hf-image-api / worker.py
kajalan's picture
Upload 14 files (#1)
5738ae5
Raw History Blame Contribute Delete
7.64 kB
"""
worker.py β€” Single background thread that processes generation jobs one at a time.
Architecture:
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” queue.put(job_id) β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ FastAPI endpoint β”‚ ────────────────────────────► β”‚ _job_queue (Queue) β”‚
β”‚ POST /generate β”‚ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β”‚ .get()
β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ worker_loop() β”‚
β”‚ (daemon thread) β”‚
β”‚ runs _process_job() β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
Only one job runs at a time β€” CPU inference is single-threaded anyway.
"""
from __future__ import annotations
import logging
import os
import queue
import threading
from datetime import datetime, timezone
from typing import List
logger = logging.getLogger("worker")
# Internal FIFO queue β€” contains job_ids as strings
_job_queue: queue.Queue[str] = queue.Queue()
# Lock used to inspect queue contents safely
_queue_lock = threading.Lock()
# Ordered list of pending job IDs (mirrors the queue for position lookups)
_pending: List[str] = []
_pending_lock = threading.Lock()
# ── Public helpers (called from FastAPI routes) ────────────────────────────────
def enqueue(job_id: str) -> None:
"""Add a job to the processing queue."""
with _pending_lock:
_pending.append(job_id)
_job_queue.put(job_id)
logger.info("Enqueued job %s (queue size: %d)", job_id, _job_queue.qsize())
def queue_position(job_id: str) -> int:
"""Return the 1-based position of job_id in the pending queue (0 if not found)."""
with _pending_lock:
try:
return _pending.index(job_id) + 1
except ValueError:
return 0
def queue_size() -> int:
"""Number of jobs still waiting (does not include the currently running job)."""
return _job_queue.qsize()
# ── Core processing ────────────────────────────────────────────────────────────
def _process_job(job_id: str) -> None:
"""Generate an image for one job. All state mutations go through job_store."""
from config import settings
from job_store import job_store
from pipeline_manager import pipeline_mgr
from schemas import JobStatus
job = job_store.get(job_id)
if not job:
logger.warning("Job %s not found in store β€” skipping.", job_id)
return
# Mark as processing
job_store.update(
job_id,
status=JobStatus.PROCESSING,
started_at=datetime.now(timezone.utc),
progress=0,
)
logger.info("Processing job %s prompt=%r", job_id, job.prompt[:60])
try:
# ── Ensure pipeline is loaded ──────────────────────────────────────────
if not pipeline_mgr.is_loaded:
logger.info("Pipeline not loaded β€” loading now (job %s will wait).", job_id)
pipeline_mgr.load()
p = job.params
# ── Step-progress callback ─────────────────────────────────────────────
def on_step(step: int, total: int) -> None:
pct = int(step / total * 95) # reserve last 5 % for save
job_store.update(job_id, progress=pct, current_step=step, total_steps=total)
# ── Inference ─────────────────────────────────────────────────────────
image, used_seed = pipeline_mgr.generate(
prompt=job.prompt,
negative_prompt=job.negative_prompt,
width=p.get("width", settings.default_width),
height=p.get("height", settings.default_height),
num_inference_steps=p.get("num_inference_steps", settings.default_steps),
guidance_scale=p.get("guidance_scale", settings.default_guidance_scale),
seed=p.get("seed", -1),
scheduler_name=p.get("scheduler", "dpm++"),
step_callback=on_step,
)
# ── Save image ────────────────────────────────────────────────────────
job_store.update(job_id, progress=97)
timestamp = datetime.now(timezone.utc).strftime("%Y%m%d_%H%M%S")
filename = f"{timestamp}_{job_id[:8]}_{used_seed}.png"
abs_path = os.path.join(settings.outputs_dir, filename)
image.save(abs_path, format="PNG", optimize=False)
# ── Compute duration ───────────────────────────────────────────────────
now = datetime.now(timezone.utc)
started = job_store.get(job_id).started_at
duration = (now - started).total_seconds() if started else None
job_store.update(
job_id,
status=JobStatus.COMPLETED,
completed_at=now,
progress=100,
image_filename=filename,
download_url=f"/jobs/{job_id}/download",
duration_seconds=round(duration, 2) if duration else None,
)
logger.info(
"Job %s completed in %.1f s β†’ %s",
job_id,
duration or 0,
filename,
)
except Exception as exc:
logger.exception("Job %s failed: %s", job_id, exc)
job_store.update(
job_id,
status=JobStatus.FAILED,
completed_at=datetime.now(timezone.utc),
error=str(exc),
progress=0,
)
# ── Worker loop (runs in its own daemon thread) ────────────────────────────────
def worker_loop() -> None:
"""Infinite loop: pull job IDs from the queue and process them."""
logger.info("Worker thread started.")
while True:
job_id = _job_queue.get() # blocks until a job arrives
with _pending_lock:
try:
_pending.remove(job_id)
except ValueError:
pass
try:
_process_job(job_id)
except Exception as exc:
# Catch-all safety net β€” the loop must never crash
logger.exception("Unhandled error in worker for job %s: %s", job_id, exc)
finally:
_job_queue.task_done()
def start_worker() -> threading.Thread:
"""Spawn and return the daemon worker thread. Call once at application startup."""
t = threading.Thread(target=worker_loop, name="image-worker", daemon=True)
t.start()
return t