[Misc][BE] Type coverage for vllm/compilation [2/3] (#31744)

This commit is contained in:
Lucas Kabela
2026-01-09 18:30:38 -05:00
committed by GitHub
parent 3adffd5b90
commit aaf4b70aae
12 changed files with 161 additions and 91 deletions
+19 -12
View File
@@ -1,6 +1,8 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import functools
from collections.abc import Callable
from typing import Any, ParamSpec, TypeVar
from torch import fx as fx
@@ -40,8 +42,11 @@ from .noop_elimination import NoOpEliminationPass
logger = init_logger(__name__)
P = ParamSpec("P")
R = TypeVar("R")
def with_pattern_match_debug(fn):
def with_pattern_match_debug(fn: Callable[P, R]) -> Callable[P, R]:
"""
Function decorator that turns on inductor pattern match debug
for the duration of the call.
@@ -49,7 +54,7 @@ def with_pattern_match_debug(fn):
"""
@functools.wraps(fn)
def wrapper(*args, **kwargs):
def wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
if (debug_val := envs.VLLM_PATTERN_MATCH_DEBUG) is not None:
# optionally check rank here
with set_env_var("TORCHINDUCTOR_PATTERN_MATCH_DEBUG", debug_val):
@@ -59,7 +64,7 @@ def with_pattern_match_debug(fn):
return wrapper
class PostGradPassManager(CustomGraphPass):
class PostGradPassManager(CustomGraphPass): # type: ignore[misc]
"""
The pass manager for post-grad passes.
It handles configuration, adding custom passes, and running passes.
@@ -74,11 +79,11 @@ class PostGradPassManager(CustomGraphPass):
This way, all passes operate on a functionalized graph.
"""
def __init__(self):
def __init__(self) -> None:
self.passes: list[InductorPass] = []
@with_pattern_match_debug
def __call__(self, graph: fx.Graph):
def __call__(self, graph: fx.Graph) -> None:
VllmInductorPass.dump_prefix = 0 # reset dump index
compile_range = get_pass_context().compile_range
@@ -98,7 +103,7 @@ class PostGradPassManager(CustomGraphPass):
self.fix_functionalization(graph)
VllmInductorPass.dump_prefix = None # Cleanup index
def configure(self, config: VllmConfig):
def configure(self, config: VllmConfig) -> None:
self.pass_config = config.compilation_config.pass_config
# Set the current vllm config to allow tracing CustomOp instances
@@ -135,23 +140,25 @@ class PostGradPassManager(CustomGraphPass):
self.post_cleanup = PostCleanupPass(config)
self.fix_functionalization = FixFunctionalizationPass(config)
def add(self, pass_: InductorPass):
def add(self, pass_: InductorPass) -> None:
assert isinstance(pass_, InductorPass)
self.passes.append(pass_)
def uuid(self):
def uuid(self) -> str:
"""
The PostGradPassManager is set as a custom pass in the Inductor and
affects compilation caching. Its uuid depends on the UUIDs of all
dependent passes and the pass config. See InductorPass for more info.
"""
state = {"pass_config": self.pass_config.compute_hash(), "passes": []}
passes = []
state: dict[str, Any] = {"pass_config": self.pass_config.compute_hash()}
for pass_ in self.passes:
state["passes"].append(pass_.uuid())
state["passes"].append(self.fix_functionalization.uuid())
passes.append(pass_.uuid())
passes.append(self.fix_functionalization.uuid())
# Include the compile range in the uuid to ensure that inductor
# recompiles the graph for the new dynamic compile range.
state["compile_range"] = str(get_pass_context().compile_range)
state["passes"] = passes
return InductorPass.hash_dict(state)