app.py

import os
import torch
import tiktoken
from model import GPT, GPTConfig
import pickle
import string
import gradio as gr
from contextlib import nullcontext

# Model and Tokenizer setup
device = 'cpu'
dtype = 'bfloat16' if device != 'cpu' and torch.cuda.is_bf16_supported() else 'float16'
device_type = 'cuda' if 'cuda' in device else 'cpu'
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)

cl100k_base = tiktoken.get_encoding("cl100k_base")
enc = tiktoken.Encoding(
    name="cl100k_im",
    pat_str=cl100k_base._pat_str,
    mergeable_ranks=cl100k_base._mergeable_ranks,
    special_tokens={
        **cl100k_base._special_tokens,
        "<EOR>": 100264,
        "<SPECIAL2>": 100265,
    }
)

# Load model from checkpoint
model_save_path = 'tuned_ckpt.pt'
if os.path.exists(model_save_path):
    checkpoint = torch.load(model_save_path, map_location=device)
    gptconf = GPTConfig(**checkpoint['model_args'])
    model = GPT(gptconf)
    model.load_state_dict(checkpoint['model'])
else:
    raise FileNotFoundError(f"Model file {model_save_path} not found")

model.eval()
model.to(device)
model = torch.compile(model)

# Function to encode and decode using the tokenizer
def encode(text):
    return enc.encode(text, allowed_special={"<EOR>"})

def decode(tokens):
    return enc.decode(tokens)

# Function to truncate output to token limit
def truncate_output(text, token_limit):
    tokens = text.split()
    if len(tokens) > token_limit:
        return ' '.join(tokens[:token_limit]) + '...'
    return text

def ensure_complete_output(output, context, max_length, temperature, top_k, top_p, repetition_penalty, eor_token_id):
    while len(output.split()) < max_length and not output.endswith('.'):
        continuation = model.generate(
            torch.tensor(encode(output), dtype=torch.long, device=device)[None, ...],
            max_new_tokens=max_length,
            temperature=temperature,
            top_k=top_k,
            top_p=top_p,
            repetition_penalty=repetition_penalty,
            eor_token_id=eor_token_id
        )
        continuation_text = decode(continuation[0].tolist())
        if eor_token_id in continuation[0].tolist():
            continuation_text = continuation_text.split("<EOR>")[0]
            output += continuation_text
            break
        else:
            output += continuation_text
            if len(output.split()) >= max_length:
                break
    return output

# Text generation function for Gradio interface
def generate_text(prompt, num_samples, max_new_tokens, temperature, top_k, top_p, repetition_penalty, eor_token_id):
    with torch.no_grad():
        with ctx:
            start_ids = encode(prompt)
            initial_prompt = torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...]
            outputs = []
            for _ in range(num_samples):
                y = model.generate(
                    initial_prompt,
                    max_new_tokens=max_new_tokens,
                    temperature=temperature,
                    top_k=top_k,
                    top_p=top_p,
                    repetition_penalty=repetition_penalty,
                    eor_token_id=eor_token_id
                )
                # Filter out tokens after the end-of-response token or similar markers
                output_ids = y[0].tolist()
                if eor_token_id in output_ids:
                    output_ids = output_ids[:output_ids.index(eor_token_id) + 1]  # Include EOR token
                else:
                    # Check for similar markers like '<E' and handle them
                    try:
                        eor_index = next(i for i, token in enumerate(output_ids) if decode([token]).startswith('<E'))
                        output_ids = output_ids[:eor_index]
                    except StopIteration:
                        pass

                # Ensure the prompt is not included in the final output
                output = decode(output_ids).replace(prompt, '').strip()
                output = ensure_complete_output(output, prompt, max_new_tokens, temperature, top_k, top_p, repetition_penalty, eor_token_id)
                truncated_output = truncate_output(output, max_new_tokens)
                outputs.append(truncated_output)
            return '\n\n'.join(outputs)

# Create a Gradio interface
demo = gr.Interface(
    fn=generate_text,
    inputs=[
        gr.Textbox(lines=2, placeholder="Enter your prompt here...", value="Write a short story about a boy:"),
        gr.Slider(minimum=1, maximum=10, step=1, value=1, label="Number of Samples"),
        gr.Slider(minimum=10, maximum=200, step=1, value=75, label="Max New Tokens"),
        gr.Slider(minimum=0.1, maximum=1.0, step=0.1, value=0.8, label="Temperature"),
        gr.Slider(minimum=1, maximum=100, step=1, value=50, label="Top-k"),
        gr.Slider(minimum=0.1, maximum=1.0, step=0.1, value=0.85, label="Top-p"),
        gr.Slider(minimum=1.0, maximum=2.0, step=0.1, value=1.1, label="Repetition Penalty"),
        gr.Number(value=100264, label="End-of-Response Token ID")
    ],
    outputs="text",
    title="GPT Text Generator",
    description="Generate text based on a prompt using a trained GPT model.",
    examples=[
        ["Write a short story about a boy:"],
        ["Explain the theory of relativity:"],
        ["What is the meaning of life?"]
    ]
)

# Launch the Gradio app
demo.launch()