[compile] Split compile/warmup monitoring (#36098)

This commit is contained in:
Richard Zou
2026-03-07 13:52:11 -08:00
committed by GitHub
parent 379689d533
commit 2dde535df1
3 changed files with 123 additions and 48 deletions
+57 -20
View File
@@ -1,46 +1,83 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import contextlib
import time
from collections.abc import Generator
from vllm.config import CompilationConfig, CompilationMode, VllmConfig
from vllm.config import CompilationMode, VllmConfig
from vllm.logger import init_logger
logger = init_logger(__name__)
context_manager = None
# Shared global so backends.py can read the start time for Dynamo timing.
torch_compile_start_time: float = 0.0
def start_monitoring_torch_compile(vllm_config: VllmConfig) -> None:
@contextlib.contextmanager
def monitor_torch_compile(
vllm_config: VllmConfig,
message: str = "torch.compile took %.2f s in total",
) -> Generator[None, None, None]:
"""Context manager that times torch.compile and manages depyf debugging.
On normal exit: logs the compile time and exits depyf.
On exception: cleans up depyf without logging (compilation failed).
"""
global torch_compile_start_time
torch_compile_start_time = time.perf_counter()
compilation_config: CompilationConfig = vllm_config.compilation_config
compilation_config = vllm_config.compilation_config
depyf_cm = None
path = vllm_config.compile_debug_dump_path()
if compilation_config.mode == CompilationMode.VLLM_COMPILE and path:
import depyf
path.mkdir(parents=True, exist_ok=True)
logger.debug("Dumping depyf output to %s", path)
global context_manager
context_manager = depyf.prepare_debug(path.as_posix())
context_manager.__enter__()
depyf_cm = depyf.prepare_debug(path.as_posix())
depyf_cm.__enter__()
try:
yield
except Exception:
raise
else:
total_compile_time = time.perf_counter() - torch_compile_start_time
if compilation_config.mode == CompilationMode.VLLM_COMPILE:
logger.info_once(message, total_compile_time, scope="local")
finally:
if depyf_cm is not None:
try:
depyf_cm.__exit__(None, None, None)
except Exception:
logger.warning("Exception during depyf cleanup.", exc_info=True)
def end_monitoring_torch_compile(vllm_config: VllmConfig) -> None:
compilation_config: CompilationConfig = vllm_config.compilation_config
total_compile_time: float = time.perf_counter() - torch_compile_start_time
if compilation_config.mode == CompilationMode.VLLM_COMPILE:
logger.info_once(
"torch.compile and initial profiling run took %.2f s in total",
total_compile_time,
scope="local",
)
global context_manager
if context_manager is not None:
context_manager.__exit__(None, None, None)
context_manager = None
@contextlib.contextmanager
def monitor_profiling_run() -> Generator[None, None, None]:
"""Context manager that times the initial profiling run.
Asserts that no backend compilation occurs during the profiling run
(all compilation should have completed before this point).
"""
from vllm.compilation.counter import compilation_counter
backend_compilations_before = compilation_counter.num_backend_compilations
start = time.perf_counter()
yield
elapsed = time.perf_counter() - start
assert (
compilation_counter.num_backend_compilations == backend_compilations_before
), (
"backend compilation occurred during the initial profiling run; "
"all compilation should be complete before the profiling run starts."
)
logger.info_once(
"Initial profiling/warmup run took %.2f s",
elapsed,
scope="local",
)
cudagraph_capturing_enabled: bool = True