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
+27 -5
View File
@@ -282,11 +282,33 @@ class EngineCore:
self.model_executor.initialize_from_config(kv_cache_configs)
elapsed = time.time() - start
logger.info_once(
"init engine (profile, create kv cache, warmup model) took %.2f seconds",
elapsed,
scope="local",
)
compile_time = vllm_config.compilation_config.compilation_time
encoder_compile_time = vllm_config.compilation_config.encoder_compilation_time
if encoder_compile_time > 0:
logger.info_once(
"init engine (profile, create kv cache, warmup model) took "
"%.2f s (compilation: %.2f s — language_model: %.2f s, "
"encoder: %.2f s)",
elapsed,
compile_time + encoder_compile_time,
compile_time,
encoder_compile_time,
scope="local",
)
elif compile_time > 0:
logger.info_once(
"init engine (profile, create kv cache, warmup model) took "
"%.2f s (compilation: %.2f s)",
elapsed,
compile_time,
scope="local",
)
else:
logger.info_once(
"init engine (profile, create kv cache, warmup model) took %.2f s",
elapsed,
scope="local",
)
return scheduler_kv_cache_config
def get_supported_tasks(self) -> tuple[SupportedTask, ...]: