ProtFaiss/config.yml at main · CECoffee/ProtFaiss · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
# 可调用 POST /admin/reload-config 立即强制重载。

gpu:
  # 主开关：false = 单卡模式（GPU 0），适用于单卡机器或调试
  multi_gpu_enabled: false
  # ESM2 编码使用的设备："auto"（选显存最多的 GPU）| "cuda:0" | "cpu"
  encoding_device: "auto"
  # FAISS 使用的 GPU 列表："auto"（所有可用 GPU）| [0, 1, 2]
  faiss_devices: "auto"
  # FAISS StandardGpuResources 临时显存限制（MB）
  faiss_temp_memory_mb: 1500
  # GPU 显存预留（MB），低于此值时不选该 GPU
  memory_reserve_mb: 500
  # IVF-PQ 检索时使用 FP16 查找表（LUT）：降低显存占用，在支持 FP16 的 GPU 上可提速
  # 代价是轻微精度损失；仅对 IVF-PQ 索引有效，Flat/HNSW 忽略此项
  fp16_lut: true

search:
  faiss_search_workers: 8
  max_concurrent_encodings: 3
  threadpool_workers: 32
  faiss_nprobe: 8
  # 闲时自动释放显存：检索完成后等待此秒数再卸载该用户的 FAISS index（0 = 禁用）
  time_release_vram: 30

build:
  encoding_batch_size: 32
  max_per_shard: 500000
  db_batch_size: 500
  # IVF-PQ 参数
  ivfpq_nlist: 256
  ivfpq_m: 48 # 5080极限
  ivfpq_nbits: 8
  # HNSW 参数
  hnsw_m: 32
  hnsw_ef_construction: 200
  # 每次向 GPU 索引 add 的向量数（避免 OOM）
  add_batch_size: 200000

scheduler:
  # 系统 GPU 总 slot 数（逻辑限流，可小于物理 GPU 数）
  total_gpu_slots: 1
  # 默认用户 GPU 配额（超管可修改）
  default_user_gpu_quota: 1
  # 调度循环间隔（秒）
  poll_interval: 0.5
  # search 任务 GPU 占用超时（秒）
  search_gpu_timeout: 60
  # build 任务 GPU 占用超时（秒）
  build_gpu_timeout: 86400
  # fair-share 每日衰减因子
  decay_factor: 0.95
  # search 基础优先级（越小越优先）
  search_base_priority: 10
  # build 基础优先级
  build_base_priority: 100
  # 同时驻留 GPU 的数据集 FAISS 缓存数
  max_cached_datasets: 4

daemon:
  # IPC TCP server address (127.0.0.1 for local dev; 0.0.0.0 in K8s/cluster)
  ipc_host: "0.0.0.0"
  # IPC TCP port (Windows 动态保留范围 9788-9887)
  ipc_port: 9002
  # PID file directory
  pid_dir: "~/.protfaiss"

api:
  # HTTP server bind address
  host: "0.0.0.0"
  # HTTP server port
  port: 8000
  # Number of persistent IPC connections to the daemon
  ipc_pool_size: 4

# PostgreSQL database connection
database:
  host: "localhost"
  port: 5432
  dbname: "protein_db"
  user: "postgres"
  # 生产环境推荐通过 DB_PASSWORD 环境变量传入，避免明文写入配置文件
  password: "0909"
  max_connections: 20

# Redis — shared task state store (required in cluster mode)
redis:
  host: "localhost"
  port: 6379
  db: 0
  password: ""

# Shared storage — must be NFS-mounted on all nodes in cluster mode
storage:
  # Overrides DATASETS_ROOT from core.config if set (leave empty for legacy path)
  datasets_root: ""
  # Overrides ESM2_MODEL_DIR from core.config if set (leave empty for legacy path)
  models_root: ""

# Distributed cluster configuration
cluster:
  # false = single-node legacy mode; true = multi-node worker dispatch mode
  enabled: true
  # Control plane IPC address that workers connect to for registration/heartbeats
  # (same as daemon.ipc_host/ipc_port — listed here for clarity)
  control_plane_host: "127.0.0.1"
  control_plane_port: 9002
  # Task result TTL in Redis (seconds)
  task_ttl: 600
  # Worker heartbeat interval (seconds)
  heartbeat_interval: 5
  # Worker considered dead after this many seconds without a heartbeat
  heartbeat_timeout: 15
  # Retries when control plane fails to reach a worker before deregistering it
  retry_num: 5

# GPU Worker node configuration (used when running app.worker)
worker:
  # Address this worker listens on for task dispatch from the control plane
  host: "0.0.0.0"
  port: 9100
  # Human-readable node identifier (defaults to hostname if empty)
  node_id: "win"