app.py

import os
from dotenv import load_dotenv
load_dotenv()

import logging

# Fix: Set Hugging Face cache to writable location
# In containerized environments, /.cache may not be writable
if "HF_HOME" not in os.environ:
    os.environ["HF_HOME"] = "/tmp/huggingface"
    print(f"Set HF_HOME to {os.environ['HF_HOME']}")

# Debug/Fix: Unset CUDA_VISIBLE_DEVICES to ensure all GPUs are visible
# Some environments (like HF Spaces) might set this to "0" by default.
if "CUDA_VISIBLE_DEVICES" in os.environ:
    # Use print because logging config might not be set yet
    print(f"Found CUDA_VISIBLE_DEVICES={os.environ['CUDA_VISIBLE_DEVICES']}. Unsetting it to enable all GPUs.")
    del os.environ["CUDA_VISIBLE_DEVICES"]
else:
    print("CUDA_VISIBLE_DEVICES not set. All GPUs should be visible.")

import torch
try:
    print(f"Startup Diagnostics: Torch version {torch.__version__}, CUDA available: {torch.cuda.is_available()}, Device count: {torch.cuda.device_count()}")
except Exception as e:
    print(f"Startup Diagnostics Error: {e}")

import asyncio
import json
import shutil
import tempfile
import time
import uuid
from contextlib import asynccontextmanager
from datetime import timedelta
from pathlib import Path
from typing import Optional

import cv2
from fastapi import BackgroundTasks, FastAPI, File, Form, HTTPException, UploadFile
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import FileResponse, HTMLResponse, JSONResponse, RedirectResponse, StreamingResponse
from fastapi.staticfiles import StaticFiles
import uvicorn

from inference import run_inference, run_grounded_sam2_tracking
from models.depth_estimators.model_loader import list_depth_estimators
from jobs.background import process_video_async
from jobs.models import JobInfo, JobStatus
from jobs.streaming import get_stream, get_stream_event
from jobs.storage import (
    get_depth_output_path,
    get_first_frame_depth_path,
    get_input_video_path,
    get_job_directory,
    get_job_storage,
    get_output_video_path,
)
from models.segmenters.model_loader import get_segmenter_detector
from pydantic import BaseModel
from inspection.router import router as inspection_router

logging.basicConfig(level=logging.INFO)

# Suppress noisy external libraries
logging.getLogger("httpx").setLevel(logging.WARNING)
logging.getLogger("huggingface_hub").setLevel(logging.WARNING)
logging.getLogger("transformers").setLevel(logging.WARNING)

async def _periodic_cleanup() -> None:
    while True:
        await asyncio.sleep(600)
        get_job_storage().cleanup_expired(timedelta(hours=1))


@asynccontextmanager
async def lifespan(_: FastAPI):
    cleanup_task = asyncio.create_task(_periodic_cleanup())
    try:
        yield
    finally:
        cleanup_task.cancel()


app = FastAPI(title="Video Object Detection", lifespan=lifespan)
app.include_router(inspection_router)
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


from fastapi import Request

@app.middleware("http")
async def add_no_cache_header(request: Request, call_next):
    """Ensure frontend assets are not cached by the browser (important for HF Spaces updates)."""
    response = await call_next(request)
    # Apply to all static files and the root page
    if request.url.path.startswith("/app") or request.url.path == "/":
        response.headers["Cache-Control"] = "no-cache, no-store, must-revalidate"
        response.headers["Pragma"] = "no-cache"
        response.headers["Expires"] = "0"
    return response

_FRONTEND_DIR = Path(__file__).with_name("frontend")
if _FRONTEND_DIR.exists():
    app.mount("/app", StaticFiles(directory=_FRONTEND_DIR, html=True), name="frontend")

# Valid detection modes
VALID_MODES = {"object_detection", "segmentation", "drone_detection"}


# ── Chat endpoint ──────────────────────────────────────────────

class ChatRequest(BaseModel):
    message: str
    mission: str = ""
    track_context: dict | None = None
    history: list[dict] = []

@app.post("/chat")
async def chat(req: ChatRequest):
    api_key = os.environ.get("OPENAI_API_KEY")
    if not api_key:
        raise HTTPException(status_code=500, detail="OPENAI_API_KEY not configured")

    import openai
    client = openai.OpenAI(api_key=api_key)

    # Build system prompt
    system_parts = [
        "You are a mission analyst for an ISR (Intelligence, Surveillance, Reconnaissance) platform. "
        "You help operators understand tracked objects, their mission relevance, and assessment results. "
        "Be concise and direct."
    ]

    if req.mission:
        system_parts.append(f"\nCurrent mission objective: {req.mission}")

    if req.track_context:
        tc = req.track_context
        system_parts.append(
            f"\nCurrently selected track context:"
            f"\n- Track ID: {tc.get('id', 'unknown')}"
            f"\n- Label: {tc.get('label', 'unknown')}"
            f"\n- Confidence: {tc.get('score', 'N/A')}"
            f"\n- Mission Relevant: {tc.get('mission_relevant', 'not assessed')}"
            f"\n- Satisfies Mission: {tc.get('satisfies', 'not assessed')}"
            f"\n- Assessment Status: {tc.get('assessment_status', 'UNASSESSED')}"
            f"\n- Reason: {tc.get('reason', 'none')}"
            f"\n- Speed: {tc.get('speed_kph', 'N/A')} kph"
            f"\n- Bounding Box: {tc.get('bbox', 'N/A')}"
        )
        features = tc.get("features", {})
        if features:
            feat_str = "\n".join(f"  - {k}: {v}" for k, v in features.items())
            system_parts.append(f"- Observable Features:\n{feat_str}")
        gpt_raw = tc.get("gpt_raw")
        if gpt_raw:
            system_parts.append(f"- Raw GPT Assessment: {json.dumps(gpt_raw)}")

    system_msg = "\n".join(system_parts)

    # Build messages array
    messages = [{"role": "system", "content": system_msg}]
    for h in req.history[-20:]:
        if h.get("role") in ("user", "assistant") and h.get("content"):
            messages.append({"role": h["role"], "content": h["content"]})
    messages.append({"role": "user", "content": req.message})

    try:
        response = await asyncio.to_thread(
            lambda: client.chat.completions.create(
                model="gpt-4o-mini",
                messages=messages,
                max_tokens=512,
                temperature=0.3,
            )
        )
        return {"response": response.choices[0].message.content}
    except Exception as e:
        logging.exception("Chat endpoint error")
        raise HTTPException(status_code=500, detail=str(e))


def _save_upload_to_tmp(upload: UploadFile) -> str:
    """Save uploaded file to temporary location."""
    suffix = Path(upload.filename or "upload.mp4").suffix or ".mp4"
    fd, path = tempfile.mkstemp(prefix="input_", suffix=suffix, dir="/tmp")
    os.close(fd)
    with open(path, "wb") as buffer:
        data = upload.file.read()
        buffer.write(data)
    return path


def _save_upload_to_path(upload: UploadFile, path: Path) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with open(path, "wb") as buffer:
        data = upload.file.read()
        buffer.write(data)


def _safe_delete(path: str) -> None:
    """Safely delete a file, ignoring errors."""
    try:
        os.remove(path)
    except FileNotFoundError:
        return
    except Exception:
        logging.exception("Failed to remove temporary file: %s", path)


def _schedule_cleanup(background_tasks: BackgroundTasks, path: str) -> None:
    """Schedule file cleanup after response is sent."""
    def _cleanup(target: str = path) -> None:
        _safe_delete(target)

    background_tasks.add_task(_cleanup)


def _default_queries_for_mode(mode: str) -> list[str]:
    if mode == "segmentation":
        return ["object"]
    if mode == "drone_detection":
        return ["drone"]
    return ["person", "car", "truck", "motorcycle", "bicycle", "bus", "train", "airplane"]


@app.get("/", response_class=HTMLResponse)
async def demo_page():
    """Redirect to Mission Console app."""
    return RedirectResponse(url="/app/index.html")


@app.post("/detect")
async def detect_endpoint(
    background_tasks: BackgroundTasks,
    video: UploadFile = File(...),
    mode: str = Form(...),
    queries: str = Form(""),
    detector: str = Form("yolo11"),
    segmenter: str = Form("GSAM2-L"),
    enable_depth: bool = Form(False),
):
    """
    Main detection endpoint.

    Args:
        video: Video file to process
        mode: Detection mode (object_detection, segmentation, drone_detection)
        queries: Comma-separated object classes for object_detection mode
        detector: Model to use (yolo11, detr_resnet50, grounding_dino)
        segmenter: Segmentation model to use (GSAM2-S/B/L, YSAM2-S/B/L)
        enable_depth: Whether to run legacy depth estimation (default: False)
        drone_detection uses the dedicated yolov8_visdrone model.

    Returns:
        - For object_detection: Processed video with bounding boxes
        - For segmentation: Processed video with masks rendered
        - For drone_detection: Processed video with bounding boxes
    """
    # Validate mode
    if mode not in VALID_MODES:
        raise HTTPException(
            status_code=400,
            detail=f"Invalid mode '{mode}'. Must be one of: {', '.join(VALID_MODES)}"
        )

    if mode == "segmentation":
        if video is None:
            raise HTTPException(status_code=400, detail="Video file is required.")

        try:
            input_path = _save_upload_to_tmp(video)
        except Exception:
            logging.exception("Failed to save uploaded file.")
            raise HTTPException(status_code=500, detail="Failed to save uploaded video.")
        finally:
            await video.close()

        fd, output_path = tempfile.mkstemp(prefix="output_", suffix=".mp4", dir="/tmp")
        os.close(fd)

        # Parse queries
        query_list = [q.strip() for q in queries.split(",") if q.strip()]
        if not query_list:
            query_list = ["object"]

        try:
            output_path = run_grounded_sam2_tracking(
                input_path,
                output_path,
                query_list,
                segmenter_name=segmenter,
                num_maskmem=7,
            )
        except ValueError as exc:
            logging.exception("Segmentation processing failed.")
            _safe_delete(input_path)
            _safe_delete(output_path)
            raise HTTPException(status_code=500, detail=str(exc))
        except Exception as exc:
            logging.exception("Segmentation inference failed.")
            _safe_delete(input_path)
            _safe_delete(output_path)
            return JSONResponse(status_code=500, content={"error": str(exc)})

        _schedule_cleanup(background_tasks, input_path)
        _schedule_cleanup(background_tasks, output_path)

        return FileResponse(
            path=output_path,
            media_type="video/mp4",
            filename="segmented.mp4",
        )

    # Handle object detection or drone detection mode
    if video is None:
        raise HTTPException(status_code=400, detail="Video file is required.")

    # Save uploaded video
    try:
        input_path = _save_upload_to_tmp(video)
    except Exception:
        logging.exception("Failed to save uploaded file.")
        raise HTTPException(status_code=500, detail="Failed to save uploaded video.")
    finally:
        await video.close()

    # Create output path
    fd, output_path = tempfile.mkstemp(prefix="output_", suffix=".mp4", dir="/tmp")
    os.close(fd)

    # Parse queries
    detector_name = (detector or "yolov8_visdrone") if mode == "drone_detection" else detector
    query_list = [q.strip() for q in queries.split(",") if q.strip()] or _default_queries_for_mode(mode)

    if mode == "drone_detection" and not query_list:
        query_list = ["drone"]

    # Run inference
    try:
        
        # Determine depth estimator
        active_depth = "depth" if enable_depth else None

        output_path, _ = run_inference(
            input_path,
            output_path,
            query_list,
            detector_name=detector_name,
            depth_estimator_name=active_depth,
            depth_scale=25.0,
        )
    except ValueError as exc:
        logging.exception("Video processing failed.")
        _safe_delete(input_path)
        _safe_delete(output_path)
        raise HTTPException(status_code=500, detail=str(exc))
    except Exception as exc:
        logging.exception("Inference failed.")
        _safe_delete(input_path)
        _safe_delete(output_path)
        return JSONResponse(status_code=500, content={"error": str(exc)})

    # Schedule cleanup
    _schedule_cleanup(background_tasks, input_path)
    _schedule_cleanup(background_tasks, output_path)

    # Return processed video
    response = FileResponse(
        path=output_path,
        media_type="video/mp4",
        filename="processed.mp4",
    )
    return response


@app.post("/detect/async")
async def detect_async_endpoint(
    video: UploadFile = File(...),
    mode: str = Form(...),
    queries: str = Form(""),
    detector: str = Form("yolo11"),
    segmenter: str = Form("GSAM2-L"),
    depth_estimator: str = Form("depth"),
    depth_scale: float = Form(25.0),
    enable_depth: bool = Form(False),
    step: int = Form(7),
    mission: str = Form(None),
):
    _ttfs_t0 = time.perf_counter()

    if mode not in VALID_MODES:
        raise HTTPException(
            status_code=400,
            detail=f"Invalid mode '{mode}'. Must be one of: {', '.join(VALID_MODES)}",
        )

    if video is None:
        raise HTTPException(status_code=400, detail="Video file is required.")

    job_id = uuid.uuid4().hex
    job_dir = get_job_directory(job_id)
    input_path = get_input_video_path(job_id)
    output_path = get_output_video_path(job_id)
    depth_output_path = get_depth_output_path(job_id)
    first_frame_depth_path = get_first_frame_depth_path(job_id)

    try:
        _save_upload_to_path(video, input_path)
    except Exception:
        logging.exception("Failed to save uploaded file.")
        raise HTTPException(status_code=500, detail="Failed to save uploaded video.")
    finally:
        await video.close()

    logging.info("[TTFS:%s] +%.1fs upload_saved", job_id, time.perf_counter() - _ttfs_t0)

    # --- Query Parsing ---
    detector_name = detector
    if mode == "drone_detection":
        detector_name = detector or "yolov8_visdrone"
    elif mode == "segmentation":
        # Segmenter registry owns detector selection (GSAM2→GDINO, YSAM2→YOLO).
        # detector_name=None so the job doesn't forward it (avoids duplicate kwarg).
        try:
            get_segmenter_detector(segmenter)
        except ValueError as exc:
            raise HTTPException(status_code=400, detail=str(exc))
        detector_name = None

    query_list = [q.strip() for q in queries.split(",") if q.strip()] or _default_queries_for_mode(mode)

    logging.info("[TTFS:%s] +%.1fs queries_parsed", job_id, time.perf_counter() - _ttfs_t0)

    available_depth_estimators = set(list_depth_estimators())
    if depth_estimator not in available_depth_estimators:
        raise HTTPException(
            status_code=400,
            detail=(
                f"Invalid depth estimator '{depth_estimator}'. "
                f"Must be one of: {', '.join(sorted(available_depth_estimators))}"
            ),
        )

    # Determine active depth estimator (Legacy)
    active_depth = depth_estimator if enable_depth else None

    job = JobInfo(
        job_id=job_id,
        status=JobStatus.PROCESSING,
        mode=mode,
        queries=query_list,
        detector_name=detector_name,
        segmenter_name=segmenter,
        input_video_path=str(input_path),
        output_video_path=str(output_path),
        depth_estimator_name=active_depth,
        depth_scale=float(depth_scale),
        depth_output_path=str(depth_output_path),
        first_frame_depth_path=str(first_frame_depth_path),
        step=step,
        ttfs_t0=_ttfs_t0,
        mission=mission,
    )
    get_job_storage().create(job)
    asyncio.create_task(process_video_async(job_id))

    return {
        "job_id": job_id,
        "status_url": f"/detect/status/{job_id}",
        "video_url": f"/detect/video/{job_id}",
        "depth_video_url": f"/detect/depth-video/{job_id}",
        "stream_url": f"/detect/stream/{job_id}",
        "status": job.status.value,
    }


@app.get("/detect/status/{job_id}")
async def detect_status(job_id: str):
    job = get_job_storage().get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found or expired.")
    return {
        "job_id": job.job_id,
        "status": job.status.value,
        "created_at": job.created_at.isoformat(),
        "completed_at": job.completed_at.isoformat() if job.completed_at else None,
        "error": job.error,
    }


@app.get("/detect/tracks/{job_id}/summary")
async def get_track_summary_endpoint(job_id: str):
    """Return per-frame detection counts for timeline heatmap."""
    from jobs.storage import get_track_summary, get_job_storage
    import cv2

    job = get_job_storage().get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found")

    summary = get_track_summary(job_id)

    total_frames = 0
    fps = 30.0
    video_path = job.output_video_path
    if video_path:
        cap = cv2.VideoCapture(video_path)
        if cap.isOpened():
            total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
            fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
            cap.release()

    if total_frames == 0 and summary:
        total_frames = max(summary.keys()) + 1

    return {
        "total_frames": total_frames,
        "fps": fps,
        "frames": summary,
    }


@app.get("/detect/tracks/{job_id}/{frame_idx}")
async def get_frame_tracks(job_id: str, frame_idx: int):
    """Retrieve detections (with tracking info) for a specific frame."""
    # This requires us to store detections PER FRAME in JobStorage or similar.
    # Currently, inference.py returns 'sorted_detections' at the end.
    # But during streaming, where is it?
    # We can peek into the 'stream_queue' logic or we need a shared store.
    # Ideally, inference should write to a map/db that we can read.
    
    # Quick fix: If job is done, we might have it. If running, it's harder absent a DB.
    # BUT, 'stream_queue' sends frames.
    # Let's use a global cache in memory for active jobs?
    # See inference.py: 'all_detections_map' is local to that function.
    
    # BETTER APPROACH for this demo:
    # Use a simple shared dictionary in jobs/storage.py or app.py used by inference.
    # We will pass a callback or shared dict to run_inference.
    
    # For now, let's just return 404 if not implemented, but I need to implement it.
    # I'll add a cache in app.py for active job tracks?
    from jobs.storage import get_track_data
    data = get_track_data(job_id, frame_idx)
    return data or []


@app.delete("/detect/job/{job_id}")
async def cancel_job(job_id: str):
    """Cancel a running job."""
    job = get_job_storage().get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found or expired.")
    if job.status != JobStatus.PROCESSING:
        return {
            "message": f"Job already {job.status.value}",
            "status": job.status.value,
        }

    get_job_storage().update(job_id, status=JobStatus.CANCELLED)
    return {
        "message": "Job cancellation requested",
        "status": "cancelled",
    }


@app.get("/detect/video/{job_id}")
async def detect_video(job_id: str):
    job = get_job_storage().get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found or expired.")
    if job.status == JobStatus.FAILED:
        raise HTTPException(status_code=500, detail=f"Job failed: {job.error}")
    if job.status == JobStatus.CANCELLED:
        raise HTTPException(status_code=410, detail="Job was cancelled")
    if job.status == JobStatus.PROCESSING:
        return JSONResponse(
            status_code=202,
            content={"detail": "Video still processing", "status": "processing"},
        )
    if not job.output_video_path or not Path(job.output_video_path).exists():
        raise HTTPException(status_code=404, detail="Video file not found.")
    return FileResponse(
        path=job.output_video_path,
        media_type="video/mp4",
        filename="processed.mp4",
    )


@app.get("/detect/depth-video/{job_id}")
async def detect_depth_video(job_id: str):
    """Return depth estimation video."""
    job = get_job_storage().get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found or expired.")
    if not job.depth_output_path:
        # Check if depth failed (partial success)
        if job.partial_success and job.depth_error:
            raise HTTPException(status_code=404, detail=f"Depth unavailable: {job.depth_error}")
        raise HTTPException(status_code=404, detail="No depth video for this job.")
    if job.status == JobStatus.FAILED:
        raise HTTPException(status_code=500, detail=f"Job failed: {job.error}")
    if job.status == JobStatus.CANCELLED:
        raise HTTPException(status_code=410, detail="Job was cancelled")
    if job.status == JobStatus.PROCESSING:
        return JSONResponse(
            status_code=202,
            content={"detail": "Video still processing", "status": "processing"},
        )
    if not Path(job.depth_output_path).exists():
        raise HTTPException(status_code=404, detail="Depth video file not found.")
    return FileResponse(
        path=job.depth_output_path,
        media_type="video/mp4",
        filename="depth.mp4",
    )


@app.get("/detect/stream/{job_id}")
async def stream_video(job_id: str):
    """MJPEG stream of the processing video (event-driven)."""
    import queue as queue_mod

    async def stream_generator():
        STREAM_FPS = 24
        FRAME_INTERVAL = 1.0 / STREAM_FPS  # ~41.7ms

        loop = asyncio.get_running_loop()
        buffered = False
        last_yield_time = 0.0

        # TTFS instrumentation
        _first_yielded = False
        _buffer_wait_logged = False
        _job = get_job_storage().get(job_id)
        _stream_t0 = _job.ttfs_t0 if _job else None

        if _stream_t0:
            logging.info("[TTFS:%s] +%.1fs stream_subscribed", job_id, time.perf_counter() - _stream_t0)

        # Get or create the asyncio.Event for this stream (must be in async context)
        event = get_stream_event(job_id)

        # Hold a local ref to the queue so we can drain it even after remove_stream()
        q = get_stream(job_id)
        if not q:
            return

        stream_removed = False

        while True:
            try:
                # Initial Buffer: Wait until we have enough frames or job is done
                if not buffered:
                    if not _buffer_wait_logged and _stream_t0:
                        logging.info("[TTFS:%s] +%.1fs stream_buffer_wait (qsize=%d)", job_id, time.perf_counter() - _stream_t0, q.qsize())
                        _buffer_wait_logged = True
                    if q.qsize() < 5 and not stream_removed:
                        await asyncio.sleep(0.1)
                        stream_removed = get_stream(job_id) is None
                        continue
                    buffered = True
                    if _stream_t0:
                        logging.info("[TTFS:%s] +%.1fs stream_buffer_ready", job_id, time.perf_counter() - _stream_t0)

                # Try to get a frame from the queue first (non-blocking)
                frame = None
                try:
                    frame = q.get_nowait()
                except queue_mod.Empty:
                    pass

                # Only block on event when queue is actually empty
                if frame is None:
                    if stream_removed:
                        break  # stream ended and queue fully drained
                    if event is not None:
                        try:
                            await asyncio.wait_for(event.wait(), timeout=1.0)
                            event.clear()
                        except asyncio.TimeoutError:
                            stream_removed = get_stream(job_id) is None
                            continue
                    else:
                        await asyncio.sleep(FRAME_INTERVAL)
                    # After waking, try the queue again
                    try:
                        frame = q.get_nowait()
                    except queue_mod.Empty:
                        stream_removed = get_stream(job_id) is None
                        continue

                # Pace output at fixed 24fps
                now = time.perf_counter()
                wait = FRAME_INTERVAL - (now - last_yield_time)
                if wait > 0:
                    await asyncio.sleep(wait)

                # Encode in thread pool to avoid blocking the event loop
                encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), 60]
                success, buffer = await loop.run_in_executor(None, cv2.imencode, '.jpg', frame, encode_param)

                if success:
                    last_yield_time = time.perf_counter()
                    if not _first_yielded:
                        _first_yielded = True
                        if _stream_t0:
                            logging.info("[TTFS:%s] +%.1fs first_yield_to_client", job_id, time.perf_counter() - _stream_t0)
                    yield (b'--frame\r\n'
                           b'Content-Type: image/jpeg\r\n\r\n' + buffer.tobytes() + b'\r\n')

            except Exception:
                await asyncio.sleep(0.1)

    return StreamingResponse(
        stream_generator(),
        media_type="multipart/x-mixed-replace; boundary=frame"
    )


@app.post("/benchmark")
async def benchmark_endpoint(
    video: UploadFile = File(...),
    queries: str = Form("person,car,truck"),
    segmenter: str = Form("GSAM2-L"),
    step: int = Form(60),
    num_maskmem: Optional[int] = Form(None),
):
    """Run instrumented GSAM2 pipeline and return latency breakdown JSON.

    This is a long-running synchronous request (may take minutes).
    Callers should set an appropriate HTTP timeout.
    """
    import threading

    # Save uploaded video to temp path
    input_path = tempfile.mktemp(suffix=".mp4", prefix="bench_in_")
    output_path = tempfile.mktemp(suffix=".mp4", prefix="bench_out_")
    try:
        with open(input_path, "wb") as f:
            shutil.copyfileobj(video.file, f)

        query_list = [q.strip() for q in queries.split(",") if q.strip()]

        metrics = {
            "end_to_end_ms": 0.0,
            "frame_extraction_ms": 0.0,
            "model_load_ms": 0.0,
            "init_state_ms": 0.0,
            "tracking_total_ms": 0.0,
            "gdino_total_ms": 0.0,
            "sam_image_total_ms": 0.0,
            "sam_video_total_ms": 0.0,
            "id_reconciliation_ms": 0.0,
            "render_total_ms": 0.0,
            "writer_total_ms": 0.0,
            "gpu_peak_mem_mb": 0.0,
        }
        lock = threading.Lock()

        await asyncio.to_thread(
            run_grounded_sam2_tracking,
            input_path,
            output_path,
            query_list,
            segmenter_name=segmenter,
            step=step,
            _perf_metrics=metrics,
            _perf_lock=lock,
            num_maskmem=num_maskmem,
        )

        # Read frame count and fps from output video
        total_frames = 0
        fps = 0.0
        cap = cv2.VideoCapture(output_path)
        if cap.isOpened():
            total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
            fps = cap.get(cv2.CAP_PROP_FPS) or 0.0
            cap.release()

        num_gpus = torch.cuda.device_count()

        return JSONResponse({
            "total_frames": total_frames,
            "fps": fps,
            "num_gpus": num_gpus,
            "num_maskmem": num_maskmem if num_maskmem is not None else 7,
            "metrics": metrics,
        })

    finally:
        for p in (input_path, output_path):
            try:
                os.remove(p)
            except OSError:
                pass


@app.get("/gpu-monitor")
async def gpu_monitor_endpoint(duration: int = 180, interval: int = 1):
    """Stream nvidia-smi dmon output for the given duration.

    Usage: curl 'http://.../gpu-monitor?duration=180&interval=1'
    Run this in one terminal while /benchmark runs in another.
    """
    import subprocess

    async def _stream():
        proc = subprocess.Popen(
            ["nvidia-smi", "dmon", "-s", "u", "-d", str(interval)],
            stdout=subprocess.PIPE,
            stderr=subprocess.STDOUT,
            text=True,
        )
        try:
            elapsed = 0
            for line in proc.stdout:
                yield line
                if interval > 0:
                    elapsed += interval
                    if elapsed > duration:
                        break
        finally:
            proc.terminate()
            proc.wait()

    return StreamingResponse(_stream(), media_type="text/plain")


# ---------------------------------------------------------------------------
# Benchmark Profiler & Roofline Analysis Endpoints
# ---------------------------------------------------------------------------

@app.get("/benchmark/hardware")
async def benchmark_hardware():
    """Return hardware specs JSON (no video needed, cached)."""
    import dataclasses
    from utils.hardware_info import get_hardware_info

    hw = await asyncio.to_thread(get_hardware_info)
    return JSONResponse(dataclasses.asdict(hw))


@app.post("/benchmark/profile")
async def benchmark_profile(
    video: UploadFile = File(...),
    mode: str = Form("detection"),
    detector: str = Form("yolo11"),
    segmenter: str = Form("GSAM2-L"),
    queries: str = Form("person,car,truck"),
    max_frames: int = Form(100),
    warmup_frames: int = Form(5),
    step: int = Form(60),
    num_maskmem: Optional[int] = Form(None),
):
    """Run profiled inference and return per-frame timing breakdown.

    Args:
        video: Video file to profile.
        mode: "detection" or "segmentation".
        detector: Detector key (for detection mode).
        segmenter: Segmenter key (for segmentation mode).
        queries: Comma-separated object classes.
        max_frames: Maximum frames to profile.
        warmup_frames: Warmup frames (detection only).
        step: Keyframe interval (segmentation only).
        num_maskmem: SAM2 memory frames (None = model default 7).
    """
    import dataclasses
    from utils.profiler import run_profiled_detection, run_profiled_segmentation

    if mode not in ("detection", "segmentation"):
        raise HTTPException(status_code=400, detail="mode must be 'detection' or 'segmentation'")

    input_path = _save_upload_to_tmp(video)
    await video.close()

    query_list = [q.strip() for q in queries.split(",") if q.strip()]

    try:
        if mode == "detection":
            result = await asyncio.to_thread(
                run_profiled_detection,
                input_path, detector, query_list,
                max_frames=max_frames, warmup_frames=warmup_frames,
            )
        else:
            result = await asyncio.to_thread(
                run_profiled_segmentation,
                input_path, segmenter, query_list,
                max_frames=max_frames, step=step,
                num_maskmem=num_maskmem,
            )
    except Exception as exc:
        _safe_delete(input_path)
        logging.exception("Profiling failed")
        raise HTTPException(status_code=500, detail=str(exc))
    finally:
        _safe_delete(input_path)

    # Serialize dataclass, handling any non-serializable fields
    out = dataclasses.asdict(result)
    # Include GSAM2 metrics if present
    gsam2 = getattr(result, "_gsam2_metrics", None)
    if gsam2:
        out["gsam2_metrics"] = gsam2
    return JSONResponse(out)


@app.post("/benchmark/analysis")
async def benchmark_analysis(
    video: UploadFile = File(...),
    mode: str = Form("detection"),
    detector: str = Form("yolo11"),
    segmenter: str = Form("GSAM2-L"),
    queries: str = Form("person,car,truck"),
    max_frames: int = Form(100),
    warmup_frames: int = Form(5),
    step: int = Form(60),
    num_maskmem: Optional[int] = Form(None),
):
    """Full roofline analysis: hardware + profiling + theoretical ceilings + bottleneck ID.

    Combines hardware extraction, profiled inference, and roofline model
    to identify bottlenecks and provide actionable recommendations.
    """
    import dataclasses
    from utils.hardware_info import get_hardware_info
    from utils.profiler import run_profiled_detection, run_profiled_segmentation
    from utils.roofline import compute_roofline

    if mode not in ("detection", "segmentation"):
        raise HTTPException(status_code=400, detail="mode must be 'detection' or 'segmentation'")

    input_path = _save_upload_to_tmp(video)
    await video.close()

    query_list = [q.strip() for q in queries.split(",") if q.strip()]

    try:
        # Get hardware info (cached, fast)
        hardware = await asyncio.to_thread(get_hardware_info)

        # Run profiling
        if mode == "detection":
            profiling = await asyncio.to_thread(
                run_profiled_detection,
                input_path, detector, query_list,
                max_frames=max_frames, warmup_frames=warmup_frames,
            )
        else:
            profiling = await asyncio.to_thread(
                run_profiled_segmentation,
                input_path, segmenter, query_list,
                max_frames=max_frames, step=step,
                num_maskmem=num_maskmem,
            )

        # Compute roofline
        roofline = compute_roofline(hardware, profiling)

    except Exception as exc:
        _safe_delete(input_path)
        logging.exception("Benchmark analysis failed")
        raise HTTPException(status_code=500, detail=str(exc))
    finally:
        _safe_delete(input_path)

    return JSONResponse({
        "hardware": dataclasses.asdict(hardware),
        "profiling": dataclasses.asdict(profiling),
        "roofline": dataclasses.asdict(roofline),
    })


if __name__ == "__main__":
    uvicorn.run("app:app", host="0.0.0.0", port=7860, reload=False)