[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
+41 -27
View File
@@ -30,7 +30,7 @@ from vllm.sequence import IntermediateTensors
from vllm.utils.import_utils import resolve_obj_by_qualname
from vllm.utils.torch_utils import is_torch_equal_or_newer
from .monitor import start_monitoring_torch_compile
from .monitor import monitor_profiling_run, monitor_torch_compile
if TYPE_CHECKING:
# Only added on nightly/2.10 so wrap
@@ -434,17 +434,24 @@ def _support_torch_compile(
cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}")
aot_compilation_path = os.path.join(cache_dir, "model")
try:
with (
set_current_vllm_config(self.vllm_config),
open(aot_compilation_path, "rb") as f,
):
start_monitoring_torch_compile(self.vllm_config)
loaded_fn = torch.compiler.load_compiled_function(
f, f_globals=self.forward.__globals__
)
_verify_source_unchanged(loaded_fn.source_info(), self.vllm_config)
if not self.compilation_config.dynamic_shapes_config.evaluate_guards:
loaded_fn.disable_guard_check()
with monitor_torch_compile(self.vllm_config):
with (
set_current_vllm_config(self.vllm_config),
open(aot_compilation_path, "rb") as f,
):
loaded_fn = torch.compiler.load_compiled_function(
f, f_globals=self.forward.__globals__
)
_verify_source_unchanged(loaded_fn.source_info(), self.vllm_config)
ds_config = self.compilation_config.dynamic_shapes_config
if not ds_config.evaluate_guards:
loaded_fn.disable_guard_check()
# Eagerly load compiled artifacts now that traced_files
# is populated by _verify_source_unchanged.
with maybe_use_cudagraph_partition_wrapper(self.vllm_config):
loaded_fn._artifacts.compiled_fn.finalize_loading(
self.vllm_config
)
self.aot_compiled_fn = loaded_fn
self.was_aot_compile_fn_loaded_from_disk = True
except Exception as e:
@@ -465,12 +472,11 @@ def _support_torch_compile(
logger.info(
"Directly load AOT compilation from path %s", aot_compilation_path
)
# Apply partition wrapper context for proper CUDA graph capture
from .monitor import end_monitoring_torch_compile
with maybe_use_cudagraph_partition_wrapper(self.vllm_config):
with (
monitor_profiling_run(),
maybe_use_cudagraph_partition_wrapper(self.vllm_config),
):
output = self.aot_compiled_fn(self, *args, **kwargs)
end_monitoring_torch_compile(self.vllm_config)
return output
if self.compiled:
@@ -489,8 +495,6 @@ def _support_torch_compile(
**kwargs,
)
# here, it is the starting point of the `torch.compile` process
start_monitoring_torch_compile(self.vllm_config)
original_code_object = self.original_code_object()
logger.debug("Start compiling function %s", original_code_object)
@@ -559,16 +563,26 @@ def _support_torch_compile(
# store the path for saving after warmup
self._aot_compilation_path = aot_compilation_path
self._aot_cache_dir = cache_dir
self.aot_compiled_fn = self.aot_compile(*args, **kwargs)
# All compilation is done at this point, save the AOT artifact.
self.save_aot_compiled_function()
output = self.aot_compiled_fn(self, *args, **kwargs)
with monitor_torch_compile(self.vllm_config):
self.aot_compiled_fn = self.aot_compile(*args, **kwargs)
# All compilation is done at this point, save the
# AOT artifact.
self.save_aot_compiled_function()
with monitor_profiling_run():
output = self.aot_compiled_fn(self, *args, **kwargs)
else:
output = TorchCompileWithNoGuardsWrapper.__call__(self, *args, **kwargs) # type: ignore[arg-type]
with monitor_torch_compile(
self.vllm_config,
"torch.compile and initial profiling/warmup "
"run together took %.2f s in total",
):
output = TorchCompileWithNoGuardsWrapper.__call__(
self, # type: ignore[arg-type]
*args,
**kwargs,
)
from .monitor import end_monitoring_torch_compile
end_monitoring_torch_compile(self.vllm_config)
self.compiled = True
return output