llava/model/multimodal_projector/builder.py

# Copyright (c) 2025 NVIDIA CORPORATION.
# Licensed under the MIT license.

# Adapted from https://github.com/NVlabs/VILA/tree/main under the Apache 2.0 license.
# LICENSE is in incl_licenses directory.

# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0

# This file is modified from https://github.com/haotian-liu/LLaVA/

import os

import torch
from transformers import PretrainedConfig, PreTrainedModel

from .base_projector import MultimodalProjector, MultimodalProjectorConfig
from .speech_base_projector import SpeechMultimodalProjector, SpeechMultimodalProjectorConfig
from .sound_base_projector import SoundMultimodalProjector, SoundMultimodalProjectorConfig


def build_speech_mm_projector(model_type_or_path: str, config: PretrainedConfig) -> PreTrainedModel:
    if model_type_or_path is None:
        return None

    ## load from pretrained model
    if config.resume_path:
        assert os.path.exists(model_type_or_path), f"Resume speech mm projector path {model_type_or_path} does not exist!"
        return SpeechMultimodalProjector.from_pretrained(model_type_or_path, config, torch_dtype=eval(config.model_dtype))
    ## build from scratch
    else:
        print("WARNING: Building speech multimodal projector from scratch!")
        speech_mm_projector_cfg = SpeechMultimodalProjectorConfig(model_type_or_path)
        speech_mm_projector = SpeechMultimodalProjector(speech_mm_projector_cfg, config).to(eval(config.model_dtype))
    return speech_mm_projector

def build_sound_mm_projector(model_type_or_path: str, config: PretrainedConfig) -> PreTrainedModel:
    if model_type_or_path is None:
        return None

    ## load from pretrained model
    if config.resume_path:
        print(config.resume_path)
        assert os.path.exists(model_type_or_path), f"Resume sound mm projector path {model_type_or_path} does not exist!"
        return SoundMultimodalProjector.from_pretrained(model_type_or_path, config, torch_dtype=eval(config.model_dtype))
    # build from scratch
    else:
        print("WARNING: Building sound multimodal projector from scratch!")
        sound_mm_projector_cfg = SoundMultimodalProjectorConfig(model_type_or_path)
        sound_mm_projector = SoundMultimodalProjector(sound_mm_projector_cfg, config).to(eval(config.model_dtype))
    return sound_mm_projector