backends.py 22 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
7
import os
import pprint
8
import time
9
10
from collections.abc import Sequence
from typing import Any, Callable, Optional
11
12
13

import torch
import torch.fx as fx
14
from torch._dispatch.python import enable_python_dispatcher
15

16
import vllm.envs as envs
17
from vllm.config import CompilationConfig, VllmConfig
18
from vllm.logger import init_logger
19
from vllm.platforms import current_platform
20
from vllm.utils import is_torch_equal_or_newer, resolve_obj_by_qualname
21

22
23
from .compiler_interface import (CompilerInterface, EagerAdaptor,
                                 InductorAdaptor, InductorStandaloneAdaptor)
24
from .counter import compilation_counter
25
26
from .inductor_pass import InductorPass
from .pass_manager import PostGradPassManager
27
28
29

logger = init_logger(__name__)

30

31
32
def make_compiler(compilation_config: CompilationConfig) -> CompilerInterface:
    if compilation_config.use_inductor:
33
34
        if envs.VLLM_USE_STANDALONE_COMPILE and is_torch_equal_or_newer(
                "2.8.0"):
35
            logger.debug("Using InductorStandaloneAdaptor")
36
37
            return InductorStandaloneAdaptor()
        else:
38
            logger.debug("Using InductorAdaptor")
39
40
            return InductorAdaptor()
    else:
41
        logger.debug("Using EagerAdaptor")
42
43
44
        return EagerAdaptor()


45
46
47
48
49
class CompilerManager:
    """
    A manager to manage the compilation process, including
    caching the compiled graph, loading the compiled graph,
    and compiling the graph.
50

51
52
53
    The cache is a dict mapping
    `(runtime_shape, graph_index, backend_name)`
    to `any_data` returned from the compiler.
54

55
56
57
    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.
58
59
    """

60
    def __init__(self, compilation_config: CompilationConfig):
61
        self.cache: dict[tuple[Optional[int], int, str], Any] = dict()
62
        self.is_cache_updated = False
63
64
        self.compilation_config = compilation_config
        self.compiler = make_compiler(compilation_config)
65

66
67
    def compute_hash(self, vllm_config: VllmConfig) -> str:
        return self.compiler.compute_hash(vllm_config)
68

69
70
    def initialize_cache(self, cache_dir: str, disable_cache: bool = False):
        self.disable_cache = disable_cache
71
        self.cache_dir = cache_dir
72
73
74
75
        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
76
            with open(self.cache_file_path) as f:
77
78
79
80
81
82
83
                # 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)
84
85

    def save_to_file(self):
86
        if self.disable_cache or not self.is_cache_updated:
87
            return
88
89
        printer = pprint.PrettyPrinter(indent=4)
        data = printer.pformat(self.cache)
90
        with open(self.cache_file_path, "w") as f:
91
92
93
94
            f.write(data)

    def load(self,
             graph: fx.GraphModule,
95
             example_inputs: list[Any],
96
97
98
99
100
101
102
             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)
103
        logger.debug(
104
105
106
107
108
109
110
111
112
113
114
115
116
            "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:
117
        if graph_index == 0:
118
119
120
121
122
123
124
125
126
127
128
129
            # 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:
130
131
132
133
134
135
136
137
            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)
138
139
140
141
            return compiled_graph

        # no compiler cached the graph, or the cache is disabled,
        # we need to compile it
142
143
144
145
146
147
        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}"
148
        compiled_graph, handle = self.compiler.compile(
149
150
            graph, example_inputs, additional_inductor_config, runtime_shape,
            maybe_key)
151
152
153
154
155
156
157

        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
158
            self.is_cache_updated = True
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
            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)
178

179
        return compiled_graph
180
181


182
183
184
@dataclasses.dataclass
class SplitItem:
    submod_name: str
185
    graph_id: int
186
187
188
189
190
    is_splitting_graph: bool
    graph: fx.GraphModule


def split_graph(graph: fx.GraphModule,
191
                ops: list[str]) -> tuple[fx.GraphModule, list[SplitItem]]:
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
    # 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(
212
        graph,
213
214
215
        None,
        lambda node: node_to_subgraph_id[node],
        keep_original_order=True)
216

217
    outputs = []
218

219
    names = [name for (name, module) in split_gm.named_modules()]
220

221
222
223
224
    for name in names:
        if "." in name or name == "":
            # recursive child module or the root module
            continue
225

226
        module = getattr(split_gm, name)
227

228
        graph_id = int(name.replace("submod_", ""))
229
230
231
232
233
        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)
234

235
    return split_gm, outputs
236
237


238
239
240
# we share the global graph pool among all the backends
global_graph_pool = None

241
242
compilation_start_time = 0.0

243
244
245
246
247
248

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.
249
250
251
252
253

    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.
254
255
256
    """

    def __init__(self, module: torch.fx.GraphModule,
257
                 compile_submod_names: list[str], vllm_config: VllmConfig,
258
                 graph_pool, vllm_backend: "VllmBackend"):
259
260
261
262
        super().__init__(module)
        from torch._guards import detect_fake_mode
        self.fake_mode = detect_fake_mode()
        self.compile_submod_names = compile_submod_names
263
        self.compilation_config = vllm_config.compilation_config
264
        self.graph_pool = graph_pool
265
        self.vllm_config = vllm_config
266
        self.vllm_backend = vllm_backend
267
268
        # When True, it annoyingly dumps the torch.fx.Graph on errors.
        self.extra_traceback = False
269
270
271
272
273
274

    def run(self, *args):
        fake_args = [
            self.fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
            for t in args
        ]
275
        with self.fake_mode, enable_python_dispatcher():
276
            return super().run(*fake_args)
277
278

    def call_module(self, target: torch.fx.node.Target,
279
280
                    args: tuple[torch.fx.node.Argument,
                                ...], kwargs: dict[str, Any]) -> Any:
281
282
283
284
        assert isinstance(target, str)
        output = super().call_module(target, args, kwargs)

        if target in self.compile_submod_names:
285
            index = self.compile_submod_names.index(target)
286
287
288
289
            submod = self.fetch_attr(target)
            sym_shape_indices = [
                i for i, x in enumerate(args) if isinstance(x, torch.SymInt)
            ]
290
            global compilation_start_time
291
292
            compiled_graph_for_general_shape = self.vllm_backend.\
                compiler_manager.compile(
293
294
                submod,
                args,
295
296
                self.compilation_config.inductor_compile_config,
                self.compilation_config,
297
298
                graph_index=index,
                num_graphs=len(self.compile_submod_names),
299
                runtime_shape=None)
300

301
302
303
            piecewise_backend = resolve_obj_by_qualname(
                current_platform.get_piecewise_backend_cls())
            self.module.__dict__[target] = piecewise_backend(
304
                submod, self.vllm_config, self.graph_pool, index,
305
                len(self.compile_submod_names), sym_shape_indices,
306
                compiled_graph_for_general_shape, self.vllm_backend)
307
308
309
310
311
312

            compilation_counter.num_piecewise_capturable_graphs_seen += 1

        return output


313
class VllmBackend:
314
    """The compilation backend for `torch.compile` with vLLM.
315
316
    It is used for compilation level of `CompilationLevel.PIECEWISE`,
    where we customize the compilation.
317

318
319
    The major work of this backend is to split the graph into
    piecewise graphs, and pass them to the piecewise backend.
320

321
322
    This backend also adds the PostGradPassManager to Inductor config,
    which handles the post-grad passes.
323
    """
324

325
326
    vllm_config: VllmConfig
    compilation_config: CompilationConfig
327
328
329
330
331
332
    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
333
    piecewise_graphs: list[SplitItem]
334
    returned_callable: Callable
335
336
    # Inductor passes to run on the graph pre-defunctionalization
    post_grad_passes: Sequence[Callable]
337
338
    sym_tensor_indices: list[int]
    input_buffers: list[torch.Tensor]
339
    compiler_manager: CompilerManager
340

341
342
    def __init__(
        self,
343
        vllm_config: VllmConfig,
344
    ):
345
346
        global global_graph_pool
        if global_graph_pool is None:
347
            global_graph_pool = current_platform.graph_pool_handle()
348
349
350
351
352

        # 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
353
354
355

        # Passes to run on the graph post-grad.
        self.post_grad_pass_manager = PostGradPassManager()
356

357
358
359
        self.sym_tensor_indices = []
        self.input_buffers = []

360
361
        self.vllm_config = vllm_config
        self.compilation_config = vllm_config.compilation_config
362

363
        self.compiler_manager: CompilerManager = CompilerManager(
364
            self.compilation_config)
365

366
367
368
        # `torch.compile` is JIT compiled, so we don't need to
        # do anything here

369
    def configure_post_pass(self):
370
        config = self.compilation_config
371
        self.post_grad_pass_manager.configure(self.vllm_config)
372

373
374
        # 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.
375
        inductor_config = config.inductor_compile_config
376
377
378
        PASS_KEY = "post_grad_custom_post_pass"
        if PASS_KEY in inductor_config:
            # Config should automatically wrap all inductor passes
379
380
381
382
383
384
            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])
385
        inductor_config[PASS_KEY] = self.post_grad_pass_manager
386

387
388
    def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:

389
        vllm_config = self.vllm_config
390
391
392
393
394
395
        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.

396
            factors = []
397
398
399
400
401
            # 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)

402
403
404
            # 1. factors come from the vllm_config (it mainly summarizes how the
            #    model is created)
            config_hash = vllm_config.compute_hash()
405
            factors.append(config_hash)
406
407
408
409
410
411
412
413
414
415
416
417

            # 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)
418
419
420
421
                if filepath == "<string>":
                    # This means the function was dynamically generated, with
                    # e.g. exec(). We can't actually check these.
                    continue
422
423
424
                with open(filepath) as f:
                    hash_content.append(f.read())
            import hashlib
425
426
            code_hash = hashlib.md5("\n".join(hash_content).encode(),
                                    usedforsecurity=False).hexdigest()
427
428
429
430
431
432
433
            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
434
435
            hash_key = hashlib.md5(str(factors).encode(),
                                   usedforsecurity=False).hexdigest()[:10]
436
437

            cache_dir = os.path.join(
438
439
440
441
442
443
                envs.VLLM_CACHE_ROOT,
                "torch_compile_cache",
                hash_key,
            )
            self.compilation_config.cache_dir = cache_dir

444
445
446
447
448
        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
449
        os.makedirs(cache_dir, exist_ok=True)
450
        self.compilation_config.cache_dir = cache_dir
451
452
453
        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}")
454
        os.makedirs(local_cache_dir, exist_ok=True)
455
        self.compilation_config.local_cache_dir = local_cache_dir
456

457
458
459
        disable_cache = envs.VLLM_DISABLE_COMPILE_CACHE

        if disable_cache:
460
461
462
            logger.info("vLLM's torch.compile cache is disabled.")
        else:
            logger.info("Using cache directory: %s for vLLM's torch.compile",
463
                        local_cache_dir)
464

465
466
        self.compiler_manager.initialize_cache(local_cache_dir, disable_cache)

467
468
        # when dynamo calls the backend, it means the bytecode
        # transform and analysis are done
469
        compilation_counter.num_graphs_seen += 1
470
471
472
        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)
473
        self.compilation_config.compilation_time += dynamo_time
474
475
476
477
478
479

        # 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
480
        self.configure_post_pass()
481
482

        self.split_gm, self.piecewise_graphs = split_graph(
483
            graph, self.compilation_config.splitting_ops)
484

485
        from torch._dynamo.utils import lazy_format_graph_code
486
487
488
489
490

        # 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)
491

492
493
494
495
496
497
498
499
500
501
        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,
502
503
                                    self.vllm_config, self.graph_pool,
                                    self).run(*example_inputs)
504

505
506
507
508
509
510
511
512
513
514
515
516
        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)

517
518
        self._called = True

519
520
        if not self.compilation_config.use_cudagraph or \
            not self.compilation_config.cudagraph_copy_inputs:
521
522
523
524
525
526
527
528
529
530
531
            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)
532
533
534
        # 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
535
536
        self.sym_tensor_indices = [
            i for i, x in enumerate(fake_args)
537
538
            if isinstance(x, torch._subclasses.fake_tensor.FakeTensor) and \
                any(is_symbolic(d) for d in x.size())
539
540
541
542
543
544
545
546
547
        ]

        # 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
        ]

youkaichao's avatar
youkaichao committed
548
549
        # this is the callable we return to Dynamo to run
        def copy_and_call(*args):
550
551
552
553
554
555
556
557
558
559
560
561
562
563
            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