853cb21542
NVIDIA fix: - Replace parse_vlm_json (requires full schema: global_summary/entities_json/ frame_details/compute_provider) with lightweight _parse_single_frame_json that only extracts per-frame fields (person/action/clothing/etc) - Root cause: NVIDIA adapter does per-frame analysis returning single-frame JSON, but parse_vlm_json rejected it for missing full-schema fields - Verified: task 297 → 6/6 frames parsed successfully, first SUCCESS Gemini + circuit breaker tuning: - Gemini timeout: 30s → 90s (multi-image vision analysis needs more time) - NVIDIA timeout: 20s → 30s (per-frame API call) - Circuit breaker threshold: 3 → 5 (less aggressive tripping) - Circuit breaker cooldown: 600s → 300s (faster recovery)
86 lines
2.7 KiB
YAML
86 lines
2.7 KiB
YAML
# FAM-Edge 配置文件 (Oracle 端) - 多模型池配置
|
||
# Tailscale: Oracle=100.74.137.126, NAS=100.70.234.39
|
||
|
||
# NAS 端回调地址
|
||
nas:
|
||
webhook_url: "http://100.70.234.39:8000/api/core/callback/event"
|
||
media_base_url: "http://100.70.234.39:8000/media"
|
||
media_token: "sentinel-media-2026"
|
||
|
||
# Oracle 端服务
|
||
server:
|
||
host: "0.0.0.0"
|
||
port: 5000
|
||
max_concurrent_tasks: 1
|
||
|
||
# 编排调度模式: fallback(顺序降级, 默认) | ensemble(并行交叉验证)
|
||
orchestrator:
|
||
mode: "fallback"
|
||
overall_timeout: 600
|
||
|
||
# 关键帧筛选参数（自适应：帧数随视频时长动态计算）
|
||
video:
|
||
candidate_per_minute: 2 # 每分钟粗抽候选帧数
|
||
candidate_min: 30 # 候选帧下限（短视频保底）
|
||
candidate_max: 120 # 候选帧上限（超长视频截断）
|
||
key_frame_interval_sec: 150 # 关键帧间隔（秒），每2.5分钟1张
|
||
min_key_frames: 5 # 关键帧下限（帧差不足时补足到此数）
|
||
max_key_frames_floor: 8 # 关键帧上限的下限（短视频保底）
|
||
max_key_frames_cap: 30 # 关键帧上限（超长视频截断）
|
||
mse_threshold: 500
|
||
jpeg_quality: 80
|
||
max_long_edge: 1024
|
||
|
||
# 超时（秒）
|
||
timeout:
|
||
download: 60
|
||
vlm_visual: 600
|
||
vlm_fusion: 300
|
||
callback: 30
|
||
overall: 1800
|
||
|
||
# 多模型池配置（新框架：本地大模型不参与视频分析，仅智能问答兜底）
|
||
#
|
||
# 视频分析链路（推送模式）:
|
||
# 云端 VLM 直接产出结构化 JSON (global_summary / entities_json / frame_details)
|
||
# -> Edge 仅做格式化/校验 (format_cloud_result) -> 直接回写 NAS，无本地融合步骤
|
||
# 视觉角色: Gemini(主) -> NVIDIA NIM(备) 顺序降级; 两云端全失败 -> 任务 FAILED 走重试
|
||
#
|
||
# 智能问答链路:
|
||
# Gemini -> NVIDIA -> 本地 Ollama (仅当两云端都失败才启用本地兜底)
|
||
models:
|
||
- provider: "gemini"
|
||
role: "vision"
|
||
enabled: true
|
||
model_name: "gemini-flash-latest" # v1beta 下 gemini-1.5-flash 会 404
|
||
api_key: "${GEMINI_API_KEY}"
|
||
timeout: 90
|
||
circuit_breaker:
|
||
enabled: true
|
||
threshold: 5
|
||
cooldown: 300
|
||
|
||
- provider: "nvidia"
|
||
role: "vision"
|
||
enabled: true
|
||
model_name: "meta/llama-3.2-11b-vision-instruct"
|
||
base_url: "https://integrate.api.nvidia.com/v1"
|
||
api_key: "${NVIDIA_API_KEY}"
|
||
timeout: 30
|
||
circuit_breaker:
|
||
enabled: true
|
||
threshold: 5
|
||
cooldown: 300
|
||
|
||
# 本地模型：纯文本 qwen2.5:7b，仅参与智能问答，作为 Gemini/NVIDIA 都失败时的兜底
|
||
- provider: "ollama"
|
||
role: "text"
|
||
usage: "qa_fallback"
|
||
enabled: true
|
||
model_name: "qwen2.5:7b"
|
||
base_url: "http://localhost:11434"
|
||
timeout: 120
|
||
num_predict: 512
|
||
circuit_breaker:
|
||
enabled: false
|