codewraith / app.py
slenk's picture
fix(zerogpu): lower @spaces.GPU duration from 300s to 120s to stay under PRO cap
ee83bcc verified
"""CodeWraith HuggingFace Spaces entry point.
Downloads the LoRA adapter from HF Hub and serves the Gradio interface.
Set HF_REPO_ID environment variable to point to your uploaded adapter.
"""
from __future__ import annotations
import os
import sys
from pathlib import Path
from typing import Any
# Ensure src/ is importable (HF Spaces runs app.py directly)
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "src"))
# Force pandas to fully initialize before transitive imports cause circular import
import pandas # noqa: F401, I001
import gradio as gr
import spaces
from codewraith import SYSTEM_MESSAGE
# --- Config ---
HF_REPO_ID = os.environ.get("HF_REPO_ID", "slenk/codewraith-lora-8b")
MODEL_KEY = os.environ.get("MODEL_KEY", "8b")
ADAPTER_DIR = "./adapter"
MODELS = {
"3b": "unsloth/Llama-3.2-3B-Instruct",
"8b": "unsloth/Llama-3.1-8B-Instruct",
}
EXAMPLE_CODE = '''\
def fibonacci(n: int) -> list[int]:
"""Generate the first n Fibonacci numbers."""
if n <= 0:
return []
sequence = [0, 1]
while len(sequence) < n:
sequence.append(sequence[-1] + sequence[-2])
return sequence[:n]
'''
# --- Global state ---
_model = None
_tokenizer = None
_retriever = None
# --- Model loading ---
def download_adapter():
"""Download the LoRA adapter from HF Hub if not already cached."""
if Path(ADAPTER_DIR).exists() and any(Path(ADAPTER_DIR).iterdir()):
print(f"Adapter already cached at {ADAPTER_DIR}")
return
from huggingface_hub import snapshot_download
print(f"Downloading adapter from {HF_REPO_ID}...")
snapshot_download(repo_id=HF_REPO_ID, local_dir=ADAPTER_DIR)
print("Download complete.")
def load_model() -> tuple[Any, Any]:
"""Load the base model with LoRA adapter."""
global _model, _tokenizer # noqa: PLW0603
if _model is not None:
return _model, _tokenizer
download_adapter()
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
model_name = MODELS[MODEL_KEY]
print(f"Loading {model_name}...")
bnb_config = BitsAndBytesConfig(load_in_4bit=True)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
quantization_config=bnb_config,
device_map="auto",
)
model = PeftModel.from_pretrained(model, ADAPTER_DIR)
model.eval()
_model, _tokenizer = model, tokenizer
return model, tokenizer
# --- RAG ---
def init_retriever():
"""Initialize retriever if ChromaDB index exists."""
global _retriever # noqa: PLW0603
if _retriever is not None:
return _retriever
try:
from codewraith.app.retriever import SpecRetriever
retriever = SpecRetriever()
if Path("data/chromadb").exists():
collection = retriever._get_collection()
if collection.count() > 0:
_retriever = retriever
print(f"RAG retriever loaded ({collection.count()} examples)")
return _retriever
except ImportError:
pass
return None
def retrieve_context(source_code: str, n_results: int = 3) -> str:
"""Retrieve similar examples as context."""
retriever = init_retriever()
if retriever is None:
return ""
examples = retriever.retrieve(source_code, n_results=n_results)
if not examples:
return ""
return retriever.format_context(examples)
# --- Inference ---
@spaces.GPU(duration=120)
def generate_spec(
source_code: str,
temperature: float = 0.7,
top_p: float = 0.9,
max_tokens: int = 2048,
use_rag: bool = True,
) -> str:
"""Generate a technical specification."""
if not source_code.strip():
return "*Please paste some Python source code.*"
try:
model, tokenizer = load_model()
user_content = source_code
if use_rag:
context = retrieve_context(source_code)
if context:
user_content = context + source_code
messages = [
{"role": "system", "content": SYSTEM_MESSAGE},
{"role": "user", "content": user_content},
]
input_text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Retry without RAG if input too long
if input_len > 6000 and use_rag:
messages = [
{"role": "system", "content": SYSTEM_MESSAGE},
{"role": "user", "content": source_code},
]
input_text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
input_len = inputs["input_ids"].shape[-1]
outputs = model.generate(
**inputs,
max_new_tokens=max_tokens,
temperature=temperature,
top_p=top_p,
do_sample=True,
)
generated = outputs[0][input_len:]
spec = tokenizer.decode(generated, skip_special_tokens=True)
return render_mermaid_images(spec)
except Exception as e:
return (
f"**Error generating specification:**\n\n```\n{e}\n```\n\n"
"Try with a shorter input or disable RAG."
)
# --- Gradio UI ---
def render_mermaid_images(spec: str) -> str:
"""Replace mermaid code blocks with rendered SVG images via mermaid.ink.
Validates the mermaid syntax first, strips malformed blocks, and
converts valid ones to inline images that render reliably regardless
of CSS/theme issues.
"""
import base64
import re
valid_starts = (
"graph ",
"graph\n",
"flowchart ",
"flowchart\n",
"classDiagram",
"sequenceDiagram",
"stateDiagram",
"erDiagram",
"gantt",
"pie",
"gitgraph",
)
def replace_block(match: re.Match) -> str:
block = match.group(1).strip()
# Must start with a valid diagram type
if not any(block.startswith(s) for s in valid_starts):
return "*[Mermaid diagram removed: unrecognized diagram type]*"
# Check balanced brackets/braces
if block.count("[") != block.count("]"):
return "*[Mermaid diagram removed: unbalanced brackets]*"
if block.count("{") != block.count("}"):
return "*[Mermaid diagram removed: unbalanced braces]*"
if block.count("(") != block.count(")"):
return "*[Mermaid diagram removed: unbalanced parentheses]*"
# Encode and return as mermaid.ink image
encoded = base64.urlsafe_b64encode(block.encode("utf-8")).decode("ascii")
url = f"https://mermaid.ink/svg/{encoded}"
return f'<img src="{url}" alt="Dependency Diagram" style="max-width: 600px;">'
return re.sub(r"```mermaid\s*\n(.*?)```", replace_block, spec, flags=re.DOTALL)
def create_app():
with gr.Blocks(
title="CodeWraith - Module-to-Spec Transformer",
) as app:
gr.Markdown(
"# CodeWraith\n"
"Generate technical specifications from Python source code.\n\n"
"Paste your Python code below, adjust sampling parameters, "
"and click **Generate Specification**."
)
code_input = gr.Code(
language="python",
label="Python Source Code",
value=EXAMPLE_CODE,
lines=15,
)
with gr.Row():
temperature = gr.Slider(0.0, 2.0, value=0.7, step=0.1, label="Temperature")
top_p = gr.Slider(0.0, 1.0, value=0.9, step=0.05, label="Top-p")
max_tokens = gr.Slider(256, 8192, value=4096, step=256, label="Max Tokens")
with gr.Row():
use_rag = gr.Checkbox(value=True, label="Use RAG (retrieve similar examples)")
generate_btn = gr.Button("Generate Specification", variant="primary")
clear_input_btn = gr.Button("Clear Input", variant="secondary")
clear_output_btn = gr.Button("Clear Output", variant="secondary")
gr.Markdown("*Model loads on first generation (~30s). Subsequent calls are fast.*")
spec_output = gr.Markdown(label="Generated Specification")
loading_msg = "*Generating specification... (loading model if first run)*"
generate_btn.click(
fn=lambda: gr.update(value=loading_msg),
outputs=spec_output,
).then(
fn=generate_spec,
inputs=[code_input, temperature, top_p, max_tokens, use_rag],
outputs=spec_output,
)
clear_input_btn.click(
fn=lambda: "",
outputs=code_input,
)
clear_output_btn.click(
fn=lambda: "",
outputs=spec_output,
)
return app
# Preload adapter on startup (CPU time, free)
print("Preloading adapter...")
download_adapter()
print("Adapter ready. Model will load on first GPU request.")
if __name__ == "__main__":
app = create_app()
app.launch()