MCPcopy Create free account
hub / github.com/NVIDIA/TensorRT-LLM / build_task

Method build_task

tensorrt_llm/llmapi/llm_utils.py:775–801  ·  view source on GitHub ↗
(engine_dir: Path)

Source from the content-addressed store, hash-verified

773 model_format = self.llm_args.model_format
774
775 def build_task(engine_dir: Path):
776 if model_format is not _ModelFormatKind.TLLM_ENGINE:
777 model_loader_kwargs = {
778 'llm_args': self.llm_args,
779 'workspace': str(self.workspace),
780 'llm_build_stats': self.llm_build_stats,
781 }
782
783 if self.llm_args.parallel_config.is_multi_gpu:
784 assert self.mpi_session
785
786 #mpi_session cannot be pickled so remove from self.llm_args
787 if self.llm_args.mpi_session:
788 del self.llm_args.mpi_session
789
790 # The engine_dir:Path will be stored to MPINodeState.state
791 build_infos = self.mpi_session.submit_sync(
792 CachedModelLoader._node_build_task,
793 engine_dir=engine_dir,
794 **model_loader_kwargs)
795 self.llm_build_stats.build_steps_info = build_infos[0]
796
797 else: # single-gpu
798 with ModelLoader(**model_loader_kwargs) as model_loader:
799 model_loader(engine_dir=engine_dir)
800
801 release_gc()
802
803 has_storage = True
804 if self.build_cache_enabled:

Callers

nothing calls this directly

Calls 3

release_gcFunction · 0.85
ModelLoaderClass · 0.70
submit_syncMethod · 0.45

Tested by

no test coverage detected