Skip to content

Latest commit

 

History

History
64 lines (54 loc) · 4.93 KB

api-reference.md

File metadata and controls

64 lines (54 loc) · 4.93 KB

Public Devmid API coverage

The client supports the following Devmid user operations. Responses remain dictionaries to preserve new fields and the distinction between omitted and null values.

HTTP Path SDK method CLI
POST /auth/login authenticate(username, password) auth login
POST /auth/refresh refresh_access_token() auth refresh
GET /models get_models() models list
GET /models/{model} get_model(model_name) models details NAME
DELETE /models/{model} delete_model(model_name) models delete NAME
POST /models/load register_model(...) models register
POST /models/{model}/scan/retry retry_model_scan(model_name) models retry-scan NAME
GET /ssh-keys get_ssh_keys() ssh list
POST /ssh-keys add_ssh_key(title, key) ssh add
DELETE /ssh-keys/{keyId} delete_ssh_key(key_id) ssh delete ID
GET /inference/gpus get_available_gpus() inference gpus
POST /inference/load load_model_for_inference(...) inference load NAME
POST /inference/unload unload_model_from_inference(session_id) inference unload ID
GET /inference/loaded-models get_loaded_models() inference loaded / status
GET /inference/user-status get_user_inference_status(...) inference user-status
GET /inference/sessions get_inference_sessions(...) inference sessions
GET /runtime/status get_runtime_status(session_id) inference session-status ID
GET /runtime/billing get_runtime_billing(session_id) inference billing ID

All routes except login require Keycloak bearer authentication. Read/write scopes are declared by the public API. User-status requires model:write even though it is GET; refresh declares both model:read and model:write.

History supports date, model, status and cursor filters plus optional scaling summaries. Runtime uses the exact query name model_inference_id. Lookup/deletion/retry use model names; unload and runtime use positive numeric session IDs. SSH-key IDs are positive integers.

Higher-level helpers: wait_for_scan, wait_for_inference, load_and_wait, iter_inference_sessions, get_usage_summary, and export_sessions_csv. GPUPlanConfig validates autoscaling bounds. Load accepts every writable public vLLM field and rejects server-controlled/unknown fields.

To use a different model revision, register it under a new name and retire the old registration when ready. update_model_revision raises NotImplementedError; there is no corresponding CLI command.

Additional helpers

Capability SDK CLI Usage
Full user overview get_dashboard_snapshot inference overview Models, GPUs, loaded deployments and windowed usage
Exact owned history row get_inference_session Via inference details Look up one session by ID
Combined session details get_session_details inference details ID Session history, runtime state and billing
Inline scaling include_scaling=True on loaded/history methods --include-scaling Requires server support
Scaling detail/pages get_scaling_history, iter_scaling_history inference scaling ID Requires the server's public OpenAPI schema
SGLang/LlamaCpp and GGUF options engine, engine_config, deployment_options --engine, --engine-config, --deployment-options Uses configuration fields declared by the server
Autoscaling targets/delays GPUPlanConfig Target/delay load options Requires server support
Complete nested JSON export export_sessions_json sessions --json-file Export all matching history pages

Engine configuration and scaling-detail discovery require a public OpenAPI schema from the server you use. The server controls feature availability and access to models and sessions. HTTP 401/403/404 errors are returned to the caller.

See custom models, usage and inference for lifecycle, billing and serving API details.