From the uDemy course on LLM engineering.
https://www.udemy.com/course/llm-engineering-master-ai-and-large-language-models
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
346 lines
13 KiB
346 lines
13 KiB
import os |
|
import io |
|
import sys |
|
import re |
|
import subprocess |
|
from dotenv import load_dotenv |
|
from openai import OpenAI |
|
from anthropic import Anthropic |
|
import gradio as gr |
|
|
|
# Load environment variables and initialize APIs |
|
load_dotenv(override=True) |
|
openai = OpenAI(api_key=os.getenv("OPENAI_API_KEY")) |
|
anthropic = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY")) |
|
MACHINE_SPEC = "MacbookPro, Apple M1 Chip" |
|
|
|
# Define global variables for HF integration |
|
# For HF chat-based CodeQwen model |
|
code_qwen = "Qwen/CodeQwen1.5-7B-Chat" |
|
CODE_QWEN_URL = "" |
|
|
|
|
|
def clean_code(code, target_language): |
|
""" |
|
Remove markdown code fences and stray language indicators. |
|
Also apply language-specific replacements. |
|
""" |
|
raw_lines = code.splitlines() |
|
cleaned_lines = [] |
|
for line in raw_lines: |
|
if "```" in line: |
|
continue |
|
if line.strip().lower() in ["c", "cpp", "c++", "rust"]: |
|
continue |
|
cleaned_lines.append(line) |
|
cleaned = "\n".join(cleaned_lines) |
|
if target_language == "C": |
|
cleaned = cleaned.replace("1U << 32", "(1ULL << 32)") |
|
if target_language == "Rust": |
|
cleaned = process_rust_code(cleaned) |
|
return cleaned |
|
|
|
# Conversion prompt functions (target language-aware) |
|
def user_prompt_for(python_code, target_language): |
|
return ( |
|
f"Rewrite this Python code in {target_language} with the fastest possible implementation that produces identical output. " |
|
f"Respond only with {target_language} code; do not explain your work. " |
|
"Pay attention to number types to ensure no int overflows. Remember to #include all necessary C++ packages such as iomanip.\n\n" |
|
+ python_code |
|
) |
|
|
|
def messages_for(python_code, target_language): |
|
system_message = ( |
|
f"You are an assistant that reimplements Python code in high performance {target_language} for an {MACHINE_SPEC}. " |
|
f"Respond only with {target_language} code; use comments sparingly. " |
|
f"The {target_language} response needs to produce an identical output in the fastest possible time." |
|
) |
|
return [ |
|
{"role": "system", "content": system_message}, |
|
{"role": "user", "content": user_prompt_for(python_code, target_language)}, |
|
] |
|
|
|
def write_output(code, target_language): |
|
"""Write the converted code to a file based on target language.""" |
|
tag = target_language.lower() if target_language is not None else "" |
|
if target_language == "C++": |
|
filename = "optimized.cpp" |
|
elif target_language == "C": |
|
filename = "optimized.c" |
|
elif target_language == "Rust": |
|
filename = "optimized.rs" |
|
else: |
|
filename = "optimized.txt" |
|
cleaned = code.replace(f"```{tag}\n", "").replace("```", "") |
|
lines = cleaned.splitlines() |
|
if lines and lines[0].strip().lower() in ["cpp", "c++", "c", "rust"]: |
|
lines = lines[1:] |
|
cleaned = "\n".join(lines) |
|
cleaned = clean_code(cleaned, target_language) |
|
with open(filename, "w") as f: |
|
f.write(cleaned) |
|
return filename |
|
|
|
# GPT integration for conversion |
|
def stream_gpt(python_code, target_language, model_version): |
|
stream = openai.chat.completions.create( |
|
model=model_version, # Use selected GPT model version |
|
messages=messages_for(python_code, target_language), |
|
stream=True, |
|
) |
|
reply = "" |
|
for chunk in stream: |
|
if not hasattr(chunk, "choices") or not chunk.choices: |
|
continue |
|
fragment = chunk.choices[0].delta.content or "" |
|
reply += fragment |
|
yield reply.replace(f"```{target_language}\n", "").replace("```", "") |
|
|
|
# Claude integration for conversion |
|
def stream_claude(python_code, target_language, model_version): |
|
prompt = user_prompt_for(python_code, target_language) |
|
response = anthropic.completions.create( |
|
prompt=prompt, |
|
model=model_version, |
|
stream=True, |
|
) |
|
reply = "" |
|
for chunk in response: |
|
fragment = chunk.get("completion", "") |
|
reply += fragment |
|
yield reply.replace(f"```{target_language}\n", "").replace("```", "") |
|
|
|
# Hugging Face integration functions |
|
def stream_code_qwen(python_code, target_language, model_version): |
|
""" |
|
HF chat-based model using CodeQwen. |
|
""" |
|
from transformers import AutoTokenizer |
|
tokenizer = AutoTokenizer.from_pretrained(code_qwen) |
|
messages = messages_for(python_code, target_language) |
|
# Convert messages to chat format as expected by Qwen. |
|
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
|
from huggingface_hub import InferenceClient |
|
client = InferenceClient(CODE_QWEN_URL, token=os.getenv("HF_TOKEN")) |
|
stream = client.text_generation(text, stream=True, details=True, max_new_tokens=3000) |
|
result = "" |
|
for r in stream: |
|
result += r.token.text |
|
yield result.replace(f"```{target_language}\n", "").replace("```", "") |
|
|
|
def stream_huggingface(python_code, target_language, model_version): |
|
""" |
|
HF single-prompt model integration. |
|
""" |
|
prompt = user_prompt_for(python_code, target_language) |
|
from huggingface_hub import InferenceClient |
|
client = InferenceClient(model_name=model_version, token=os.getenv("HF_TOKEN")) |
|
stream = client.text_generation(prompt, stream=True, details=True, max_new_tokens=3000) |
|
reply = "" |
|
for chunk in stream: |
|
reply += chunk.token.text |
|
yield reply.replace(f"```{target_language}\n", "").replace("```", "") |
|
|
|
|
|
def optimize(python_code, combined_model, target_language): |
|
""" |
|
combined_model is a string like "GPT: gpt-4o", "CLAUDE: claude-3-5-sonnet-20240620" or "HF: model_name" |
|
""" |
|
provider, model_version = [x.strip() for x in combined_model.split(":")] |
|
if provider == "GPT": |
|
for partial in stream_gpt(python_code, target_language, model_version): |
|
yield partial |
|
elif provider == "CLAUDE": |
|
for partial in stream_claude(python_code, target_language, model_version): |
|
yield partial |
|
elif provider == "HF": |
|
if "CodeQwen" in model_version: |
|
for partial in stream_code_qwen(python_code, target_language, model_version): |
|
yield partial |
|
else: |
|
for partial in stream_huggingface(python_code, target_language, model_version): |
|
yield partial |
|
else: |
|
raise ValueError("Unknown model provider") |
|
|
|
def execute_python(code): |
|
"""Execute Python code and return its output.""" |
|
env = {} # Dedicated global namespace |
|
try: |
|
output = io.StringIO() |
|
sys.stdout = output |
|
exec(code, env) |
|
finally: |
|
sys.stdout = sys.__stdout__ |
|
return output.getvalue() |
|
|
|
def execute_cpp(code): |
|
write_output(code, target_language="C++") |
|
try: |
|
compile_cmd = [ |
|
"clang++", "-Ofast", "-std=c++17", "-march=armv8.5-a", |
|
"-mtune=apple-m1", "-mcpu=apple-m1", "-o", "optimized", "optimized.cpp" |
|
] |
|
subprocess.run(compile_cmd, check=True, text=True, capture_output=True) |
|
run_cmd = ["./optimized"] |
|
run_result = subprocess.run(run_cmd, check=True, text=True, capture_output=True) |
|
return run_result.stdout |
|
except subprocess.CalledProcessError as e: |
|
return f"Error:\n{e.stderr}" |
|
|
|
def execute_c(code): |
|
cleaned_code = clean_code(code, "C") |
|
with open("optimized.c", "w") as f: |
|
f.write(cleaned_code) |
|
try: |
|
compile_cmd = ["clang", "-O2", "-std=c11", "-o", "optimized_c", "optimized.c"] |
|
subprocess.run(compile_cmd, check=True, text=True, capture_output=True) |
|
run_cmd = ["./optimized_c"] |
|
run_result = subprocess.run(run_cmd, check=True, text=True, capture_output=True) |
|
return run_result.stdout |
|
except subprocess.CalledProcessError as e: |
|
return f"Error:\n{e.stderr}" |
|
|
|
def process_rust_code(code): |
|
code = code.replace("{:.6f}", "{:.6}") |
|
code = re.sub( |
|
r'(println!$begin:math:text$"Execution Time: \\{\\:\\.6\\} seconds", duration\\.as_secs_f64)(\\s*)$', |
|
r'\\1())', |
|
code, |
|
flags=re.MULTILINE, |
|
) |
|
code = code.replace("max_val - min_val as u32 + 1", "((max_val - min_val + 1) as u32)") |
|
code = code.replace("1 << 32", "1u64 << 32") |
|
code = re.sub(r'($end:math:text$\s*as i64)\)', r'\1', code) |
|
return code |
|
|
|
def execute_rust(code): |
|
code = code.replace("```rust\n", "").replace("```", "") |
|
lines = code.split('\n', 1) |
|
if lines and lines[0].strip().lower() == "rust": |
|
code = lines[1] if len(lines) > 1 else "" |
|
code = process_rust_code(code) |
|
with open("optimized.rs", "w") as f: |
|
f.write(code) |
|
try: |
|
compile_cmd = ["rustc", "optimized.rs", "-O", "-o", "optimized_rust"] |
|
subprocess.run(compile_cmd, check=True, text=True, capture_output=True) |
|
run_cmd = ["./optimized_rust"] |
|
run_result = subprocess.run(run_cmd, check=True, text=True, capture_output=True) |
|
return run_result.stdout |
|
except subprocess.CalledProcessError as e: |
|
return f"Error:\n{e.stderr}" |
|
|
|
def execute_target_code(code, target_language): |
|
"""Select the appropriate execution function based on target language.""" |
|
if target_language == "C++": |
|
return execute_cpp(code) |
|
elif target_language == "C": |
|
return execute_c(code) |
|
elif target_language == "Rust": |
|
return execute_rust(code) |
|
else: |
|
return "Unsupported language" |
|
|
|
# Gradio UI setup |
|
css = """ |
|
.python {background-color: #306998;} |
|
.code {background-color: #050;} |
|
""" |
|
|
|
def launch_ui(): |
|
with gr.Blocks(css=css) as ui: |
|
gr.Markdown("## Convert Python Code to C/C++/Rust") |
|
with gr.Row(): |
|
python_box = gr.Textbox(label="Python code:", value=PYTHON_HARD, lines=10) |
|
converted_box = gr.Textbox(label="Converted Code:", lines=10) |
|
with gr.Row(): |
|
model_dropdown = gr.Dropdown( |
|
["GPT: gpt-4o", "GPT: gpt-4o-mini", "CLAUDE: claude-3-5-sonnet-20240620", "CLAUDE: claude-3-haiku-20240307", "HF: CodeQwen1.5-7B-Chat", "HF: bigcode/starcoder"], |
|
label="Select Model", |
|
value="GPT: gpt-4o" |
|
) |
|
target_lang_dropdown = gr.Dropdown( |
|
["C++", "C", "Rust"], |
|
label="Select target language", |
|
value="C++" |
|
) |
|
with gr.Row(): |
|
convert_btn = gr.Button("Convert code") |
|
with gr.Row(): |
|
python_run_btn = gr.Button("Run Python") |
|
run_converted_btn = gr.Button("Run Converted Code") |
|
with gr.Row(): |
|
python_out = gr.TextArea(label="Python result:", elem_classes=["python"]) |
|
converted_out = gr.TextArea(label="Converted Code result:", elem_classes=["code"]) |
|
convert_btn.click( |
|
optimize, |
|
inputs=[python_box, model_dropdown, target_lang_dropdown], |
|
outputs=[converted_box], |
|
) |
|
python_run_btn.click(execute_python, inputs=[python_box], outputs=[python_out]) |
|
run_converted_btn.click( |
|
execute_target_code, |
|
inputs=[converted_box, target_lang_dropdown], |
|
outputs=[converted_out], |
|
) |
|
ui.launch() |
|
|
|
# Example Python code blocks |
|
PYTHON_HARD = """ |
|
# Support large number sizes |
|
def lcg(seed, a=1664525, c=1013904223, m=2**32): |
|
value = seed |
|
while True: |
|
value = (a * value + c) % m |
|
yield value |
|
def max_subarray_sum(n, seed, min_val, max_val): |
|
lcg_gen = lcg(seed) |
|
random_numbers = [next(lcg_gen) % (max_val - min_val + 1) + min_val for _ in range(n)] |
|
max_sum = float('-inf') |
|
for i in range(n): |
|
current_sum = 0 |
|
for j in range(i, n): |
|
current_sum += random_numbers[j] |
|
if current_sum > max_sum: |
|
max_sum = current_sum |
|
return max_sum |
|
def total_max_subarray_sum(n, initial_seed, min_val, max_val): |
|
total_sum = 0 |
|
lcg_gen = lcg(initial_seed) |
|
for _ in range(20): |
|
seed = next(lcg_gen) |
|
total_sum += max_subarray_sum(n, seed, min_val, max_val) |
|
return total_sum |
|
n = 10000 |
|
initial_seed = 42 |
|
min_val = -10 |
|
max_val = 10 |
|
import time |
|
start_time = time.time() |
|
result = total_max_subarray_sum(n, initial_seed, min_val, max_val) |
|
end_time = time.time() |
|
print("Total Maximum Subarray Sum (20 runs):", result) |
|
print("Execution Time: {:.6f} seconds".format(end_time - start_time)) |
|
""" |
|
|
|
if __name__ == "__main__": |
|
import argparse |
|
parser = argparse.ArgumentParser( |
|
description="Single script with multiple executable sections and target language support" |
|
) |
|
parser.add_argument( |
|
"--mode", |
|
choices=["direct", "ui"], |
|
default="ui", |
|
help="Run direct conversion or launch Gradio UI", |
|
) |
|
args = parser.parse_args() |
|
|
|
if args.mode == "direct": |
|
print("\nExecuting Python code (PYTHON_HARD)...") |
|
exec(PYTHON_HARD) |
|
for partial in optimize(PYTHON_HARD, "GPT: gpt-4o", "C++"): |
|
print(partial, end="") |
|
elif args.mode == "ui": |
|
launch_ui() |