HuggingFace Hub Inference Jobs API.
Jobs let you run arbitrary Docker containers or HF Spaces as serverless
compute tasks. Billing is per-second based on the flavor (hardware tier).
Usage
{:ok, job} = HuggingfaceClient.Hub.Jobs.run("my-org",
docker_image: "pytorch/pytorch:2.3.0-cuda12.1-cudnn8-runtime",
flavor: "gpu-t4-medium",
command: ["python", "train.py"],
env: %{"EPOCHS" => "10"},
access_token: System.get_env("HF_TOKEN")
)
# Poll until done
{:ok, result} = HuggingfaceClient.Hub.Jobs.wait(job["id"],
access_token: System.get_env("HF_TOKEN")
)Summary
Functions
Cancels a running job.
Fetches the current status of a job by job_id.
Lists all jobs under namespace.
Returns the available hardware flavors for running jobs.
Submits a new inference job under namespace (user or org slug).
Polls a job until it reaches a terminal state ("completed" or "failed").
Functions
@spec cancel(String.t(), String.t(), keyword()) :: {:ok, map()} | {:error, Exception.t()}
Cancels a running job.
@spec get(String.t(), String.t(), keyword()) :: {:ok, map()} | {:error, Exception.t()}
Fetches the current status of a job by job_id.
@spec list( String.t(), keyword() ) :: {:ok, list()} | {:error, Exception.t()}
Lists all jobs under namespace.
@spec list_flavors(keyword()) :: {:ok, list()} | {:error, Exception.t()}
Returns the available hardware flavors for running jobs.
@spec run( String.t(), keyword() ) :: {:ok, map()} | {:error, Exception.t()}
Submits a new inference job under namespace (user or org slug).
Required option (one of):
:docker_image— Docker image to run (e.g."pytorch/pytorch:latest"):space_id— HF Space to run as a job (e.g."my-org/my-space")
Optional options
:flavor— hardware tier:"cpu-basic","gpu-t4-medium", etc.:command— override default command, list of strings:env— map of environment variables:secrets— map of secret names to values:timeout— max runtime seconds:access_token
@spec wait(String.t(), String.t(), keyword()) :: {:ok, map()} | {:error, Exception.t()}
Polls a job until it reaches a terminal state ("completed" or "failed").
Returns {:ok, job_map} when done or {:error, reason} on timeout/failure.
Options
:timeout_s— max total wait in seconds (default: 600):interval_ms— polling interval ms (default: 5_000):namespace— job namespace (required unless included injob_id):access_token