code/inference.py

import json
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from typing import List, Dict
from accelerate import load_checkpoint_and_dispatch

# Global variables to persist the model and tokenizer between invocations
model = None
tokenizer = None

# Function to format chat messages using Qwen's chat template
def format_chat(messages: List[Dict[str, str]], tokenizer) -> str:
    """
    Format chat messages using Qwen's chat template.
    """
    return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Model loading function for SageMaker with tensor parallelism and offloading
def model_fn(model_dir, context=None):
    """
    Load the model and tokenizer from the model directory for inference.
    Supports tensor parallelism across multiple GPUs with offloading.
    The model is loaded only once and stored in a global variable.
    """
    global model, tokenizer  # Declare model and tokenizer as global to persist across invocations
    
    if model is None:  # Check if the model is already loaded
        print("Loading the model and tokenizer...")
        # Define an offload directory for any model components that can't fit in GPU memory
        offload_dir = "/tmp/offload_dir"  # Ensure SageMaker has write access to this directory

        # Load and dispatch the model across multiple GPUs
        model = AutoModelForCausalLM.from_pretrained(model_dir, torch_dtype=torch.float16)
        model = load_checkpoint_and_dispatch(
            model, 
            model_dir, 
            device_map="auto",  # Automatically map model layers across devices
            offload_folder=offload_dir,  # Offload parts of the model to disk if GPU memory is insufficient
        )
        
        # Load the tokenizer
        tokenizer = AutoTokenizer.from_pretrained(model_dir)

    return model, tokenizer

# Custom predict function for SageMaker
def predict_fn(input_data, model_and_tokenizer):
    """
    Generate predictions for the input data.
    """
    try:
        model, tokenizer = model_and_tokenizer
        data = json.loads(input_data)

        # Format the prompt using Qwen's chat template
        messages = data.get("messages", [])
        formatted_prompt = format_chat(messages, tokenizer)

        # Tokenize the input
        inputs = tokenizer([formatted_prompt], return_tensors="pt").to("cuda:0")  # Send input to GPU 0 for generation

        # Generate output
        outputs = model.generate(
            inputs['input_ids'],
            max_new_tokens=data.get("max_new_tokens", 512),
            temperature=data.get("temperature", 0.7),
            top_p=data.get("top_p", 0.9),
            repetition_penalty=data.get("repetition_penalty", 1.0),
            length_penalty=data.get("length_penalty", 1.0),
            do_sample=True
        )

        # Decode the output
        generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]

        # Build response
        response = {
            "id": "chatcmpl-uuid",
            "object": "chat.completion",
            "model": "qwen-72b",
            "choices": [{
                "index": 0,
                "message": {
                    "role": "assistant",
                    "content": generated_text
                },
                "finish_reason": "stop"
            }],
            "usage": {
                "prompt_tokens": len(inputs['input_ids'][0]),
                "completion_tokens": len(outputs[0]),
                "total_tokens": len(inputs['input_ids'][0]) + len(outputs[0])
            }
        }
        return response

    except Exception as e:
        return {"error": str(e), "details": repr(e)}

# Define input format for SageMaker
def input_fn(serialized_input_data, content_type,context=None):
    """
    Prepare the input data for inference.
    """
    return serialized_input_data

# Define output format for SageMaker
def output_fn(prediction_output, accept, context=None):
    """
    Convert the model output to a JSON response.
    """
    return json.dumps(prediction_output)