A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 02:15 UTC
[Bugfix] Upgrade depyf to 0.19 and streamline custom pass logging (#20777)
Signed-off-by: Luka Govedic <lgovedic@redhat.com> Signed-off-by: luka <lgovedic@redhat.com>
Luka Govedič committed
Jul 11, 2025 at 03:15 UTC
762be26a8ee0de15638fa21a59d85efedacec847
4 files changed
+13
-36
requirements/common.txt
+1
-1
@@ -40,7 +40,7 @@ six>=1.16.0; python_version > '3.11' # transitive dependency of pandas that need
40
setuptools>=77.0.3,<80; python_version > '3.11' # Setuptools is used by triton, we need to ensure a modern version is installed for 3.12+ so that it does not try to import distutils, which was removed in 3.12
41
einops # Required for Qwen2-VL.
42
compressed-tensors == 0.10.2 # required for compressed-tensors
43
-depyf==0.18.0 # required for profiling and debugging with compilation config
43
+depyf==0.19.0 # required for profiling and debugging with compilation config
44
cloudpickle # allows pickling lambda functions in model_executor/models/registry.py
45
watchfiles # required for http server to monitor the updates of TLS files
46
python-json-logger # Used by logging as per examples/others/logging_configuration.md
tests/compile/test_full_graph.py
+6
@@ -3,6 +3,7 @@
3
4
from __future__ import annotations
5
6
+import tempfile
7
from typing import Any, Optional, Union
8
9
import pytest
@@ -111,6 +112,11 @@ def test_full_graph(
112
pass_config=PassConfig(enable_fusion=True,
113
enable_noop=True)), model)
114
for model in models_list(keywords=["FP8-dynamic", "quantized.w8a8"])
115
+ ] + [
116
+ # Test depyf integration works
117
+ (CompilationConfig(level=CompilationLevel.PIECEWISE,
118
+ debug_dump_path=tempfile.gettempdir()),
119
+ ("facebook/opt-125m", {})),
120
])
121
# only test some of the models
122
@create_new_process_for_each_test()
vllm/compilation/vllm_inductor_pass.py
+4
-24
@@ -6,13 +6,7 @@ import time
6
import torch
7
from torch._dynamo.utils import lazy_format_graph_code
8
9
-from vllm.config import PassConfig, VllmConfig
10
-# yapf: disable
11
-from vllm.distributed import get_tensor_model_parallel_rank as get_tp_rank
12
-from vllm.distributed import (
13
- get_tensor_model_parallel_world_size as get_tp_world_size)
14
-from vllm.distributed import model_parallel_is_initialized as p_is_init
15
-# yapf: enable
9
+from vllm.config import VllmConfig
10
from vllm.logger import init_logger
11
12
from .inductor_pass import InductorPass
@@ -34,22 +28,9 @@ class VllmInductorPass(InductorPass):
28
else None
29
self.pass_name = self.__class__.__name__
30
37
- def dump_graph(self, graph: torch.fx.Graph, stage: str, always=False):
31
+ def dump_graph(self, graph: torch.fx.Graph, stage: str):
32
lazy_format_graph_code(stage, graph.owning_module)
33
40
- if stage in self.pass_config.dump_graph_stages or always:
41
- # Make sure filename includes rank in the distributed setting
42
- parallel = p_is_init() and get_tp_world_size() > 1
43
- rank = f"-{get_tp_rank()}" if parallel else ""
44
- filepath = self.pass_config.dump_graph_dir / f"{stage}{rank}.py"
45
-
46
- logger.info("%s printing graph to %s", self.pass_name, filepath)
47
- with open(filepath, "w") as f:
48
- src = graph.python_code(root_module="self", verbose=True).src
49
- # Add imports so it's not full of errors
50
- print("import torch; from torch import device", file=f)
51
- print(src, file=f)
52
-
34
def begin(self):
35
self._start_time = time.perf_counter_ns()
36
@@ -61,10 +42,9 @@ class VllmInductorPass(InductorPass):
42
43
class PrinterInductorPass(VllmInductorPass):
44
64
- def __init__(self, name: str, config: PassConfig, always=False):
45
+ def __init__(self, name: str, config: VllmConfig):
46
super().__init__(config)
47
self.name = name
67
- self.always = always
48
49
def __call__(self, graph: torch.fx.Graph):
70
- self.dump_graph(graph, self.name, always=self.always)
50
+ self.dump_graph(graph, self.name)
vllm/config.py
+2
-11
@@ -16,7 +16,6 @@ from dataclasses import (MISSING, Field, asdict, field, fields, is_dataclass,
16
replace)
17
from functools import cached_property
18
from importlib.util import find_spec
19
-from pathlib import Path
19
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Literal, Optional,
20
Protocol, TypeVar, Union, cast, get_args)
21
@@ -3953,11 +3952,6 @@ class PassConfig:
3952
don't all have access to full configuration - that would create a cycle as
3953
the `PassManager` is set as a property of config."""
3954
3956
- dump_graph_stages: list[str] = field(default_factory=list)
3957
- """List of stages for which we want to dump the graph. Each pass defines
3958
- its own stages (before, after, maybe in-between)."""
3959
- dump_graph_dir: Path = Path(".")
3960
- """Directory to dump the graphs."""
3955
enable_fusion: bool = field(default_factory=lambda: not envs.VLLM_USE_V1)
3956
"""Whether to enable the custom fusion (RMSNorm/SiluMul+quant) pass."""
3957
enable_attn_fusion: bool = False
@@ -3975,12 +3969,9 @@ class PassConfig:
3969
"""
3970
Produces a hash unique to the pass configuration.
3971
Any new fields that affect compilation should be added to the hash.
3978
- Do not include dump_graph_* in the hash - they don't affect
3979
- compilation.
3972
+ Any future fields that don't affect compilation should be excluded.
3973
"""
3981
- exclude = {"dump_graph_stages", "dump_graph_dir"}
3982
- dict_ = {k: v for k, v in asdict(self).items() if k not in exclude}
3983
- return InductorPass.hash_dict(dict_)
3974
+ return InductorPass.hash_dict(asdict(self))
3975
3976
def __post_init__(self) -> None:
3977
if not self.enable_noop: