Update app.py

app.py

@@ -1,93 +1,136 @@
 import os
-import logging
+import json
+import subprocess
 from threading import Thread
+import logging
 from logging.handlers import RotatingFileHandler
+
 import torch
+import spaces
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, pipeline
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TextIteratorStreamer
+
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
-# Logging setup
 log_file = '/tmp/app_debug.log'
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
 file_handler = RotatingFileHandler(log_file, maxBytes=10*1024*1024, backupCount=5)
-file_handler.setFormatter(logging.Formatter('%(asctime)s - %(levelname)s - %(message)s'))
+file_handler.setLevel(logging.DEBUG)
+formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
+file_handler.setFormatter(formatter)
 logger.addHandler(file_handler)
 
 logger.debug("Application started")
 
-# Define model parameters
 MODEL_ID = "Qwen/Qwen2.5-Coder-7B-Instruct"
+CHAT_TEMPLATE = "ChatML"
+MODEL_NAME = MODEL_ID.split("/")[-1]
 CONTEXT_LENGTH = 16000
 
-# Configuration for 4-bit quantization
-quantization_config = BitsAndBytesConfig(
-    load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16
-)
+COLOR = "blue"
+EMOJI = "🤖"
+DESCRIPTION = f"This is the {MODEL_NAME} model designed for coding assistance and general AI tasks."
 
-# Load tokenizer and model
-if torch.cuda.is_available():
-    logger.debug("GPU is available. Proceeding with GPU setup.")
-    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
-    model = AutoModelForCausalLM.from_pretrained(
-        MODEL_ID,
-        device_map="auto",
-        quantization_config=quantization_config,
-        trust_remote_code=True,
-    )
-else:
-    logger.warning("GPU is not available. Proceeding with CPU setup.")
-    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
-    model = AutoModelForCausalLM.from_pretrained(
-        MODEL_ID,
-        device_map="auto",
-        trust_remote_code=True,
-        low_cpu_mem_usage=True,
-    )
+@spaces.GPU()
+def predict(message, history, system_prompt, temperature, max_new_tokens, top_k, repetition_penalty, top_p):
+    logger.debug(f"Received prediction request: message='{message}', system_prompt='{system_prompt}'")
+    if CHAT_TEMPLATE == "Auto":
+        stop_tokens = [tokenizer.eos_token_id]
+        instruction = system_prompt + "\n\n"
+        for user, assistant in history:
+            instruction += f"User: {user}\nAssistant: {assistant}\n"
+        instruction += f"User: {message}\nAssistant:"
+    elif CHAT_TEMPLATE == "ChatML":
+        stop_tokens = ["<|endoftext|>", "<|im_end|>"]
+        instruction = '<|im_start|>system\n' + system_prompt + '\n<|im_end|>\n'
+        for user, assistant in history:
+            instruction += f'<|im_start|>user\n{user}\n<|im_end|>\n<|im_start|>assistant\n{assistant}\n<|im_end|>\n'
+        instruction += f'<|im_start|>user\n{message}\n<|im_end|>\n<|im_start|>assistant\n'
+    elif CHAT_TEMPLATE == "Mistral Instruct":
+        stop_tokens = ["</s>", "[INST]", "[INST] ", "<s>", "[/INST]", "[/INST] "]
+        instruction = f'<s>[INST] {system_prompt}\n'
+        for user, assistant in history:
+            instruction += f'{user} [/INST] {assistant}</s>[INST]'
+        instruction += f' {message} [/INST]'
+    else:
+        raise Exception("Incorrect chat template, select 'Auto', 'ChatML' or 'Mistral Instruct'")
+    print(instruction)
+    
+    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+    enc = tokenizer(instruction, return_tensors="pt", padding=True, truncation=True)
+    input_ids, attention_mask = enc.input_ids, enc.attention_mask
 
-# Create Hugging Face pipeline
-pipe = pipeline(
-    "text-generation",
-    model=model,
-    tokenizer=tokenizer,
-    max_length=CONTEXT_LENGTH,
-    temperature=0.7,
-    top_k=50,
-    top_p=0.9,
-    repetition_penalty=1.2,
-)
+    if input_ids.shape[1] > CONTEXT_LENGTH:
+        input_ids = input_ids[:, -CONTEXT_LENGTH:]
+        attention_mask = attention_mask[:, -CONTEXT_LENGTH:]
 
-# Prediction function using the model directly
-def predict(
-    message,
-    temperature,
-    max_new_tokens,
-    top_k,
-    repetition_penalty,
-    top_p,
-):
+    generate_kwargs = dict(
+        input_ids=input_ids.to(device),
+        attention_mask=attention_mask.to(device),
+        streamer=streamer,
+        do_sample=True,
+        temperature=temperature,
+        max_new_tokens=max_new_tokens,
+        top_k=top_k,
+        repetition_penalty=repetition_penalty,
+        top_p=top_p
+    )
+    t = Thread(target=model.generate, kwargs=generate_kwargs)
+    t.start()
+    outputs = []
     try:
-        result = pipe(message, temperature=temperature, max_length=max_new_tokens, top_k=top_k, top_p=top_p, repetition_penalty=repetition_penalty)
-        return result[0]['generated_text']
+        for new_token in streamer:
+            outputs.append(new_token)
+            if new_token in stop_tokens:
+                break
+            yield "".join(outputs)
+        logger.debug(f"Prediction completed successfully for message: '{message}'")
     except Exception as e:
-        logger.exception(f"Error during prediction: {e}")
-        return "An error occurred."
+        logger.exception(f"Error during prediction for message '{message}': {str(e)}")
+        yield "An error occurred during processing."
+
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.bfloat16
+)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    device_map="auto",
+    quantization_config=quantization_config,
+    attn_implementation="flash_attention_2",
+)
 
-# Gradio UI
-interface = gr.Interface(
-    fn=predict,
-    inputs=[
-        gr.Textbox(label="User input"),
-        gr.Slider(0, 1, 0.7, label="Temperature"),
-        gr.Slider(128, 2048, 1024, label="Max new tokens"),
+logger.debug("Model and tokenizer loaded successfully")
+
+gr.ChatInterface(
+    predict,
+    title=EMOJI + " " + MODEL_NAME,
+    description=DESCRIPTION,
+    examples=[
+       ["Can you solve the equation 2x + 3 = 11 for x in Python?"],
+       ["Write a Java program that checks if a number is even or odd."],
+       ["How can I reverse a string in JavaScript?"],
+       ["Create a C++ function to find the factorial of a number."],
+       ["Write a Python list comprehension to generate a list of squares of numbers from 1 to 10."],
+       ["How do I implement a binary search algorithm in C?"],
+       ["Write a Ruby script to read a file and count the number of lines in it."],
+       ["Create a Swift class to represent a bank account with deposit and withdrawal methods."],
+       ["How do I find the maximum element in an array using Kotlin?"],
+       ["Write a Rust program to generate the Fibonacci sequence up to the 10th number."]
+    ],
+    additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False),
+    additional_inputs=[
+        gr.Textbox("You are a code assistant.", label="System prompt"),
+        gr.Slider(0, 1, 0.3, label="Temperature"),
+        gr.Slider(128, 4096, 1024, label="Max new tokens"),
         gr.Slider(1, 80, 40, label="Top K sampling"),
         gr.Slider(0, 2, 1.1, label="Repetition penalty"),
         gr.Slider(0, 1, 0.95, label="Top P sampling"),
     ],
-    outputs="text",
-    live=True,
-)
-
-interface.launch()
+    theme=gr.themes.Soft(primary_hue=COLOR),
+).queue().launch()
 
-logger.debug("Chat interface initialized and launched")
+logger.debug("Chat interface initialized and launched")