log 相关
triton 提供四种等级的 log:info warning error verbose,每一个选项可以分别选择是否打开。--log-verbose <integer>--log-info <boolean>--log-warning <boolean>--log-error <boolean>
模型相关
--model-store 和 --model-repository 一样,指定模型仓库地址,可以多次指定--backend-directory: 指定 backend 的位置--repoagent-directory: 指定 agent 的位置--strict-model-config: true 表示一定要配置文件,false 则表示可以尝试自动填充--model-control-mode: none, poll, explicit 三种--repository-poll-secs: 轮询时长--load-model: 配合 explicit 使用,指定启动的模型--backend-config <<string>,<string>=<string>>: 给某个 backend 加上特定的选项
服务相关
--id: 指定服务器标识符--strict-readiness: true 表示服务器和模型启动完毕才可以访问 /v2/health/ready,false 表示服务器启动即可--allow-http: 开启 HTTP 服务器--allow-grpc: 开启 GRPC 服务器--allow-sagemaker: 监听 Sagemaker 的请求。--allow-metrics: 开启端口提供 prometheus 格式的使用数据--allow-gpu-metrics: 提供 GPU 相关的数据
上面每一种 allow,都有配套的其他选项,比如 http 可以设置处理线程数量,grpc 可设置是否开启 ssl 等,这里就不一一列举了,请直接 –help 查看。
trace 相关
记录一条请求在 triton 中执行的整个过程,记录接受请求,请求处理,排队,计算,发送请求等的时间戳。
--trace-file: 输出的位置--trace-level: 等级,max 会记录更多信息,比如将计算拆解成输入、输出、推理等部分。--trace-rate: 频率,比如多少个请求采样一次
cuda 相关
--pinned-memory-pool-byte-size: 锁页内存的大小,可以加速 host 和 device 的数据传输,默认 256MB--cuda-memory-pool-byte-size <<integer>:<integer>>: 第一个数字 gpu 卡号,第二个数字显存大小,默认 64MB。--min-supported-compute-capability <float>: 最低的 compute capability,小于这个的就不使用。
其他
--exit-on-error: 发送错误的时候退出--exit-timeout-secs <integer>: 当退出服务器的时候,有的请求还没处理完,这个选项指定了超时时间。--rate-limit 和 --rate-limit-resource: 用于限制了请求分发到模型实例上,resource 选项用来分配资源,资源满足就执行--response-cache-byte-size <integer>: 响应缓存的大小。·--buffer-manager-thread-count: 指定线程数量,可以加速在输入输出 tensor 上的操作。--host-policy <<string>,<string>=<string>: NUMA( 多个物理 CPU 共享内存 ) 相关的选项,具体看文档