app.py

import json
import gradio as gr
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    AutoModelForSeq2SeqLM,
    pipeline
)

# ── 1) Pick your model ────────────────────────────────────────────────
# ▸ For a causal code model (no T5 errors):
MODEL_ID = "Salesforce/codegen-350M-multi"
# ▸ Or, for a seq‑to‑seq model:
# MODEL_ID = "google/flan-t5-base"
# MODEL_ID = "google/flan-t5-small"
# MODEL_ID = "Salesforce/codegen-350M-multi"

# ── 2) Load tokenizer + model ────────────────────────────────────────
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# Detect T5 vs causal:
if "t5" in MODEL_ID.lower() or MODEL_ID.startswith("google/"):
    # seq‑to‑seq
    model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)
    task  = "text2text-generation"
else:
    # causal
    model = AutoModelForCausalLM.from_pretrained(MODEL_ID)
    task  = "text-generation"

# ── 3) Build the pipeline ────────────────────────────────────────────
pipe = pipeline(
    task,
    model=model,
    tokenizer=tokenizer,
    device=-1,           # CPU
    max_new_tokens=256,
    temperature=0.2
)

# ── 4) Your review function ──────────────────────────────────────────
def review_code(diff: str, guidelines: str):
    prompt = (
        "You are an expert code reviewer.  Return *only* a valid JSON array of objects\n"
        "with two fields each: `line` (number) and `comment` (string).\n\n"
        f"DIFF:\n{diff}\n\n"
        f"GUIDELINES:\n{guidelines}\n\n"
        "OUTPUT FORMAT EXAMPLE:\n"
        '[{"line":12,"comment":"…"}]\n\n'
        "<<<END_JSON>>>\n"
    )

    # Run the model
    out = pipe(prompt)[0]["generated_text"]

    # Truncate at our stop marker, if present
    if "<<<END_JSON>>>" in out:
        out = out.split("<<<END_JSON>>>")[0]

    start = out.find("[")
    end   = out.rfind("]") + 1
    if start < 0 or end < 0:
        return {"error": "No JSON array found", "raw": out}

    snippet = out[start:end]
    try:
        return json.loads(snippet)
    except json.JSONDecodeError:
        return {"error": "JSON parse failed", "raw": snippet}

# ── 5) Gradio interface ──────────────────────────────────────────────
iface = gr.Interface(
    fn=review_code,
    inputs=[
        gr.Textbox(lines=10, label="Git Diff"),
        gr.Textbox(lines=5,  label="Review Guidelines")
    ],
    outputs=gr.JSON(label="Comments"),
    title="🤖 Code Review LLM",
    description="Paste your git diff and guidelines; get back JSON comments."
)

if __name__ == "__main__":
    local_url = iface.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )
    # print so you can curl it immediately
    print(f"🔗 endpoint → {local_url}/predict")