Measure encoder compile time seperate from llm backbone (#39240)

Signed-off-by: Lucas Kabela <lucaskabela@meta.com>
This commit is contained in:
Lucas Kabela
2026-04-14 10:52:49 -04:00
committed by GitHub
parent f02b3269e7
commit c08f3b2a62
9 changed files with 186 additions and 147 deletions
+6 -3
View File
@@ -56,7 +56,7 @@ from vllm.v1.outputs import (
)
from vllm.v1.utils import compute_iteration_details, report_usage_stats
from vllm.v1.worker.utils import is_residual_scattered_for_sp
from vllm.v1.worker.worker_base import WorkerBase
from vllm.v1.worker.worker_base import CompilationTimes, WorkerBase
from vllm.v1.worker.workspace import init_workspace_manager
from ...model_executor.model_loader import TensorizerLoader
@@ -547,7 +547,7 @@ class Worker(WorkerBase):
self.model_runner._init_kv_zero_meta()
@instrument(span_name="Warmup (GPU)")
def compile_or_warm_up_model(self) -> float:
def compile_or_warm_up_model(self) -> CompilationTimes:
warmup_sizes: list[int] = []
if self.vllm_config.compilation_config.mode == CompilationMode.VLLM_COMPILE:
@@ -689,7 +689,10 @@ class Worker(WorkerBase):
# the model initialization and profiling.
set_random_seed(self.model_config.seed)
return self.compilation_config.compilation_time
return CompilationTimes(
language_model=self.compilation_config.compilation_time,
encoder=self.compilation_config.encoder_compilation_time,
)
def reset_mm_cache(self) -> None:
self.model_runner.reset_mm_cache()