app.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from safetensors import safe_open
import json
import gradio as gr
from PIL import Image
import numpy as np
from huggingface_hub import snapshot_download
from mistral_common.protocol.instruct.messages import UserMessage, TextChunk, ImageChunk
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
import spaces

# Download model files
model_path = snapshot_download(repo_id="mistral-community/pixtral-12b-240910")

# Load model parameters and tokenizer configuration
with open(f'{model_path}/params.json', 'r') as f:
    params = json.load(f)

with open(f'{model_path}/tekken.json', 'r') as f:
    tokenizer_config = json.load(f)

class GELU(nn.Module):
    def __init__(self, dim_in, dim_out, approximate='none', bias=True):
        super().__init__()
        self.linear = nn.Linear(dim_in, dim_out, bias=bias)
        self.approximate = approximate

    def forward(self, x):
        if self.approximate == 'tanh':
            return 0.5 * x * (1 + torch.tanh(np.sqrt(2 / np.pi) * (x + 0.044715 * torch.pow(x, 3))))
        else:
            return F.gelu(self.linear(x))

class Rope2D(nn.Module):
    def __init__(self, dim, max_position_embeddings=1024, base=10000):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)
        self.max_seq_len_cached = max_position_embeddings
        t = torch.arange(self.max_seq_len_cached, dtype=self.inv_freq.dtype)
        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)

    def forward(self, x, seq_len=None):
        if seq_len > self.max_seq_len_cached:
            self.max_seq_len_cached = seq_len
            t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
            self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
            self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
        return (
            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
        )

class VisionEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.embed = nn.Conv2d(config['num_channels'], config['hidden_size'], kernel_size=config['patch_size'], stride=config['patch_size'])
        self.rope = Rope2D(config['hidden_size'] // config['num_attention_heads'], base=config['rope_theta'])
        self.layers = nn.ModuleList([nn.TransformerEncoderLayer(d_model=config['hidden_size'], nhead=config['num_attention_heads'], dim_feedforward=config['intermediate_size']) for _ in range(config['num_hidden_layers'])])
        self.norm = nn.LayerNorm(config['hidden_size'])
        self.gelu = GELU(config['hidden_size'], config['hidden_size'])

    def forward(self, pixel_values):
        x = self.embed(pixel_values)
        b, c, h, w = x.shape
        x = x.flatten(2).transpose(1, 2)
        cos, sin = self.rope(x, seq_len=h*w)
        for layer in self.layers:
            x = layer(x)
        x = self.norm(x)
        x = self.gelu(x)
        return x

class PixtralModel(nn.Module):
    def __init__(self, params):
        super().__init__()
        self.vision_encoder = VisionEncoder(params['vision_encoder'])
        # Add text generation components here

    def forward(self, image):
        vision_output = self.vision_encoder(image)
        # Add text generation logic here
        return vision_output

def load_model(params, model_path):
    model = PixtralModel(params)
    
    with safe_open(f'{model_path}/consolidated.safetensors', framework="pt", device="cpu") as f:
        for name, param in model.named_parameters():
            if name in f.keys():
                param.data = f.get_tensor(name)
    
    model.eval()
    return model

# Initialize the model
model = load_model(params, model_path)

# Initialize the tokenizer
tokenizer = MistralTokenizer.from_model("pixtral")

@spaces.GPU
def process_image_and_text(image, prompt):
    # Prepare the image
    image = image.convert('RGB')
    image = image.resize((params['vision_encoder']['image_size'], params['vision_encoder']['image_size']))
    image_tensor = torch.tensor(np.array(image)).permute(2, 0, 1).unsqueeze(0).float() / 255.0
    image_tensor = image_tensor.cuda()

    # Tokenize the input
    tokenized = tokenizer.encode_chat_completion(
        ChatCompletionRequest(
            messages=[
                UserMessage(
                    content=[
                        TextChunk(text=prompt),
                        ImageChunk(image=image),
                    ]
                )
            ],
            model="pixtral",
        )
    )
    tokens, text, images = tokenized.tokens, tokenized.text, tokenized.images

    # Process the image and generate text
    with torch.no_grad():
        model.cuda()  # Move model to GPU only when processing
        vision_output = model(image_tensor)
        model.cpu()  # Move model back to CPU after processing
        # Add text generation logic here
        generated_text = f"Generated text based on the image and prompt: {prompt}"

    return generated_text, len(tokens), len(images)

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# Pixtral Image-to-Text Model Demo")
    gr.Markdown("Upload an image and provide a prompt to generate text based on it.")
    
    with gr.Row():
        with gr.Column(scale=1):
            input_image = gr.Image(type="pil")
            input_prompt = gr.Textbox(label="Prompt")
            submit_btn = gr.Button("Generate Text")
        
        with gr.Column(scale=1):
            output_text = gr.Textbox(label="Generated Text")
            token_count = gr.Number(label="Number of Tokens")
            image_count = gr.Number(label="Number of Images")
    
    submit_btn.click(
        fn=process_image_and_text,
        inputs=[input_image, input_prompt],
        outputs=[output_text, token_count, image_count]
    )
    
    gr.Markdown("## How it works")
    gr.Markdown("1. The image is processed by a Vision Encoder using 2D ROPE (Rotary Position Embedding).")
    gr.Markdown("2. The encoder uses GELU activation in its layers.")
    gr.Markdown("3. The encoded image and the prompt are used to generate descriptive text.")
    
    gr.Markdown("## Model Details")
    gr.Markdown(f"- Vision Encoder Hidden Size: {params['vision_encoder']['hidden_size']}")
    gr.Markdown(f"- Number of Vision Encoder Layers: {params['vision_encoder']['num_hidden_layers']}")
    gr.Markdown(f"- Number of Attention Heads: {params['vision_encoder']['num_attention_heads']}")
    gr.Markdown(f"- Image Size: {params['vision_encoder']['image_size']}x{params['vision_encoder']['image_size']}")
    gr.Markdown(f"- Patch Size: {params['vision_encoder']['patch_size']}x{params['vision_encoder']['patch_size']}")

if __name__ == "__main__":
    demo.launch()