modeling_llava_phi3.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.


from typing import List, Optional, Tuple, Union

import torch
import torch.nn as nn
import math 
import sys
import pdb
from typing import Dict, Any

from transformers import AutoConfig, AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
                        #  MistralConfig, MistralModel, MistralForCausalLM
                         

from transformers.modeling_outputs import CausalLMOutputWithPast


from transformers.cache_utils import Cache, DynamicCache


from .llava_arch import LlavaMetaModel, LlavaMetaForCausalLM
from .modeling_phi3 import Phi3ForCausalLM, Phi3Model, Phi3Config
from .generation_utils import build_allava_input




################ Phi ###############################

class LlavaPhi3Config(Phi3Config):
    model_type = "llava_phi3"

class LlavaPhi3Model(LlavaMetaModel, Phi3Model):
    config_class = LlavaPhi3Config

    def __init__(self, config: Phi3Config):
        super(LlavaPhi3Model, self).__init__(config)



class LlavaPhi3ForCausalLM(Phi3ForCausalLM, LlavaMetaForCausalLM):
    config_class = LlavaPhi3Config

    def __init__(self, config, init_vision_encoder_from_ckpt=True):
        config.flash_attn = True
        config.flash_rotary = True
        config.fused_dense = True
        config._attn_implementation = "flash_attention_2"

        super(Phi3ForCausalLM, self).__init__(config)
        # self.model is used in LlavaMetaForCausalLM.get_model(); self.transformer is used in PhiForCausalLM.forward()
        self.model = LlavaPhi3Model(config)
        # self.model.embd = 
        if hasattr(self.model, '_use_flash_attention_2'):
            assert self.model._use_flash_attention_2, 'flash attn is not enabled. check it out!'
        # self.pretraining_tp = config.pretraining_tp
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        
        if init_vision_encoder_from_ckpt:
            vision_tower = self.get_vision_tower()
            print(f'loading from CLIP first. This should only be used at inference!!!')
            vision_tower.load_model() # 
            
        # Initialize weights and apply final processing
        self.post_init()

    # ############ these two methods are missing in modeling_phi.py
    # def get_input_embeddings(self) -> nn.Embedding:
    #     return self.model.embd.wte

    # def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
    #     self.model.embd.wte = new_embeddings
    # ############ these two methods are missing in modeling_phi.py

    def get_model(self):
        return self.model
    
    def get_tokenizer(self):
        return self.tokenizer

    def get_processor(self):
        return self.model.vision_tower.image_processor

    def set_tokenizer_eos_id(self):
        eos_token_id = 30027 # only for llava_phi3
        self.tokenizer.eos_token_id = eos_token_id


    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

        # pdb.set_trace()
        if inputs_embeds is None:
            (
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                inputs_embeds,
                labels
            # ) = self.prepare_inputs_labels_for_multimodal(
            ) = self.prepare_inputs_labels_for_multimodal_new(
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                labels,
                images
            )
        

        return super().forward(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            labels=labels,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict
        )

    @torch.no_grad()
    def generate(
        self,
        inputs: Optional[torch.Tensor] = None,
        images: Optional[torch.Tensor] = None,
        **kwargs,
    ) :
        position_ids = kwargs.pop("position_ids", None)
        attention_mask = kwargs.pop("attention_mask", None)
        if "inputs_embeds" in kwargs:
            raise NotImplementedError("`inputs_embeds` is not supported")

        if images is not None:
            (
                inputs,
                position_ids,
                attention_mask,
                _,
                inputs_embeds,
                _
            ) = self.prepare_inputs_labels_for_multimodal_new(
                inputs,
                position_ids,
                attention_mask,
                None,
                None,
                images
            )
        else:
            inputs_embeds = self.get_model().embed_tokens(inputs)

        # print(inputs_embeds.shape)
        return super().generate(
            position_ids=None,
            attention_mask=None,
            inputs_embeds=inputs_embeds,
            **kwargs
        )
    
    
    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs):
        '''
        This function is called for each token at inference
        '''
        # pdb.set_trace()
        images = kwargs.pop("images", None)

        ####################################################
        # lines from modeling_phi.py
        ####################################################

        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                cache_length = past_key_values.get_seq_length()
                past_length = past_key_values.seen_tokens
                max_cache_length = past_key_values.get_max_length()
            else:
                cache_length = past_length = past_key_values[0][0].shape[2]
                max_cache_length = None

            # Keep only the unprocessed tokens:
            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
            # input)
            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
            # input_ids based on the past_length.
            elif past_length < input_ids.shape[1]:
                input_ids = input_ids[:, past_length:]
            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
            elif past_length >= input_ids.shape[1]:
                input_ids = input_ids[:, [-1]] # only keep the last one!

            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
            if (
                max_cache_length is not None
                and attention_mask is not None
                and cache_length + input_ids.shape[1] > max_cache_length
            ):
                attention_mask = attention_mask[:, -max_cache_length:]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}
        
        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
            }
        )
        ####################################################
        # end of lines from modeling_phi.py
        ####################################################


        if images is not None:
            model_inputs['images'] = images
        return model_inputs


    # def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
    #     images = kwargs.pop("images", None)
    #     _inputs = super().prepare_inputs_for_generation(
    #         input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
    #     )
    #     if images is not None:
    #         _inputs['images'] = images
    #     return _inputs

    def chat(
        self, 
        texts: Optional[str | list[list[str, str]]], 
        images: Optional[str | list[str]] = None, 
        history: Optional[list[str]] = None, 
        stream = False, 
        return_history = False, 
        **kwargs
    ):
        '''
        texts: if `str`, then generate for a single round; if list[dict], 
        images: str (optional), local path to an image.
        '''
        use_cache = kwargs.pop('use_cache', True)
        
        if 'eos_token_id' in kwargs:
            _ = kwargs.pop('eos_token_id', None)
            print(f'eos_token_id {_} from gen_kwargs is popped since it is not needed.')
        # pdb.set_trace()


        ############################
        # merge history
        ############################            
        input_ids, image_tensors, history = build_allava_input(
            tokenizer = self.get_tokenizer(), 
            processor = self.get_processor(), 
            texts = texts, 
            images = images, 
            history=history,
            return_history=return_history,
            device = self.device
        )

        ############################
        # generate response
        ############################
        # with torch.autocast(device_type='cuda'):
        if 'cuda' in str(self.device):
            device_type = 'cuda'
        else:
            device_type = 'cpu'

        with torch.autocast(device_type=device_type, dtype=self.dtype):
            output_ids = self.generate(
                inputs=input_ids,
                images=image_tensors,
                use_cache=use_cache,
                **kwargs)

        answer = self.get_tokenizer().decode(output_ids[0, :], skip_special_tokens=True).strip()

        if return_history:
            history[-1][-1] = answer
            return answer, history
        return answer


AutoConfig.register("llava_phi3", LlavaPhi3Config)
AutoModelForCausalLM.register(LlavaPhi3Config, LlavaPhi3ForCausalLM)