operators/llm_op.py

from dora import DoraStatus
import pylcs
import os
import pyarrow as pa
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

import re
import time

CHATGPT = False
MODEL_NAME_OR_PATH = "TheBloke/deepseek-coder-6.7B-instruct-GPTQ"

CODE_MODIFIER_TEMPLATE = """
### Instruction
Respond with one block of modified code only in ```python block. No explaination.

```python
{code}
```

{user_message}

### Response:
"""


model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME_OR_PATH,
    device_map="auto",
    trust_remote_code=True,
    revision="main",
    max_length=1024,
).to("cuda:0")


tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME_OR_PATH, use_fast=True)


def extract_python_code_blocks(text):
    """
    Extracts Python code blocks from the given text that are enclosed in triple backticks with a python language identifier.

    Parameters:
    - text: A string that may contain one or more Python code blocks.

    Returns:
    - A list of strings, where each string is a block of Python code extracted from the text.
    """
    pattern = r"```python\n(.*?)\n```"
    matches = re.findall(pattern, text, re.DOTALL)
    if len(matches) == 0:
        pattern = r"```python\n(.*?)(?:\n```|$)"
        matches = re.findall(pattern, text, re.DOTALL)
        if len(matches) == 0:
            return [text]
        else:
            matches = [remove_last_line(matches[0])]

    return matches


def remove_last_line(python_code):
    """
    Removes the last line from a given string of Python code.

    Parameters:
    - python_code: A string representing Python source code.

    Returns:
    - A string with the last line removed.
    """
    lines = python_code.split("\n")  # Split the string into lines
    if lines:  # Check if there are any lines to remove
        lines.pop()  # Remove the last line
    return "\n".join(lines)  # Join the remaining lines back into a string


def calculate_similarity(source, target):
    """
    Calculate a similarity score between the source and target strings.
    This uses the edit distance relative to the length of the strings.
    """
    edit_distance = pylcs.edit_distance(source, target)
    max_length = max(len(source), len(target))
    # Normalize the score by the maximum possible edit distance (the length of the longer string)
    similarity = 1 - (edit_distance / max_length)
    return similarity


def find_best_match_location(source_code, target_block):
    """
    Find the best match for the target_block within the source_code by searching line by line,
    considering blocks of varying lengths.
    """
    source_lines = source_code.split("\n")
    target_lines = target_block.split("\n")

    best_similarity = 0
    best_start_index = 0
    best_end_index = -1

    # Iterate over the source lines to find the best matching range for all lines in target_block
    for start_index in range(len(source_lines) - len(target_lines) + 1):
        for end_index in range(start_index + len(target_lines), len(source_lines) + 1):
            current_window = "\n".join(source_lines[start_index:end_index])
            current_similarity = calculate_similarity(current_window, target_block)
            if current_similarity > best_similarity:
                best_similarity = current_similarity
                best_start_index = start_index
                best_end_index = end_index

    # Convert line indices back to character indices for replacement
    char_start_index = len("\n".join(source_lines[:best_start_index])) + (
        1 if best_start_index > 0 else 0
    )
    char_end_index = len("\n".join(source_lines[:best_end_index]))

    return char_start_index, char_end_index


def replace_code_in_source(source_code, replacement_block: str):
    """
    Replace the best matching block in the source_code with the replacement_block, considering variable block lengths.
    """
    replacement_block = extract_python_code_blocks(replacement_block)[0]
    start_index, end_index = find_best_match_location(source_code, replacement_block)
    if start_index != -1 and end_index != -1:
        # Replace the best matching part with the replacement block
        new_source = (
            source_code[:start_index] + replacement_block + source_code[end_index:]
        )
        return new_source
    else:
        return source_code


class Operator:

    def on_event(
        self,
        dora_event,
        send_output,
    ) -> DoraStatus:
        global model, tokenizer
        if dora_event["type"] == "INPUT" and dora_event["id"] == "text":
            input = dora_event["value"][0].as_py()
            # Path to the current file
            current_file_path = __file__

            # Directory of the current file
            current_directory = os.path.dirname(current_file_path)
            path = current_directory + "/policy.py"

            with open(path, "r", encoding="utf8") as f:
                code = f.read()

            user_message = input
            start_llm = time.time()

            output = self.ask_llm(
                CODE_MODIFIER_TEMPLATE.format(code=code, user_message=user_message)
            )

            source_code = replace_code_in_source(code, output)
            print("response time:", time.time() - start_llm, flush=True)

            print("response: ", output, flush=True)
            with open(path, "w") as file:
                file.write(source_code)
            del model
            del tokenizer
            # model will still be on cache until its place is taken by other objects so also execute the below lines
            import gc  # garbage collect library

            gc.collect()
            torch.cuda.empty_cache()
            time.sleep(9)
            send_output("init", pa.array([]))

            ## Stopping to liberate GPU space
            return DoraStatus.STOP

        return DoraStatus.CONTINUE

    def ask_llm(self, prompt):

        # Generate output
        # prompt = PROMPT_TEMPLATE.format(system_message=system_message, prompt=prompt))
        input = tokenizer(prompt, return_tensors="pt")
        input_ids = input.input_ids.cuda()

        # add attention mask here
        attention_mask = input.attention_mask.cuda()

        output = model.generate(
            inputs=input_ids,
            temperature=0.7,
            do_sample=True,
            top_p=0.95,
            top_k=40,
            max_new_tokens=512,
            attention_mask=attention_mask,
            eos_token_id=tokenizer.eos_token_id,
        )
        # Get the tokens from the output, decode them, print them

        # Get text between im_start and im_end
        return tokenizer.decode(output[0], skip_special_tokens=True)[len(prompt) :]


if __name__ == "__main__":
    op = Operator()

    # Path to the current file
    current_file_path = __file__

    # Directory of the current file
    current_directory = os.path.dirname(current_file_path)

    path = current_directory + "/policy.py"
    with open(path, "r", encoding="utf8") as f:
        raw = f.read()

    op.on_event(
        {
            "type": "INPUT",
            "id": "text",
            "value": pa.array(
                [
                    {
                        "path": path,
                        "user_message": "go to the living room, ask the model if there is people, if there is, say i'm going to go get coffee for you, then go to the kitchen, when you reach the kitchen, check with the model if there is a person and say can i have a coffee please, then wait 10 sec and go back to the living room",
                    },
                ]
            ),
            "metadata": [],
        },
        print,
    )