feat(video-render): providers/veo.py — Veo 3.1 Vertex AI client (SP-7)

predictLongRunning → fetchPredictOperation 폴링 (12초 × 50).
결과 gs://bucket/veo/{task_id}/sample_0.mp4 → google-cloud-storage SDK로
다운로드 → NAS SMB. GOOGLE_PROJECT_ID/LOCATION/GCS_BUCKET/APPLICATION_CREDENTIALS env.
Plan-B-Video Phase 2.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-19 08:37:45 +09:00
parent b4bec9d51b
commit 8d246b5b32

View File

@@ -0,0 +1,176 @@
"""Veo 3.1 video generation — Google Vertex AI.
POST .../models/{MODEL}:predictLongRunning → POST :fetchPredictOperation 폴링 →
결과 gs://bucket/path/sample_0.mp4 → google-cloud-storage로 다운로드 → NAS SMB.
"""
from __future__ import annotations
import logging
import os
import subprocess
import time
from typing import Optional
import requests
from nas_client import webhook_update_task
logger = logging.getLogger(__name__)
VIDEO_MEDIA_ROOT = os.getenv("VIDEO_MEDIA_ROOT", "/mnt/nas/webpage/data/video")
VIDEO_MEDIA_URL_PREFIX = os.getenv("VIDEO_MEDIA_URL_PREFIX", "/media/video")
POLL_INTERVAL = 12 # Veo는 30~120초 소요
POLL_MAX_ATTEMPTS = 50 # 최대 ~10분
DEFAULT_MODEL = "veo-3.1-fast-generate-001"
def _gcloud_access_token() -> Optional[str]:
"""GOOGLE_APPLICATION_CREDENTIALS service account JSON으로 access token 발행.
google-auth가 컨테이너 안에서 자동 인증 — Bearer 토큰을 GCS SDK가 직접 사용.
REST API 호출용으로는 명시적 token이 필요 → google.auth로 발행.
"""
try:
from google.auth import default as google_default_auth
from google.auth.transport.requests import Request as GoogleAuthRequest
credentials, _ = google_default_auth(
scopes=["https://www.googleapis.com/auth/cloud-platform"],
)
credentials.refresh(GoogleAuthRequest())
return credentials.token
except Exception:
logger.exception("Google credentials refresh 실패")
return None
def _download_gcs(gcs_uri: str, local_path: str) -> bool:
"""gs://bucket/path/file.mp4 → local_path 다운로드. 성공 여부 반환."""
try:
from google.cloud import storage as gcs_storage
if not gcs_uri.startswith("gs://"):
return False
without_scheme = gcs_uri[len("gs://"):]
bucket_name, blob_path = without_scheme.split("/", 1)
client = gcs_storage.Client(project=os.getenv("GOOGLE_PROJECT_ID"))
bucket = client.bucket(bucket_name)
blob = bucket.blob(blob_path)
blob.download_to_filename(local_path)
return True
except Exception:
logger.exception("GCS 다운로드 실패: %s", gcs_uri)
return False
def run_veo_generation(task_id: str, params: dict) -> None:
"""Veo 3.1로 영상 생성 → GCS → NAS SMB → webhook."""
try:
project_id = os.getenv("GOOGLE_PROJECT_ID", "")
location = os.getenv("GOOGLE_LOCATION", "us-central1")
gcs_bucket = os.getenv("GOOGLE_GCS_BUCKET", "")
if not project_id or not gcs_bucket:
webhook_update_task(task_id, "failed", 0, "",
error="GOOGLE_PROJECT_ID 또는 GOOGLE_GCS_BUCKET 미설정")
return
token = _gcloud_access_token()
if not token:
webhook_update_task(task_id, "failed", 0, "",
error="Google access token 발행 실패 (서비스 계정 JSON 확인)")
return
webhook_update_task(task_id, "processing", 5, "Veo API 호출 중...")
model_id = params.get("model") or DEFAULT_MODEL
endpoint_base = (
f"https://{location}-aiplatform.googleapis.com/v1/projects/{project_id}"
f"/locations/{location}/publishers/google/models/{model_id}"
)
headers = {
"Authorization": f"Bearer {token}",
"Content-Type": "application/json",
}
body = {
"instances": [{"prompt": params["prompt"]}],
"parameters": {
"storageUri": f"gs://{gcs_bucket}/veo/{task_id}/",
"sampleCount": 1,
"aspectRatio": params.get("aspect_ratio") or "16:9",
},
}
if params.get("duration"):
body["parameters"]["duration"] = params["duration"]
if params.get("negative_prompt"):
body["parameters"]["negativePrompt"] = params["negative_prompt"]
resp = requests.post(f"{endpoint_base}:predictLongRunning",
headers=headers, json=body, timeout=30)
if resp.status_code != 200:
webhook_update_task(task_id, "failed", 0, "",
error=f"Veo API 오류: {resp.status_code} {resp.text[:300]}")
return
op_name = resp.json().get("name", "")
if not op_name:
webhook_update_task(task_id, "failed", 0, "", error="Veo 응답에 operation name 없음")
return
webhook_update_task(task_id, "processing", 15, f"Veo 작업 시작됨")
# 폴링 — fetchPredictOperation
gcs_uri = None
for attempt in range(POLL_MAX_ATTEMPTS):
time.sleep(POLL_INTERVAL)
fetch = requests.post(
f"{endpoint_base}:fetchPredictOperation",
headers=headers,
json={"operationName": op_name},
timeout=30,
)
if fetch.status_code != 200:
continue
fd = fetch.json()
done = fd.get("done", False)
scaled = min(15 + int((attempt / POLL_MAX_ATTEMPTS) * 65), 79)
webhook_update_task(task_id, "processing", scaled, "Veo 생성 중...")
if done:
if "error" in fd:
webhook_update_task(task_id, "failed", 0, "",
error=f"Veo 작업 실패: {fd['error'].get('message','?')}")
return
videos = (fd.get("response") or {}).get("videos") or []
if not videos:
webhook_update_task(task_id, "failed", 0, "", error="Veo 완료했으나 videos 비어 있음")
return
gcs_uri = videos[0].get("gcsUri", "")
break
else:
webhook_update_task(task_id, "failed", 0, "", error="Veo 폴링 timeout (10분)")
return
if not gcs_uri:
webhook_update_task(task_id, "failed", 0, "", error="Veo 응답에 gcsUri 없음")
return
webhook_update_task(task_id, "processing", 85, "GCS에서 mp4 다운로드 중...")
filename = f"{task_id}.mp4"
os.makedirs(VIDEO_MEDIA_ROOT, exist_ok=True)
file_path = os.path.join(VIDEO_MEDIA_ROOT, filename)
ok = _download_gcs(gcs_uri, file_path)
if not ok:
webhook_update_task(task_id, "failed", 0, "", error=f"GCS 다운로드 실패: {gcs_uri}")
return
local_url = f"{VIDEO_MEDIA_URL_PREFIX}/{filename}"
webhook_update_task(task_id, "succeeded", 100, "Veo 생성 완료", video_url=local_url)
except requests.Timeout:
webhook_update_task(task_id, "failed", 0, "", error="Veo API 타임아웃")
except Exception as e:
logger.exception("Veo generation error task=%s", task_id)
webhook_update_task(task_id, "failed", 0, "", error=str(e))