Skip to main content

Models API

Prefix: /api/v1/models · UI: Model Management

Model lifecycle

MethodPathDescription
POST/downloadDownload a model from HuggingFace (202); body: repo_id, quantization, trust_remote_code
POST/{id}/redownloadRe-download / re-quantize an existing model (202)
DELETE/{id}/cancelCancel an in-progress download (revokes the running task)
GET``List models
GET/local-cache/listEnumerate models present in the local HF cache
GET/{id}Get model details
GET/{id}/architectureDiscovered layer/hook structure (dynamic layer discovery output)
PATCH/{id}Update model metadata
DELETE/{id}Delete model (204). Returns 409 if a training references it
GET/tasks/{task_id}Raw Celery task status

Activation extraction (Stage 1)

These run the base model over a tokenized dataset and cache raw activations — see the extraction pipeline for how this differs from SAE feature extraction.

MethodPathDescription
POST/{id}/estimate-extractionEstimate VRAM/disk for an extraction config before running it
POST/{id}/extract-activationsStart activation extraction (202); config includes layers, hook types, and optional gpu_id
GET/{id}/extractionsList the model's activation extractions
GET/{id}/extractions/activeThe currently running extraction, if any (200 with null data when idle)
POST/{id}/extractions/{eid}/cancelCancel a running extraction
POST/{id}/extractions/{eid}/retryRetry a failed extraction
DELETE/{id}/extractionsDelete the model's extraction records/artifacts

Progress channels: models/{id}/progress (events model:progress|completed|error) and models/{id}/extraction.