load(self,
         graph: fx.GraphModule,
         example_inputs: list[Any],
         graph_index: int,
         runtime_shape: Optional[int] = None) -> Optional[Callable]:
    if (runtime_shape, graph_index, self.compiler.name) not in self.cache:
        return None
    handle = self.cache[(runtime_shape, graph_index, self.compiler.name)]
    compiled_graph = self.compiler.load(handle, graph, example_inputs,
                                        graph_index, runtime_shape)
    logger.debug(
        "Directly load the %s-th graph for shape %s from %s via "
        "handle %s", graph_index, str(runtime_shape), self.compiler.name,
        handle)
    return compiled_graph

save_to_file 

save_to_file()
Source code in vllm/compilation/backends.py 
82
83
84
85
86
87
88
def save_to_file(self):
    if self.disable_cache or not self.is_cache_updated:
        return
    printer = pprint.PrettyPrinter(indent=4)
    data = printer.pformat(self.cache)
    with open(self.cache_file_path, "w") as f:
        f.write(data)

PiecewiseCompileInterpreter

Bases: Interpreter

Code adapted from torch.fx.passes.shape_prop.ShapeProp. It runs the given graph with fake inputs, and compile some submodules specified by compile_submod_names with the given compilation configs.

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.

Source code in vllm/compilation/backends.py 
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
class PiecewiseCompileInterpreter(torch.fx.Interpreter):
    """Code adapted from `torch.fx.passes.shape_prop.ShapeProp`.
    It runs the given graph with fake inputs, and compile some
    submodules specified by `compile_submod_names` with the given
    compilation configs.

    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.
    """

    def __init__(self, module: torch.fx.GraphModule,
                 compile_submod_names: list[str], vllm_config: VllmConfig,
                 graph_pool, vllm_backend: "VllmBackend"):
        super().__init__(module)
        from torch._guards import detect_fake_mode
        self.fake_mode = detect_fake_mode()
        self.compile_submod_names = compile_submod_names
        self.compilation_config = vllm_config.compilation_config
        self.graph_pool = graph_pool
        self.vllm_config = vllm_config
        self.vllm_backend = vllm_backend
        # When True, it annoyingly dumps the torch.fx.Graph on errors.
        self.extra_traceback = False

    def run(self, *args):
        fake_args = [
            self.fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
            for t in args
        ]
        with self.fake_mode:
            return super().run(*fake_args)

    def call_module(self, target: torch.fx.node.Target,
                    args: tuple[torch.fx.node.Argument,
                                ...], kwargs: dict[str, Any]) -> Any:
        assert isinstance(target, str)
        output = super().call_module(target, args, kwargs)

        if target in self.compile_submod_names:
            index = self.compile_submod_names.index(target)
            submod = self.fetch_attr(target)
            sym_shape_indices = [
                i for i, x in enumerate(args) if isinstance(x, torch.SymInt)
            ]
            global compilation_start_time
            compiled_graph_for_general_shape = self.vllm_backend.\
                compiler_manager.compile(
                submod,
                args,
                self.compilation_config.inductor_compile_config,
                self.compilation_config,
                graph_index=index,
                num_graphs=len(self.compile_submod_names),
                runtime_shape=None)

            piecewise_backend = resolve_obj_by_qualname(
                current_platform.get_piecewise_backend_cls())
            self.module.__dict__[target] = piecewise_backend(
                submod, self.vllm_config, self.graph_pool, index,
                len(self.compile_submod_names), sym_shape_indices,
                compiled_graph_for_general_shape, self.vllm_backend)

            compilation_counter.num_piecewise_capturable_graphs_seen += 1

        return output

compilation_config instance-attribute 

compilation_config = compilation_config

compile_submod_names instance-attribute 

compile_submod_names = compile_submod_names

extra_traceback instance-attribute 

extra_traceback = False

fake_mode instance-attribute 

fake_mode = detect_fake_mode()

graph_pool instance-attribute 

graph_pool = graph_pool

vllm_backend instance-attribute 

vllm_backend = vllm_backend

vllm_config instance-attribute 

vllm_config = vllm_config

__init__ 

__init__(
    module: GraphModule,
    compile_submod_names: list[str],
    vllm_config: VllmConfig,
    graph_pool,
    vllm_backend: VllmBackend,
)
Source code in vllm/compilation/backends.py 
253
254
255
256
257
258
259
260
261
262
263
264
265
def __init__(self, module: torch.fx.GraphModule,
             compile_submod_names: list[str], vllm_config: VllmConfig,
             graph_pool, vllm_backend: "VllmBackend"):
    super().__init__(module)
    from torch._guards import detect_fake_mode
    self.fake_mode = detect_fake_mode()
    self.compile_submod_names = compile_submod_names
    self.compilation_config = vllm_config.compilation_config
    self.graph_pool = graph_pool
    self.vllm_config = vllm_config
    self.vllm_backend = vllm_backend
    # When True, it annoyingly dumps the torch.fx.Graph on errors.
    self.extra_traceback = False

call_module 

call_module(
    target: Target,
    args: tuple[Argument, ...],
    kwargs: dict[str, Any],
) -> Any
Source code in vllm/compilation/backends.py 
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
def call_module(self, target: torch.fx.node.Target,
                args: tuple[torch.fx.node.Argument,
                            ...], kwargs: dict[str, Any]) -> Any:
    assert isinstance(target, str)
    output = super().call_module(target, args, kwargs)

    if target in self.compile_submod_names:
        index = self.compile_submod_names.index(target)
        submod = self.fetch_attr(target)
        sym_shape_indices = [
            i for i, x in enumerate(args) if isinstance(x, torch.SymInt)
        ]
        global compilation_start_time
        compiled_graph_for_general_shape = self.vllm_backend.\
            compiler_manager.compile(
            submod,
            args,
            self.compilation_config.inductor_compile_config,
            self.compilation_config,
            graph_index=index,
            num_graphs=len(self.compile_submod_names),
            runtime_shape=None)

        piecewise_backend = resolve_obj_by_qualname(
            current_platform.get_piecewise_backend_cls())
        self.module.__dict__[target] = piecewise_backend(
            submod, self.vllm_config, self.graph_pool, index,
            len(self.compile_submod_names), sym_shape_indices,
            compiled_graph_for_general_shape, self.vllm_backend)

        compilation_counter.num_piecewise_capturable_graphs_seen += 1

    return output

run 

run(*args)
Source code in vllm/compilation/backends.py 
267
268
269
270
271
272
273
def run(self, *args):
    fake_args = [
        self.fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
        for t in args
    ]
    with self.fake_mode:
        return super().run(*fake_args)

SplitItem dataclass

Source code in vllm/compilation/backends.py 
179
180
181
182
183
184
@dataclasses.dataclass
class SplitItem:
    submod_name: str
    graph_id: int
    is_splitting_graph: bool
    graph: fx.GraphModule

graph instance-attribute 

graph: GraphModule

graph_id instance-attribute 

graph_id: int

is_splitting_graph instance-attribute 

is_splitting_graph: bool

submod_name instance-attribute 

submod_name: str

__init__ 

__init__(
    submod_name: str,
    graph_id: int,
    is_splitting_graph: bool,
    graph: GraphModule,
) -> None

VllmBackend

The compilation backend for torch.compile with vLLM. It is used for compilation level of CompilationLevel.PIECEWISE, where we customize the compilation.

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

This backend also adds the PostGradPassManager to Inductor config, which handles the post-grad passes.

Source code in vllm/compilation/backends.py 
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
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
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
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
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
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
class VllmBackend:
    """The compilation backend for `torch.compile` with vLLM.
    It is used for compilation level of `CompilationLevel.PIECEWISE`,
    where we customize the compilation.

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

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

    vllm_config: VllmConfig
    compilation_config: CompilationConfig
    graph_pool: Any
    _called: bool = False
    # the graph we compiled
    graph: fx.GraphModule
    # the stiching graph module for all the piecewise graphs
    split_gm: fx.GraphModule
    piecewise_graphs: list[SplitItem]
    returned_callable: Callable
    # Inductor passes to run on the graph pre-defunctionalization
    post_grad_passes: Sequence[Callable]
    sym_tensor_indices: list[int]
    input_buffers: list[torch.Tensor]
    compiler_manager: CompilerManager

    def __init__(
        self,
        vllm_config: VllmConfig,
    ):
        global global_graph_pool
        if global_graph_pool is None:
            global_graph_pool = current_platform.graph_pool_handle()

        # TODO: in the future, if we want to use multiple
        # streams, it might not be safe to share a global pool.
        # only investigate this when we use multiple streams
        self.graph_pool = global_graph_pool

        # Passes to run on the graph post-grad.
        self.post_grad_pass_manager = PostGradPassManager()

        self.sym_tensor_indices = []
        self.input_buffers = []

        self.vllm_config = vllm_config
        self.compilation_config = vllm_config.compilation_config

        self.compiler_manager: CompilerManager = CompilerManager(
            self.compilation_config)

        # `torch.compile` is JIT compiled, so we don't need to
        # do anything here

    def configure_post_pass(self):
        config = self.compilation_config
        self.post_grad_pass_manager.configure(self.vllm_config)

        # 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.
        inductor_config = config.inductor_compile_config
        PASS_KEY = "post_grad_custom_post_pass"
        if PASS_KEY in inductor_config:
            # Config should automatically wrap all inductor passes
            if isinstance(inductor_config[PASS_KEY], PostGradPassManager):
                assert (inductor_config[PASS_KEY].uuid() ==
                        self.post_grad_pass_manager.uuid())
            else:
                assert isinstance(inductor_config[PASS_KEY], InductorPass)
                self.post_grad_pass_manager.add(inductor_config[PASS_KEY])
        inductor_config[PASS_KEY] = self.post_grad_pass_manager

    def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:

        vllm_config = self.vllm_config
        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.

            factors = []
            # 0. factors come from the env, for example, The values of
            # VLLM_PP_LAYER_PARTITION will affects the computation graph.
            env_hash = envs.compute_hash()
            factors.append(env_hash)

            # 1. factors come from the vllm_config (it mainly summarizes how the
            #    model is created)
            config_hash = vllm_config.compute_hash()
            factors.append(config_hash)

            # 2. factors come from the code files that are traced by Dynamo (
            #    it mainly summarizes how the model is used in forward pass)
            forward_code_files = list(
                sorted(self.compilation_config.traced_files))
            self.compilation_config.traced_files.clear()
            logger.debug(
                "Traced files (to be considered for compilation cache):\n%s",
                "\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
                with open(filepath) as f:
                    hash_content.append(f.read())
            import hashlib
            code_hash = hashlib.md5("\n".join(hash_content).encode(),
                                    usedforsecurity=False).hexdigest()
            factors.append(code_hash)

            # 3. compiler hash
            compiler_hash = self.compiler_manager.compute_hash(vllm_config)
            factors.append(compiler_hash)

            # combine all factors to generate the cache dir
            hash_key = hashlib.md5(str(factors).encode(),
                                   usedforsecurity=False).hexdigest()[:10]

            cache_dir = os.path.join(
                envs.VLLM_CACHE_ROOT,
                "torch_compile_cache",
                hash_key,
            )
            self.compilation_config.cache_dir = cache_dir

        if compilation_counter.num_graphs_seen > 0:
            cache_dir = self.compilation_config.cache_dir + \
                f'-{compilation_counter.num_graphs_seen}'
        else:
            cache_dir = self.compilation_config.cache_dir
        os.makedirs(cache_dir, exist_ok=True)
        self.compilation_config.cache_dir = cache_dir
        rank = vllm_config.parallel_config.rank
        dp_rank = vllm_config.parallel_config.data_parallel_rank
        local_cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}")
        os.makedirs(local_cache_dir, exist_ok=True)
        self.compilation_config.local_cache_dir = local_cache_dir

        disable_cache = envs.VLLM_DISABLE_COMPILE_CACHE

        if disable_cache:
            logger.info("vLLM's torch.compile cache is disabled.")
        else:
            logger.info("Using cache directory: %s for vLLM's torch.compile",
                        local_cache_dir)

        self.compiler_manager.initialize_cache(local_cache_dir, disable_cache)

        # when dynamo calls the backend, it means the bytecode
        # transform and analysis are done
        compilation_counter.num_graphs_seen += 1
        from .monitor import torch_compile_start_time
        dynamo_time = time.time() - torch_compile_start_time
        logger.info("Dynamo bytecode transform time: %.2f s", dynamo_time)
        self.compilation_config.compilation_time += dynamo_time

        # 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
        self.configure_post_pass()

        self.split_gm, self.piecewise_graphs = split_graph(
            graph, self.compilation_config.splitting_ops)

        from torch._dynamo.utils import lazy_format_graph_code

        # 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)

        compilation_counter.num_piecewise_graphs_seen += len(
            self.piecewise_graphs)
        submod_names_to_compile = [
            item.submod_name for item in self.piecewise_graphs
            if not item.is_splitting_graph
        ]

        # propagate the split graph to the piecewise backend,
        # compile submodules with symbolic shapes
        PiecewiseCompileInterpreter(self.split_gm, submod_names_to_compile,
                                    self.vllm_config, self.graph_pool,
                                    self).run(*example_inputs)

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

            logger.debug("Computation graph saved to %s", graph_path)

        self._called = True

        if not self.compilation_config.use_cudagraph or \
            not self.compilation_config.cudagraph_copy_inputs:
            return self.split_gm

        # if we need to copy input buffers for cudagraph
        from torch._guards import detect_fake_mode
        fake_mode = detect_fake_mode()
        fake_args = [
            fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
            for t in example_inputs
        ]

        # index of tensors that have symbolic shapes (batch size)
        # 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
        self.sym_tensor_indices = [
            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())
        ]

        # compiler managed cudagraph input buffers
        # we assume the first run with symbolic shapes
        # has the maximum size among all the tensors
        self.input_buffers = [
            example_inputs[x].clone() for x in self.sym_tensor_indices
        ]

        # this is the callable we return to Dynamo to run
        def copy_and_call(*args):
            list_args = list(args)
            for i, index in enumerate(self.sym_tensor_indices):
                runtime_tensor = list_args[index]
                runtime_shape = runtime_tensor.shape[0]
                static_tensor = self.input_buffers[i][:runtime_shape]

                # copy the tensor to the static buffer
                static_tensor.copy_(runtime_tensor)

                # replace the tensor in the list_args to the static buffer
                list_args[index] = static_tensor
            return self.split_gm(*list_args)

        return copy_and_call

_called class-attribute instance-attribute 

_called: bool = False

compilation_config instance-attribute 

compilation_config: CompilationConfig = compilation_config

compiler_manager instance-attribute 

compiler_manager: CompilerManager = CompilerManager(
    compilation_config
)

graph instance-attribute 

graph: GraphModule

graph_pool instance-attribute 

graph_pool: Any = global_graph_pool

input_buffers instance-attribute 

input_buffers: list[Tensor] = []

piecewise_graphs instance-attribute 

piecewise_graphs: list[SplitItem]

post_grad_pass_manager instance-attribute 

post_grad_pass_manager = PostGradPassManager()

post_grad_passes instance-attribute 

post_grad_passes: Sequence[Callable]

returned_callable instance-attribute 

returned_callable: Callable

split_gm instance-attribute 

split_gm: GraphModule

sym_tensor_indices instance-attribute 

sym_tensor_indices: list[int] = []

vllm_config instance-attribute 

vllm_config: VllmConfig = vllm_config

__call__ 

__call__(graph: GraphModule, example_inputs) -> Callable
Source code in vllm/compilation/backends.py 
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
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
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
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
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:

    vllm_config = self.vllm_config
    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.

        factors = []
        # 0. factors come from the env, for example, The values of
        # VLLM_PP_LAYER_PARTITION will affects the computation graph.
        env_hash = envs.compute_hash()
        factors.append(env_hash)

        # 1. factors come from the vllm_config (it mainly summarizes how the
        #    model is created)
        config_hash = vllm_config.compute_hash()
        factors.append(config_hash)

        # 2. factors come from the code files that are traced by Dynamo (
        #    it mainly summarizes how the model is used in forward pass)
        forward_code_files = list(
            sorted(self.compilation_config.traced_files))
        self.compilation_config.traced_files.clear()
        logger.debug(
            "Traced files (to be considered for compilation cache):\n%s",
            "\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
            with open(filepath) as f:
                hash_content.append(f.read())
        import hashlib
        code_hash = hashlib.md5("\n".join(hash_content).encode(),
                                usedforsecurity=False).hexdigest()
        factors.append(code_hash)

        # 3. compiler hash
        compiler_hash = self.compiler_manager.compute_hash(vllm_config)
        factors.append(compiler_hash)

        # combine all factors to generate the cache dir
        hash_key = hashlib.md5(str(factors).encode(),
                               usedforsecurity=False).hexdigest()[:10]

        cache_dir = os.path.join(
            envs.VLLM_CACHE_ROOT,
            "torch_compile_cache",
            hash_key,
        )
        self.compilation_config.cache_dir = cache_dir

    if compilation_counter.num_graphs_seen > 0:
        cache_dir = self.compilation_config.cache_dir + \
            f'-{compilation_counter.num_graphs_seen}'
    else:
        cache_dir = self.compilation_config.cache_dir
    os.makedirs(cache_dir, exist_ok=True)
    self.compilation_config.cache_dir = cache_dir
    rank = vllm_config.parallel_config.rank
    dp_rank = vllm_config.parallel_config.data_parallel_rank
    local_cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}")
    os.makedirs(local_cache_dir, exist_ok=True)
    self.compilation_config.local_cache_dir = local_cache_dir

    disable_cache = envs.VLLM_DISABLE_COMPILE_CACHE

    if disable_cache:
        logger.info("vLLM's torch.compile cache is disabled.")
    else:
        logger.info("Using cache directory: %s for vLLM's torch.compile",
                    local_cache_dir)

    self.compiler_manager.initialize_cache(local_cache_dir, disable_cache)

    # when dynamo calls the backend, it means the bytecode
    # transform and analysis are done
    compilation_counter.num_graphs_seen += 1
    from .monitor import torch_compile_start_time
    dynamo_time = time.time() - torch_compile_start_time
    logger.info("Dynamo bytecode transform time: %.2f s", dynamo_time)
    self.compilation_config.compilation_time += dynamo_time

    # 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
    self.configure_post_pass()

    self.split_gm, self.piecewise_graphs = split_graph(
        graph, self.compilation_config.splitting_ops)

    from torch._dynamo.utils import lazy_format_graph_code

    # 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)

    compilation_counter.num_piecewise_graphs_seen += len(
        self.piecewise_graphs)
    submod_names_to_compile = [
        item.submod_name for item in self.piecewise_graphs
        if not item.is_splitting_graph
    ]

    # propagate the split graph to the piecewise backend,
    # compile submodules with symbolic shapes
    PiecewiseCompileInterpreter(self.split_gm, submod_names_to_compile,
                                self.vllm_config, self.graph_pool,
                                self).run(*example_inputs)

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

        logger.debug("Computation graph saved to %s", graph_path)

    self._called = True

    if not self.compilation_config.use_cudagraph or \
        not self.compilation_config.cudagraph_copy_inputs:
        return self.split_gm

    # if we need to copy input buffers for cudagraph
    from torch._guards import detect_fake_mode
    fake_mode = detect_fake_mode()
    fake_args = [
        fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
        for t in example_inputs
    ]

    # index of tensors that have symbolic shapes (batch size)
    # 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
    self.sym_tensor_indices = [
        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())
    ]

    # compiler managed cudagraph input buffers
    # we assume the first run with symbolic shapes
    # has the maximum size among all the tensors
    self.input_buffers = [
        example_inputs[x].clone() for x in self.sym_tensor_indices
    ]

    # this is the callable we return to Dynamo to run
    def copy_and_call(*args):
        list_args = list(args)
        for i, index in enumerate(self.sym_tensor_indices):
            runtime_tensor = list_args[index]
            runtime_shape = runtime_tensor.shape[0]
            static_tensor = self.input_buffers[i][:runtime_shape]

            # copy the tensor to the static buffer
            static_tensor.copy_(runtime_tensor)

            # replace the tensor in the list_args to the static buffer
            list_args[index] = static_tensor
        return self.split_gm(*list_args)

    return copy_and_call

__init__ 

__init__(vllm_config: VllmConfig)
Source code in vllm/compilation/backends.py 
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
def __init__(
    self,
    vllm_config: VllmConfig,
):
    global global_graph_pool
    if global_graph_pool is None:
        global_graph_pool = current_platform.graph_pool_handle()

    # TODO: in the future, if we want to use multiple
    # streams, it might not be safe to share a global pool.
    # only investigate this when we use multiple streams
    self.graph_pool = global_graph_pool

    # Passes to run on the graph post-grad.
    self.post_grad_pass_manager = PostGradPassManager()

    self.sym_tensor_indices = []
    self.input_buffers = []

    self.vllm_config = vllm_config
    self.compilation_config = vllm_config.compilation_config

    self.compiler_manager: CompilerManager = CompilerManager(
        self.compilation_config)

configure_post_pass 

configure_post_pass()
Source code in vllm/compilation/backends.py 
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
def configure_post_pass(self):
    config = self.compilation_config
    self.post_grad_pass_manager.configure(self.vllm_config)

    # 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.
    inductor_config = config.inductor_compile_config
    PASS_KEY = "post_grad_custom_post_pass"
    if PASS_KEY in inductor_config:
        # Config should automatically wrap all inductor passes
        if isinstance(inductor_config[PASS_KEY], PostGradPassManager):
            assert (inductor_config[PASS_KEY].uuid() ==
                    self.post_grad_pass_manager.uuid())
        else:
            assert isinstance(inductor_config[PASS_KEY], InductorPass)
            self.post_grad_pass_manager.add(inductor_config[PASS_KEY])
    inductor_config[PASS_KEY] = self.post_grad_pass_manager

make_compiler 

make_compiler(
    compilation_config: CompilationConfig,
) -> CompilerInterface
Source code in vllm/compilation/backends.py 
29
30
31
32
33
34
35
36
37
38
39
def make_compiler(compilation_config: CompilationConfig) -> CompilerInterface:
    if compilation_config.use_inductor:
        if envs.VLLM_TEST_STANDALONE_COMPILE:
            logger.info("Using InductorStandaloneAdaptor")
            return InductorStandaloneAdaptor()
        else:
            logger.info("Using InductorAdaptor")
            return InductorAdaptor()
    else:
        logger.info("Using EagerAdaptor")
        return EagerAdaptor()

split_graph 

split_graph(
    graph: GraphModule, ops: list[str]
) -> tuple[GraphModule, list[SplitItem]]
Source code in vllm/compilation/backends.py 
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
def split_graph(graph: fx.GraphModule,
                ops: list[str]) -> tuple[fx.GraphModule, list[SplitItem]]:
    # split graph by ops
    subgraph_id = 0
    node_to_subgraph_id = {}
    split_op_graphs = []
    for node in graph.graph.nodes:
        if node.op in ("output", "placeholder"):
            continue
        if node.op == 'call_function' and str(node.target) in ops:
            subgraph_id += 1
            node_to_subgraph_id[node] = subgraph_id
            split_op_graphs.append(subgraph_id)
            subgraph_id += 1
        else:
            node_to_subgraph_id[node] = subgraph_id

    # `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
    split_gm = torch.fx.passes.split_module.split_module(
        graph,
        None,
        lambda node: node_to_subgraph_id[node],
        keep_original_order=True)

    outputs = []

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

    for name in names:
        if "." in name or name == "":
            # recursive child module or the root module
            continue

        module = getattr(split_gm, name)

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

    # sort by intetger graph_id, rather than string name
    outputs.sort(key=lambda x: x.graph_id)

    return split_gm, outputs