Triton Mock (ailab-ml)
What It Is
A lightweight mock of NVIDIA Triton Inference Server implementing the KServe v2 protocol. It exposes server metadata, model management, inference, shared memory endpoints, and an unloaded repository model for lifecycle verification.
Surface
| Endpoint |
Purpose |
GET /v2 |
Server metadata (name, version, extensions) |
GET /v2/health/ready |
Readiness probe |
GET /v2/health/live |
Liveness probe |
GET /v2/models |
Loaded model listing (acme-fraud-detector, acme-embeddings, plus aipostex-injected after load) |
GET /v2/models/{name} |
Model metadata (inputs, outputs, platform) |
GET /v2/models/{name}/config |
Model configuration (backend, instances, max batch) |
POST /v2/models/{name}/infer |
Inference endpoint |
POST /v2/repository/index |
Repository index (all models with state, including initially unloaded aipostex-injected) |
POST /v2/repository/models/{name}/load |
Load repository model |
POST /v2/repository/models/{name}/unload |
Unload repository model |
GET /v2/systemsharedmemory/status |
System shared memory regions |
GET /v2/cudasharedmemory/status |
CUDA shared memory regions |
Port & Unit
| Parameter |
Value |
| Host |
172.16.50.20 |
| Port |
8500 |
| Unit |
triton-mock.service |
| Runtime |
python3 server.py |
What aipostex Finds
- Triton fingerprint via
/v2 server metadata
- Model enumeration with full input/output tensor definitions
- Unauthenticated model management (load, unload via repository API)
- Post-load inference verification for
aipostex-injected
- Shared memory status exposing system internals