`nemo_export.multimodal.build`#

Module Contents#

`build_trtllm_engine`	Build TRTLLM engine by nemo export.
`build_mllama_trtllm_engine`	Build mllama TRTLLM engine from HF.
`export_visual_wrapper_onnx`	Export visual wrapper to ONNX.
`build_trt_engine`	Build TRT engine from onnx.
`build_neva_engine`
`build_video_neva_engine`	Build video neva visual engine.
`build_mllama_visual_engine`	Build mllama visual engine.
`build_visual_engine`	Build visual engine.
`extract_lora_ckpt`	Extrace lora from checkpoint.
`build_mllama_engine`	Build mllama engine.

nemo_export.multimodal.build.build_trtllm_engine( model_dir: str, visual_checkpoint_path: str, llm_checkpoint_path: str = None, model_type: str = 'neva', llm_model_type: str = 'llama', tensor_parallelism_size: int = 1, max_input_len: int = 256, max_output_len: int = 256, max_batch_size: int = 1, max_multimodal_len: int = 1024, dtype: str = 'bfloat16', use_lora_plugin: str = None, lora_target_modules: List[str] = None, max_lora_rank: int = 64, lora_ckpt_list: List[str] = None, )#: Build TRTLLM engine by nemo export.

nemo_export.multimodal.build.build_mllama_trtllm_engine( model_dir: str, hf_model_path: str, tensor_parallelism_size: int = 1, max_input_len: int = 256, max_output_len: int = 256, max_batch_size: int = 1, max_multimodal_len: int = 1024, dtype: str = 'bfloat16', use_lora_plugin: str = None, lora_target_modules: List[str] = None, max_lora_rank: int = 64, lora_ckpt_list: List[str] = None, )#: Build mllama TRTLLM engine from HF.

nemo_export.multimodal.build.export_visual_wrapper_onnx( visual_wrapper, input, output_dir, input_names=['input'], dynamic_axes={'input': {0: 'batch'}}, )#: Export visual wrapper to ONNX.

nemo_export.multimodal.build.build_trt_engine( model_type, input_sizes, output_dir, vision_max_batch_size, dtype=torch.bfloat16, image_size=None, num_frames=None, nemo_config=None, part_name='visual_encoder', )#: Build TRT engine from onnx.

nemo_export.multimodal.build.build_neva_engine( model_type: str, model_dir: str, visual_checkpoint_path: str, vision_max_batch_size: int = 1, )#

nemo_export.multimodal.build.build_video_neva_engine( model_dir: str, visual_checkpoint_path: str, vision_max_batch_size: int = 1, )#: Build video neva visual engine.

nemo_export.multimodal.build.build_mllama_visual_engine( model_dir: str, hf_model_path: str, processor_name: str = 'meta-llama/Llama-3.2-11B-Vision-Instruct', vision_max_batch_size: int = 1, )#: Build mllama visual engine.

nemo_export.multimodal.build.build_visual_engine( model_dir: str, visual_checkpoint_path: str, model_type: str = 'neva', vision_max_batch_size: int = 1, )#: Build visual engine.

nemo_export.multimodal.build.extract_lora_ckpt(lora_ckpt: str, output_dir: str)#: Extrace lora from checkpoint.

nemo_export.multimodal.build.build_mllama_engine( model_dir: str, checkpoint_path: str, processor_name: str = 'meta-llama/Llama-3.2-11B-Vision-Instruct', vision_max_batch_size: int = 1, tensor_parallelism_size: int = 1, max_input_len: int = 256, max_output_len: int = 256, max_batch_size: int = 1, max_multimodal_len: int = 1024, dtype: str = 'bfloat16', )#: Build mllama engine.