llava/model/builder.py

#    Copyright 2023 Haotian Liu
#
#    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.

import os
import warnings
import shutil

from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig
import torch
from llava.model import *
from llava.constants import DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN




def map_keys(model, pretrained_ckpt_loc):
    
    ckpt = torch.load(pretrained_ckpt_loc, map_location='cpu')
    print(ckpt.keys())
    print(ckpt['proj'].size())
    i = 0
    for name, param in model.named_parameters():
        # print(ckpt.keys())
        i+=1
        print(name, param.size())
        # if param.requires_grad:
    print(i)
    exit()
    with torch.no_grad():
        for i in range(4):
            for p in range(2):
                self.downsample_layers[i][p].weight.copy_(ckpt[f'downsample_layers.{i}.{p}.weight'])
                self.downsample_layers[i][p].bias.copy_(ckpt[f'downsample_layers.{i}.{p}.bias'])
       
        
        for j in range(4):
            for k in range(stt[j]):
                self.stages[j][k].gamma.copy_(ckpt[f'stages.{j}.{k}.gamma'])
                self.stages[j][k].dwconv.weight.copy_(ckpt[f'stages.{j}.{k}.dwconv.weight'])
                self.stages[j][k].dwconv.bias.copy_(ckpt[f'stages.{j}.{k}.dwconv.bias'])
                self.stages[j][k].norm.weight.copy_(ckpt[f'stages.{j}.{k}.norm.weight'])
                self.stages[j][k].norm.bias.copy_(ckpt[f'stages.{j}.{k}.norm.bias'])
                self.stages[j][k].pwconv1.weight.copy_(ckpt[f'stages.{j}.{k}.pwconv1.weight'])
                self.stages[j][k].pwconv1.bias.copy_(ckpt[f'stages.{j}.{k}.pwconv1.bias'])
                self.stages[j][k].pwconv2.weight.copy_(ckpt[f'stages.{j}.{k}.pwconv2.weight'])
                self.stages[j][k].pwconv2.bias.copy_(ckpt[f'stages.{j}.{k}.pwconv2.bias'])


class ClipVisionModel(torch.nn.Module):
    def __init__(self, model, normalize, all_tokens=False, proj=True):
        super().__init__()
        self.model = model
        self.normalize = normalize
        self.proj = model.proj
        if all_tokens:
            self.model.output_tokens = True
        if not proj:
            self.model.proj = None

    def forward(self, vision_, output_normalize):
        embedding = self.model(self.normalize(vision_))
        if output_normalize:
            embedding = F.normalize(embedding, dim=-1)

        if self.model.output_tokens:
            # flatten and concatenate all tokens
            return torch.hstack([embedding[0].flatten(1), embedding[1].flatten(1)])
        else:
            return embedding



def load_pretrained_model(model_path, model_base, model_name, pretrained_rob_path=None, dtype=None, device_map="auto", device="cuda"):
    kwargs = {"device_map": device_map}
    load_8bit=False
    load_4bit=False
    if load_8bit:
        kwargs['load_in_8bit'] = True
    elif load_4bit:
        kwargs['load_in_4bit'] = True
        kwargs['quantization_config'] = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_compute_dtype=torch.float16,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type='nf4'
        )
    else:
        assert dtype is not None
        if dtype == 'float16':
            kwargs['torch_dtype'] = torch.float16
        elif dtype == 'float32':
            kwargs['torch_dtype'] = torch.float32
        else:
            raise ValueError(f"Unknown dtype {dtype}, must be float16 or float32")

    if 'llava' in model_name.lower():
        # Load LLaVA model
        if 'lora' in model_name.lower() and model_base is None:
            warnings.warn('There is `lora` in model name but no `model_base` is provided. If you are loading a LoRA model, please provide the `model_base` argument. Detailed instruction: https://github.com/haotian-liu/LLaVA#launch-a-model-worker-lora-weights-unmerged.')
        if 'lora' in model_name.lower() and model_base is not None:
            lora_cfg_pretrained = AutoConfig.from_pretrained(model_path)
            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
            print('Loading LLaVA from base model...')
            model = LlavaLlamaForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=lora_cfg_pretrained, **kwargs)
            token_num, tokem_dim = model.lm_head.out_features, model.lm_head.in_features
            if model.lm_head.weight.shape[0] != token_num:
                model.lm_head.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))
                model.model.embed_tokens.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))

            print('Loading additional LLaVA weights...')
            if os.path.exists(os.path.join(model_path, 'non_lora_trainables.bin')):
                non_lora_trainables = torch.load(os.path.join(model_path, 'non_lora_trainables.bin'), map_location='cpu')
            else:
                # this is probably from HF Hub
                from huggingface_hub import hf_hub_download
                def load_from_hf(repo_id, filename, subfolder=None):
                    cache_file = hf_hub_download(
                        repo_id=repo_id,
                        filename=filename,
                        subfolder=subfolder)
                    return torch.load(cache_file, map_location='cpu')
                non_lora_trainables = load_from_hf(model_path, 'non_lora_trainables.bin')
            non_lora_trainables = {(k[11:] if k.startswith('base_model.') else k): v for k, v in non_lora_trainables.items()}
            if any(k.startswith('model.model.') for k in non_lora_trainables):
                non_lora_trainables = {(k[6:] if k.startswith('model.') else k): v for k, v in non_lora_trainables.items()}
            model.load_state_dict(non_lora_trainables, strict=False)

            from peft import PeftModel
            print('Loading LoRA weights...')
            model = PeftModel.from_pretrained(model, model_path)
            print('Merging LoRA weights...')
            model = model.merge_and_unload()
            print('Model is loaded...')
        elif model_base is not None:
            # this may be mm projector only
            print('Loading LLaVA from base model...')
            if 'mpt' in model_name.lower():
                if not os.path.isfile(os.path.join(model_path, 'configuration_mpt.py')):
                    shutil.copyfile(os.path.join(model_base, 'configuration_mpt.py'), os.path.join(model_path, 'configuration_mpt.py'))
                tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=True)
                cfg_pretrained = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
                model = LlavaMPTForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs)
            else:
                tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
                cfg_pretrained = AutoConfig.from_pretrained(model_path)
                model = LlavaLlamaForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs)

            mm_projector_weights = torch.load(os.path.join(model_path, 'mm_projector.bin'), map_location='cpu')
            mm_projector_weights = {k: v.to(kwargs["torch_dtype"]) for k, v in mm_projector_weights.items()}
            model.load_state_dict(mm_projector_weights, strict=False)
        else:
            if 'mpt' in model_name.lower():
                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
                model = LlavaMPTForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
            else:
                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
                model = LlavaLlamaForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
    else:
        # Load language model
        if model_base is not None:
            # PEFT model
            from peft import PeftModel
            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
            model = AutoModelForCausalLM.from_pretrained(model_base, torch_dtype=kwargs["torch_dtype"], low_cpu_mem_usage=True, device_map="auto")
            print(f"Loading LoRA weights from {model_path}")
            model = PeftModel.from_pretrained(model, model_path)
            print(f"Merging weights")
            model = model.merge_and_unload()
            if kwargs["torch_dtype"] == torch.float16:
                print('Convert to FP16...')
                model.to(torch.float16)
        else:
            use_fast = False
            if 'mpt' in model_name.lower():
                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
                model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, trust_remote_code=True, **kwargs)
            else:
                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
                model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)

    image_processor = None

    if 'llava' in model_name.lower():
        mm_use_im_start_end = getattr(model.config, "mm_use_im_start_end", False)
        mm_use_im_patch_token = getattr(model.config, "mm_use_im_patch_token", True)
        if mm_use_im_patch_token:
            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
        if mm_use_im_start_end:
            tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
        model.resize_token_embeddings(len(tokenizer))

        vision_tower = model.get_vision_tower()
        # vision_tower.set_device(device)
        non_llava = True if pretrained_rob_path not in [None, 'None', 'none'] else False
        if not vision_tower.is_loaded:
            vision_tower.load_model(non_llava, pretrained_rob_path)#.to(device=device)

        # print(vision_tower.vision_tower)
        vision_tower.to(device=device, dtype=kwargs["torch_dtype"])
        image_processor = vision_tower.image_processor

    if hasattr(model.config, "max_sequence_length"):
        context_len = model.config.max_sequence_length
    else:
        context_len = 2048

    return model, image_processor, tokenizer, context_len