modeling_llamavision.py

import torch
import torch.nn as nn
from transformers import (
    PreTrainedModel,
    AutoModelForCausalLM,
    AutoModel,
    SiglipImageProcessor,
)
from .configuration_llamavision import LlamavisionConfig


class ProjectionModule(nn.Module):
    def __init__(self, mm_hidden_size=1152, hidden_size=4096):
        super(ProjectionModule, self).__init__()

        # Directly set up the sequential model
        self.model = nn.Sequential(
            nn.Linear(mm_hidden_size, hidden_size),
            nn.GELU(),
            nn.Linear(hidden_size, hidden_size),
        )

    def forward(self, x):
        return self.model(x)


class Llamavision(PreTrainedModel):
    config_class = LlamavisionConfig

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

        self.text_model = AutoModelForCausalLM.from_config(config.text_config)
        self.vision_model = AutoModel.from_config(config.vision_config)
        self.processor = SiglipImageProcessor()
        self.mm_projector = ProjectionModule()

    @property
    def device(self):
        return self.text_model.device

    def tokenizer_image_token(
        self, prompt, tokenizer, image_token_index=-200, return_tensors=None
    ):
        prompt_chunks = [
            tokenizer(chunk).input_ids for chunk in prompt.split("<image>")
        ]

        def insert_separator(X, sep):
            return [ele for sublist in zip(X, [sep] * len(X)) for ele in sublist][:-1]

        input_ids = []
        offset = 0
        if (
            len(prompt_chunks) > 0
            and len(prompt_chunks[0]) > 0
            and prompt_chunks[0][0] == tokenizer.bos_token_id
        ):
            offset = 1
            input_ids.append(prompt_chunks[0][0])

        for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
            input_ids.extend(x[offset:])

        return torch.tensor(input_ids, dtype=torch.long)

    def process_tensors(self, input_ids, image_features, embedding_layer):
        # Find the index of -200 in input_ids
        split_index = (input_ids == -200).nonzero(as_tuple=True)[1][0]

        # Split the input_ids at the index found, excluding -200
        input_ids_1 = input_ids[:, :split_index]
        input_ids_2 = input_ids[:, split_index + 1 :]

        # Convert input_ids to embeddings
        embeddings_1 = embedding_layer(input_ids_1)
        embeddings_2 = embedding_layer(input_ids_2)

        device = image_features.device
        token_embeddings_part1 = embeddings_1.to(device)
        token_embeddings_part2 = embeddings_2.to(device)

        # Concatenate the token embeddings and image features
        concatenated_embeddings = torch.cat(
            [token_embeddings_part1, image_features, token_embeddings_part2], dim=1
        )

        # Create the corrected attention mask
        attention_mask = torch.ones(
            concatenated_embeddings.shape[:2], dtype=torch.long, device=device
        )
        return concatenated_embeddings, attention_mask

    def answer_question(self, image, question, tokenizer, **kwargs):
        question = "<image>" + question

        prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{question}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"

        input_ids = (
            self.tokenizer_image_token(prompt, tokenizer, -200, return_tensors="pt")
            .unsqueeze(0)
            .to(self.device)
        )
        terminators = [
            tokenizer.eos_token_id,
            tokenizer.convert_tokens_to_ids("<|eot_id|>"),
        ]
        with torch.inference_mode():
            image_inputs = self.processor(
                images=[image],
                return_tensors="pt",
                do_resize=True,
                size={"height": 384, "width": 384},
            )

            image_inputs = image_inputs["pixel_values"].to(
                device=self.device, dtype=self.dtype
            )

            image_forward_outs = self.vision_model(
                image_inputs,
                output_hidden_states=True,
            )

            image_features = image_forward_outs.hidden_states[-2]

            projected_embeddings = self.mm_projector(image_features).to(self.device)

            embedding_layer = self.text_model.get_input_embeddings()
            # text_embeddings = embedding_layer(input_ids)

            new_embeds, attn_mask = self.process_tensors(
                input_ids, projected_embeddings, embedding_layer
            )

            attn_mask = attn_mask.to(self.device)
            new_embeds = new_embeds.to(self.device)
            answer = self.text_model.generate(
                inputs_embeds=new_embeds,
                attention_mask=attn_mask,
                eos_token_id=terminators,
                temperature=0.2,
                do_sample=True,
                **kwargs,
            )[0]

            return answer