server.py

#!/usr/bin/env python3
"""FastAPI server for the Korean pest detector.

Wraps the validated Unsloth FastVisionModel + PEFT runtime LoRA setup
(load_in_4bit=True by default → ~8.7 GB VRAM).

Endpoints:
    GET  /health          → {"status": "ok", "model_loaded": bool}
    GET  /classes         → ["검거세미밤나방", ...] (19 classes)
    GET  /                → minimal HTML upload form
    POST /classify        → multipart file OR JSON {"image": "<base64>"}
                            returns {"pred": ..., "raw": ..., "elapsed_s": ..., "all_classes": [...]}

Env:
    BASE_MODEL       default: unsloth/Qwen3.5-9B
    ADAPTER          default: pfox1995/pest-detector-deploy
    LOAD_IN_4BIT     "true"/"false" (default: true)
    PORT             default: 8080

Usage:
    python server.py
"""
import base64
import io
import os
import time
from contextlib import asynccontextmanager
from typing import Optional

import torch
import uvicorn
from fastapi import FastAPI, File, UploadFile, HTTPException
from fastapi.responses import HTMLResponse, JSONResponse
from PIL import Image
from pydantic import BaseModel

# ─── Constants from training (DO NOT change) ─────────────────────────────
# 클래스 목록은 학습 시점의 어댑터에 따라 다름:
#   - pfox1995/pest-detector-deploy : 19 classes (비단노린재, 알락수염노린재 포함)
#   - WizWix/kor-pest-detector      : 20 classes (복숭아혹진딧물, 열대거세미나방, 홍비단노린재 추가)
# 아래는 WizWix 어댑터의 evaluation/metrics.json 의 캐노니컬 목록 (20개, 한글 정렬).
PEST_CLASSES = [
    "검거세미밤나방", "꽃노랑총채벌레", "담배가루이", "담배거세미나방",
    "담배나방", "도둑나방", "먹노린재", "목화바둑명나방", "무잎벌",
    "배추좀나방", "배추흰나비", "벼룩잎벌레", "복숭아혹진딧물",
    "썩덩나무노린재", "열대거세미나방", "정상", "큰28점박이무당벌레",
    "톱다리개미허리노린재", "파밤나방", "홍비단노린재",
]
SYSTEM_MSG = (
    "당신은 작물 해충 식별 전문가입니다. "
    "사진을 보고 해충의 이름만 한국어로 답하세요. "
    '해충이 없으면 "정상"이라고만 답하세요. '
    "부가 설명 없이 이름만 출력하세요."
)
USER_PROMPT = "이 사진에 있는 해충의 이름을 알려주세요."
LETTERBOX_SIZE = 512
LETTERBOX_FILL = (128, 128, 128)


def letterbox(img: Image.Image, size: int = LETTERBOX_SIZE) -> Image.Image:
    img = img.convert("RGB")
    w, h = img.size
    scale = size / max(w, h)
    nw, nh = int(round(w * scale)), int(round(h * scale))
    resized = img.resize((nw, nh), Image.Resampling.LANCZOS)
    canvas = Image.new("RGB", (size, size), LETTERBOX_FILL)
    canvas.paste(resized, ((size - nw) // 2, (size - nh) // 2))
    return canvas


# ─── Model state ─────────────────────────────────────────────────────────
class ModelState:
    model = None
    tokenizer = None
    text_tokenizer = None  # underlying transformers tokenizer (for stop_strings=)


STATE = ModelState()


def load_model():
    from unsloth import FastVisionModel
    from peft import PeftModel
    from huggingface_hub import snapshot_download

    base = os.environ.get("BASE_MODEL", "unsloth/Qwen3.5-9B")
    adapter = os.environ.get("ADAPTER", "pfox1995/pest-detector-runpod")
    four_bit = os.environ.get("LOAD_IN_4BIT", "true").lower() == "true"

    if os.environ.get("HF_TOKEN"):
        from huggingface_hub import login
        login(token=os.environ["HF_TOKEN"], add_to_git_credential=False)

    print(f"[startup] FastVisionModel.from_pretrained({base}, load_in_4bit={four_bit})", flush=True)
    t0 = time.time()
    model, tok = FastVisionModel.from_pretrained(base, load_in_4bit=four_bit)
    print(f"[startup]   loaded base in {time.time()-t0:.1f}s; vram={torch.cuda.memory_allocated()/1e9:.1f} GB", flush=True)

    adapter_dir = adapter if os.path.isdir(adapter) else snapshot_download(repo_id=adapter)
    print(f"[startup] attaching LoRA: {adapter_dir}", flush=True)
    model = PeftModel.from_pretrained(model, adapter_dir)
    FastVisionModel.for_inference(model)
    model.eval()
    print(f"[startup]   ready; vram={torch.cuda.memory_allocated()/1e9:.1f} GB", flush=True)

    STATE.model = model
    STATE.tokenizer = tok
    STATE.text_tokenizer = tok.tokenizer if hasattr(tok, "tokenizer") else tok


def classify_image(img: Image.Image) -> dict:
    if STATE.model is None:
        raise RuntimeError("Model not loaded")
    image = letterbox(img)
    messages = [
        {"role": "system", "content": [{"type": "text", "text": SYSTEM_MSG}]},
        {"role": "user", "content": [
            {"type": "image", "image": image},
            {"type": "text",  "text": USER_PROMPT},
        ]},
    ]
    text = STATE.tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, enable_thinking=False,
    )
    inputs = STATE.tokenizer(
        image, text, add_special_tokens=False, return_tensors="pt",
    ).to("cuda")

    t0 = time.time()
    with torch.inference_mode():
        out = STATE.model.generate(
            **inputs,
            max_new_tokens=10,
            use_cache=True,
            stop_strings=["\n"],
            tokenizer=STATE.text_tokenizer,
        )
    elapsed = time.time() - t0
    raw = STATE.tokenizer.decode(
        out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True,
    ).strip()
    pred = raw if raw in PEST_CLASSES else None
    if pred is None:
        for c in sorted(PEST_CLASSES, key=len, reverse=True):
            if raw.startswith(c):
                pred = c
                break
        if pred is None:
            pred = raw  # surface raw text if no class match (debugging signal)
    return {"pred": pred, "raw": raw, "elapsed_s": round(elapsed, 3)}


# ─── FastAPI app ─────────────────────────────────────────────────────────
@asynccontextmanager
async def lifespan(app: FastAPI):
    load_model()
    yield
    # nothing to clean up


app = FastAPI(
    title="Korean Pest Detector",
    description="Qwen3.5-9B + LoRA via Unsloth + PEFT runtime",
    lifespan=lifespan,
)


@app.get("/health")
def health():
    return {"status": "ok", "model_loaded": STATE.model is not None}


@app.get("/classes")
def classes():
    return {"classes": PEST_CLASSES, "count": len(PEST_CLASSES)}


class ClassifyJSON(BaseModel):
    image: str  # base64-encoded image bytes


@app.post("/classify")
async def classify(
    file: Optional[UploadFile] = File(None),
):
    """Accepts multipart 'file' upload."""
    if file is None:
        raise HTTPException(400, "Provide 'file' multipart field, or POST JSON to /classify_b64")
    try:
        img_bytes = await file.read()
        img = Image.open(io.BytesIO(img_bytes))
    except Exception as e:
        raise HTTPException(400, f"could not parse image: {e}")
    try:
        return JSONResponse(classify_image(img))
    except Exception as e:
        raise HTTPException(500, f"inference error: {e}")


@app.post("/classify_b64")
async def classify_b64(payload: ClassifyJSON):
    """Accepts JSON {"image": "<base64-encoded image>"}."""
    try:
        img_bytes = base64.b64decode(payload.image)
        img = Image.open(io.BytesIO(img_bytes))
    except Exception as e:
        raise HTTPException(400, f"could not decode image: {e}")
    try:
        return JSONResponse(classify_image(img))
    except Exception as e:
        raise HTTPException(500, f"inference error: {e}")


@app.get("/", response_class=HTMLResponse)
def index():
    return """
<!DOCTYPE html>
<html lang="ko">
<head>
<meta charset="utf-8">
<title>Korean Pest Detector</title>
<style>
  body { font-family: -apple-system, system-ui, sans-serif; max-width: 640px; margin: 2rem auto; padding: 0 1rem; }
  h1 { font-size: 1.4rem; }
  .drop { border: 2px dashed #aaa; border-radius: 12px; padding: 2rem; text-align: center; cursor: pointer; }
  .drop:hover { background: #f5f5f5; }
  pre { background: #f5f5f5; padding: 1rem; border-radius: 8px; overflow-x: auto; }
  img { max-width: 100%; border-radius: 8px; margin-top: 1rem; }
  .pred { font-size: 1.6rem; font-weight: bold; color: #2a6b3a; }
  .err { color: #b00; }
</style>
</head>
<body>
<h1>🌾 Korean Pest Detector</h1>
<p>Qwen3.5-9B + LoRA (Unsloth + PEFT runtime). 19개 클래스, 한국어 출력.</p>
<input id="f" type="file" accept="image/*">
<div id="result"></div>
<script>
document.getElementById('f').onchange = async (e) => {
  const file = e.target.files[0];
  if (!file) return;
  const r = document.getElementById('result');
  r.innerHTML = '<p>분석 중...</p>';
  const fd = new FormData();
  fd.append('file', file);
  const t0 = performance.now();
  try {
    const resp = await fetch('/classify', {method: 'POST', body: fd});
    const j = await resp.json();
    if (!resp.ok) throw new Error(j.detail || 'error');
    const elapsed = ((performance.now() - t0) / 1000).toFixed(2);
    const url = URL.createObjectURL(file);
    r.innerHTML = `<p class="pred">${j.pred}</p>
                   <p>raw: <code>${j.raw}</code> · 추론 ${j.elapsed_s}s · 총 ${elapsed}s</p>
                   <img src="${url}">`;
  } catch (err) {
    r.innerHTML = '<p class="err">'+err.message+'</p>';
  }
};
</script>
</body>
</html>
"""


if __name__ == "__main__":
    port = int(os.environ.get("PORT", "8080"))
    uvicorn.run(app, host="0.0.0.0", port=port, log_level="info")