| method | path | purpose |
|---|---|---|
| POST | /v1/infer | one image; return=json, txt, coco or annotated JPEG |
| POST | /v1/infer/batch | several files in one multipart body, independent batch-1 jobs |
| POST | /v1/video | video upload, NDJSON stream of per-frame results |
| WS | /v1/stream | binary frames in, JSON per frame out, keep-latest backpressure |
| GET | /v1/models, /v1/models/{id}/load, /unload | model registry |
| GET | /v1/stats, /metrics | rolling latency percentiles, Prometheus exposition |
| GET | /healthz, /readyz | liveness and readiness |
Responses carry per-stage timings (queue, decode, pre, infer, post). Overload returns 503 with Retry-After; deadlines return 504; shutdown drains in-flight work.
| model | backend | model time | request p50 (1 client) | throughput (8 clients, 4 workers) |
|---|---|---|---|---|
| v0.1-N fp16 | TensorRT | 1.8 ms | 7.7 ms | 379 img/s |
| EsMoE-N fp16 | TensorRT | 1.6 ms | 6.9 ms | 359 img/s |
| v0.1-N fp32 | ONNX Runtime CUDA | 4.3 ms | 9.8 ms | 378 img/s |
| v0.1-N fp32 | ncnn CPU (6 threads) | 106 ms | 118 ms |
HTTP overhead is about 1 ms per request; detections are numerically identical to the bare CLI runtime. Full tables: API_SERVER_RESULTS.md in the repository.
/opt/yolomaster/server.json lists the models (id, path, backend, device, precision, threads, workers, conf, iou). Mount your own over it, or pass -m id=path,backend=trt,device=cuda,precision=fp16 on the command line. Extra models go under
a mounted /opt/yolomaster/models/. Full reference: docs/API.md and deploy/README.md in the repository.
Base nvidia/cuda:12.9.1-cudnn-runtime-ubuntu22.04; needs the NVIDIA container toolkit for GPU use. Includes the yolomaster_edge CLI (docker run ... cli --help). Built with Bazel and rules_oci. Third-party licenses under
/opt/yolomaster/LICENSES/. Version 1.2.0 tracks yolo-master-edge main.
The numbers are the MEASURED ones from the results document; nothing in it claims upstream capabilities as ours beyond the model names.
Content type
Image
Digest
sha256:65771d282…
Size
4.9 GB
Last updated
10 days ago
docker pull skywalker0501/yolomaster-api