Python SDK#
The Python SDK is the application entry point for automating MEGA Hub from Python. It talks to the same Hub API used by the web application and CLI; it does not proxy repository operations through Hugging Face.
Install the package and configure an access token:
pip install megatensorsexport MEGA_TOKEN=YOUR_MEGA_TOKENSet MEGA_ENDPOINT only for a staging or self-hosted Hub. Production defaults
to https://mega.tensorplay.cn.
Native Hub client#
MegaHubClient is the typed client for the current MEGA public contract:
from megatensors.hub import MegaHubClient client = MegaHubClient()client.create_repo("research/demo", repo_type="model", private=True, exist_ok=True)client.upload_file("research/demo", "./config.json", path_in_repo="config.json")Use it for repository files and revisions, discussions, webhooks, account keys, Storage Buckets, and Jobs. The complete REST request and response shapes are in the Hub API and the live OpenAPI Explorer.
Hub-compatible imports#
megatensors.mega_hub provides the evaluated Hugging Face Hub import surface
while retaining MEGA transport semantics:
from megatensors.mega_hub import HfApi, HfFileSystem, hf_hub_download api = HfApi()local_path = hf_hub_download("research/demo", "config.json")HfApi, HfFileSystem, HfFileMetadata, HfUri, cache, OAuth, URL, and
TensorBoard names are aliases of the corresponding MEGA implementation. For
example, hf_hub_download and hf_hub_url resolve against MEGA's configured
artifact service. New integrations can use the canonical Mega* spellings.
This compatibility layer covers Hub repository workflows; it does not claim a managed Inference Endpoint lifecycle product.
Job methods#
Use the live hardware catalogue before dispatching a Job. The catalogue is the source of truth for CPU, RAM, accelerator, and price; an SDK never invents a GPU flavor that is not currently available.
from megatensors.hub import MegaHubClient client = MegaHubClient()hardware = client.list_jobs_hardware()job = client.run_job( image="python:3.12-slim", command=["python", "-c", "print('hello')"], labels={"lane": "release"},)for metric in client.fetch_job_metrics(job.id): print(metric["cpu_usage_pct"])final = client.wait_for_job(job.id, timeout=900)Jobs support dispatch, logs, cancellation, labels, schedules, mounted
repositories or Buckets, and opt-in private SSH according to the current
service contract. fetch_job_metrics streams CPU, memory, and network samples
for a running Job. Create a long-running Job with ssh=True; while it is
RUNNING, job.status.ssh_url is populated and the CLI can connect with
mega jobs ssh JOB_ID. The connection requires a registered account SSH key,
write ownership of the Job, and the authenticated SSH URL returned for that
Job. MEGA does not provide Hugging Face's free-form Job name field.
See Jobs, Job Configuration, and Storage Buckets for current limits and volume semantics.
For a UV script, MegaApi.run_uv_job and create_scheduled_uv_job build an
ordinary Job or scheduled Job. They use the same live CPU flavor, Bucket
mounting, secret, and SSH rules as MegaHubClient; they do not add GPU or
public port support.
Space methods#
The compatibility client exposes Space metadata, runtime controls, variables, and secrets:
from megatensors import MegaApi api = MegaApi()space = api.space_info("research/demo-space")runtime = api.get_space_runtime("research/demo-space")Call list_spaces_hardware() before requesting hardware. It returns the live
MEGA catalogue; compatibility enum values are not availability guarantees.
See Spaces and Space Hardware.
Routed inference SDK#
Inference stays in the SDK, separate from Hub repository operations and from the CLI/MCP Hub surface:
from megatensors import InferenceClient client = InferenceClient(provider="auto")response = client.chat.completions.create( model="mega/gpt-5.4-mini", messages=[{"role": "user", "content": "Hello"}],)InferenceClient uses MEGA's OpenAI-compatible inference router. See the
Inference Provider guides
for routing and provider-specific behavior.