wrapper.py 13.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
6
import os
import sys
from abc import abstractmethod
7
from collections.abc import Callable, Generator
8
from contextlib import contextmanager, nullcontext
9
from types import CodeType
10
from typing import Any, ParamSpec, TypeVar
11
12
13

import torch

14
import vllm.envs as envs
15
from vllm.config import CompilationMode, CUDAGraphMode, get_current_vllm_config
16
from vllm.config.compilation import DynamicShapesType
17
from vllm.logger import init_logger
18
from vllm.utils.nvtx_pytorch_hooks import layerwise_nvtx_marker_context
19
20

logger = init_logger(__name__)
21

22
23
R = TypeVar("R")
P = ParamSpec("P")
24

25

26
@contextmanager
27
def _compilation_context() -> Generator[None, None, None]:
28
29
30
31
32
33
    """Context manager for compilation settings.

    This manager sets higher dynamo cache limits for compilation.
    (Needed for qwen2_5_vl see test_qwen2_5_vl_evs_functionality).
    Generally a recompilation can happen whenever we use a new
    backend instance in torch.compile.
34
    """
35
36
37
38
39
40
41
42
43
44
45
46
47
    original_cache_size = torch._dynamo.config.cache_size_limit
    original_accumulated_cache = torch._dynamo.config.accumulated_cache_size_limit

    try:
        torch._dynamo.config.cache_size_limit = 2048
        torch._dynamo.config.accumulated_cache_size_limit = 8192
        yield
    finally:
        torch._dynamo.config.cache_size_limit = original_cache_size
        torch._dynamo.config.accumulated_cache_size_limit = original_accumulated_cache


class TorchCompileWithNoGuardsWrapper:
48
    """
49
50
51
52
53
54
55
    A wrapper class for torch.compile, it ensures that all guards are dropped
    when CompilationMode is not CompilationMode.STOCK_TORCH_COMPILE.
    When guards are dropped, the first time __call__ is invoked, a single
    compilation is triggered. Dynamo should never be traced again after that
    since we drop all guards.
    """

56
    def check_invariants_and_forward(self, *args: Any, **kwargs: Any) -> Any:
57
58
59
60
61
        assert hasattr(self, "_check_shape_invariants")
        self._check_shape_invariants(*args, **kwargs)

        return self.forward(*args, **kwargs)

62
63
64
    def _call_with_optional_nvtx_range(
        self, callable_fn: Callable[P, R], *args: P.args, **kwargs: P.kwargs
    ) -> Any:
65
66
67
68
69
70
71
72
73
74
75
76
77
        if self.layerwise_nvtx_tracing_enabled:
            args_list = list(args)
            kwargs_dict = dict(kwargs)
            with layerwise_nvtx_marker_context(
                "Torch Compiled Module (input):{}".format(self.__class__.__name__),
                self,
                in_tensor=args_list,
                kwargs=kwargs_dict,
            ) as ctx:
                ctx.result = callable_fn(*args, **kwargs)
            return ctx.result
        return callable_fn(*args, **kwargs)

78
    def __init__(self) -> None:
79
        self.compiled = False
80

81
82
        vllm_config = get_current_vllm_config()
        self.vllm_config = vllm_config
83
        mode = vllm_config.compilation_config.mode
84
85
86
        self.layerwise_nvtx_tracing_enabled = (
            vllm_config.observability_config.enable_layerwise_nvtx_tracing
        )
87
88
89
90
91
92
93
94
95
        if mode is None:
            raise RuntimeError("Compilation mode cannot be NO_COMPILATION")

        backend = vllm_config.compilation_config.init_backend(vllm_config)
        options = {}

        if isinstance(backend, str) and backend == "inductor":
            options = vllm_config.compilation_config.inductor_compile_config

96
97
98
99
100
101
102
        self.first_compile = True
        self.evaluate_guards = (
            vllm_config.compilation_config.dynamic_shapes_config.evaluate_guards
        )

        ds_type = vllm_config.compilation_config.dynamic_shapes_config.type

103
104
        if mode != CompilationMode.STOCK_TORCH_COMPILE:
            # Drop all the guards.
105
106
107
108
109
            if self.evaluate_guards:
                assert not envs.VLLM_USE_BYTECODE_HOOK, (
                    "compilation_config.dynamic_shapes_config.evaluate_guards "
                    "requires VLLM_USE_BYTECODE_HOOK=0. "
                )
110

111
112
113
114
                options["guard_filter_fn"] = lambda x: [
                    entry.guard_type == "SHAPE_ENV" for entry in x
                ]
            else:
115
                options["guard_filter_fn"] = torch.compiler.skip_all_guards_unsafe
116
117

        compiled_ptr: Any = self.forward
118
119
        # Validate that unbacked dynamic shapes require VLLM_USE_BYTECODE_HOOK=False

120
        if ds_type == DynamicShapesType.UNBACKED:
121
122
123
124
125
126
127
128
129
130
            # reason is that bytecode does torch._dynamo.eval_frame.
            # remove_from_cache(self.original_code_object()) to force a new
            # re-compilation. And if we use
            # compiled_ptr = self.check_invariants_and_forward
            # it will reset all entries.
            assert not envs.VLLM_USE_BYTECODE_HOOK, (
                "UNBACKED dynamic shapes requires VLLM_USE_BYTECODE_HOOK=0. "
            )
            assert not self.evaluate_guards, "UNBACKED dynamic shapes do not add guards"

131
132
            compiled_ptr = self.check_invariants_and_forward

133
        aot_context = nullcontext()
134
135
        if envs.VLLM_USE_AOT_COMPILE:
            if hasattr(torch._dynamo.config, "enable_aot_compile"):
136
                aot_context = torch._dynamo.config.patch(enable_aot_compile=True)
137
138
139
140
141
142
            else:
                msg = "torch._dynamo.config.enable_aot_compile is not "
                msg += "available. AOT compile is disabled and please "
                msg += "upgrade PyTorch version to use AOT compile."
                logger.warning(msg)

143
144
145
146
147
148
149
150
        with aot_context:
            self._compiled_callable = torch.compile(
                compiled_ptr,
                fullgraph=True,
                dynamic=False,
                backend=backend,
                options=options,
            )
151

152
153
        if envs.VLLM_USE_BYTECODE_HOOK and mode != CompilationMode.STOCK_TORCH_COMPILE:
            torch._dynamo.convert_frame.register_bytecode_hook(self.bytecode_hook)
154
            self._compiled_bytecode: CodeType | None = None
155

156
    def aot_compile(self, *args: Any, **kwargs: Any) -> Any:
157
        if not hasattr(self._compiled_callable, "aot_compile"):
158
159
            raise RuntimeError(
                "aot_compile is not supported by the current configuration. "
160
161
                "Please make sure torch.compile is enabled with the latest "
                f"version of PyTorch (current using torch: {torch.__version__})"
162
            )
163
        return self._compiled_callable.aot_compile((args, kwargs))
164

165
    def __call__(self, *args: Any, **kwargs: Any) -> Any:
166
167
168
169
170
171
172
173
174
175
176
        if envs.VLLM_USE_BYTECODE_HOOK:
            if (
                self.vllm_config.compilation_config.mode
                == CompilationMode.STOCK_TORCH_COMPILE
            ):
                return self._compiled_callable(*args, **kwargs)

            if not self._compiled_bytecode:
                # Make sure a compilation is triggered by clearing dynamo
                # cache.
                torch._dynamo.eval_frame.remove_from_cache(self.original_code_object())
177
178
179
                return self._call_with_optional_nvtx_range(
                    self._compiled_callable, *args, **kwargs
                )
180
181
            else:
                with self._dispatch_to_compiled_code():
182
183
184
                    return self._call_with_optional_nvtx_range(
                        self.forward, *args, **kwargs
                    )
185
        else:
186
187
188
189
190
191
192
            ctx = (
                nullcontext()
                if self.first_compile or not self.evaluate_guards
                else torch.compiler.set_stance("fail_on_recompile")
            )
            self.first_compile = False
            with _compilation_context(), ctx:
193
194
195
                return self._call_with_optional_nvtx_range(
                    self._compiled_callable, *args, **kwargs
                )
196
197

    @abstractmethod
198
    def forward(self, *args: Any, **kwargs: Any) -> Any: ...
199

200
201
202
203
    def original_code_object(self) -> CodeType:
        """Return the original code object of the forward method."""
        return self.__class__.forward.__code__

204
    def bytecode_hook(self, old_code: CodeType, new_code: CodeType) -> None:
205
        """Hook to save the compiled bytecode for direct execution."""
206
        if old_code is not self.original_code_object():
207
208
209
            return
        # code borrowed from https://github.com/thuml/depyf/blob/f4ad79fadee27ea113b4c75202db1eb1a11c0dbc/depyf/explain/enable_debugging.py#L25
        frame = sys._getframe()
210
        while frame and frame.f_back:
211
212
213
214
215
216
217
218
219
220
221
            frame = frame.f_back
            code_name = frame.f_code.co_name
            file_name = frame.f_code.co_filename.split(os.path.sep)[-1]
            if code_name == "_compile" and file_name == "convert_frame.py":
                break
        frame = frame.f_locals["frame"]
        assert frame.f_code == old_code

        if frame.f_locals["self"] is not self:
            return

222
        self._compiled_bytecode = new_code
223
224
225
226
227

        path = self.vllm_config.compile_debug_dump_path()
        if path:
            decompiled_file = path / "transformed_code.py"
            if not decompiled_file.exists():
228
229
230
231
232
                try:
                    # usually the decompilation will succeed for most models,
                    # as we guarantee a full-graph compilation in Dynamo.
                    # but there's no 100% guarantee, since decompliation is
                    # not a reversible process.
233
                    import depyf
234

235
                    src = depyf.decompile(new_code)
236

237
238
239
                    with open(decompiled_file, "w") as f:
                        f.write(src)

240
                    logger.debug("Dynamo transformed code saved to %s", decompiled_file)
241
242
                except Exception:
                    pass
243

244
245
246
247
        if (
            self.vllm_config.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and "update" in new_code.co_names
        ):
248
            import depyf
249

250
            src = depyf.decompile(new_code)
251
            msg = (
252
253
254
255
256
257
                "Assigning / modifying buffers of nn.Module during forward pass is not "
                "allowed when using cudagraph inside the compiler because it will "
                "cause silent errors. Please use eager mode or fix the code. The "
                "following code contains clues about which buffer is being modified "
                f"(please search for the usage of the function `update`):\n{src}"
            )
258
259
            raise RuntimeError(msg)

260
    @contextmanager
261
    def _dispatch_to_compiled_code(self) -> Generator[None, None, None]:
262
263
264
        # noqa: E501
        """
        Context manager to dispatch to internally compiled code for torch<2.8.
265
266
267
268
269
        Why does this work? Because Dynamo guarantees that the compiled
        bytecode has exactly the same arguments, cell variables, and free
        variables as the original code. Therefore we can directly switch
        the code object in the function and call it.

270
271
272
273
274
275
276
277
278
        See https://dev-discuss.pytorch.org/t/what-is-the-relationship-requirement-among-original-bytecode-transformed-bytecode-and-bytecode-returned-by-hooks-in-dynamo/1693/7 for more details.
        """  # noqa: E501 line too long
        original = self.original_code_object()
        assert self._compiled_bytecode is not None
        self.__class__.forward.__code__ = self._compiled_bytecode
        try:
            yield
        finally:
            self.__class__.forward.__code__ = original
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
308


def reset_compile_wrapper(model: torch.nn.Module) -> None:
    """
    Clean up compiled model and captured CUDA graphs for elastic EP.
    """
    if not isinstance(model, TorchCompileWithNoGuardsWrapper) and hasattr(
        model, "model"
    ):
        model = model.model
    if not isinstance(model, TorchCompileWithNoGuardsWrapper):
        return
    # model.do_not_compile is set by the @support_torch_compile decorator
    if hasattr(model, "do_not_compile") and model.do_not_compile:
        return
    from vllm.compilation.counter import compilation_counter

    # reset the compilation counter
    compilation_counter.num_models_seen = 0
    compilation_counter.num_graphs_seen = 0
    compilation_counter.num_piecewise_graphs_seen = 0
    compilation_counter.num_piecewise_capturable_graphs_seen = 0
    compilation_counter.num_backend_compilations = 0
    compilation_counter.num_gpu_runner_capture_triggers = 0
    compilation_counter.num_cudagraph_captured = 0
    compilation_counter.num_inductor_compiles = 0
    compilation_counter.num_eager_compiles = 0
    compilation_counter.num_cache_entries_updated = 0
    compilation_counter.num_compiled_artifacts_saved = 0
    compilation_counter.stock_torch_compile_count = 0
309
310
311
    compilation_counter.num_aot_compiles = 0
    compilation_counter.num_aot_artifacts_saved = 0
    compilation_counter.num_aot_artifacts_loaded = 0
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330

    # Clear the AOT compiled function so the model is forced to
    # recompile on the next call. Without this, decorators.py
    # __call__ uses the stale aot_compiled_fn whose torchinductor
    # kernels have old parameters (expert_map size for example)
    # baked in as compile-time constants.
    if hasattr(model, "aot_compiled_fn"):
        model.aot_compiled_fn = None
    if hasattr(model, "was_aot_compile_fn_loaded_from_disk"):
        model.was_aot_compile_fn_loaded_from_disk = False

    # Reset the cache_dir so VllmBackend recomputes the hash
    # (data_parallel_size changed, so the config hash differs).
    compilation_config = model.vllm_config.compilation_config
    compilation_config.cache_dir = ""
    compilation_config.local_cache_dir = ""

    model.__class__.forward.__code__ = model.original_code_object()
    TorchCompileWithNoGuardsWrapper.__init__(model)