Hub Python SDK#
MegaHubClient is the typed Python client for MEGA Hub service operations. It is separate from the local Tensor Runtime API, which opens and loads .mega artifacts.
Create a client#
from megatensors.hub import MegaHubClient client = MegaHubClient()print(client.whoami())By default, the client uses MEGA_ENDPOINT, MEGA_TOKEN, or the token selected by mega auth login.
Override configuration explicitly when needed:
import os from megatensors.hub import MegaHubClient client = MegaHubClient( endpoint="https://mega.tensorplay.cn", token=os.environ["MEGA_TOKEN"],)Pass token=False only for public, unauthenticated reads.
Hugging Face Hub source compatibility#
The complete Hub-style surface is available from megatensors._hub; the usual
imports are also re-exported by megatensors.mega_hub. The conventional names
remain available for migration, but execute the corresponding MEGA
implementation and use MEGA_ENDPOINT:
from megatensors.mega_hub import HfApi, hf_hub_download api = HfApi()local_path = hf_hub_download("research/demo", "config.json")HfApi, HfFileSystem, HfFileMetadata, HfUri, cache, OAuth, URL, and
TensorBoard compatibility names are aliases of their Mega* equivalents. For
example, hf_hub_url(...) returns a mega.tensorplay.cn artifact URL; it never
redirects repository operations to Hugging Face. New MEGA code should use the
Mega* spellings. Managed Inference Endpoint lifecycle APIs are outside this
compatibility scope.
Repository methods#
| Workflow | Methods |
|---|---|
| Metadata | create_repo, list_repos, repo_info, update_repo, move_repo, duplicate_repo, delete_repo |
| Files | list_files, upload_file, upload_folder, delete_file, copy_file, copy_files |
| Downloads | download_file, download_files, snapshot_download, iter_snapshot_files |
| Revisions | list_refs, create_branch, delete_branch, create_tag, delete_tag, list_commits, get_commit, create_commit |
| Community | list_discussions, get_discussion, create_discussion, reply_to_discussion, update_discussion, merge_pull_request, message and reaction methods |
Create and publish a repository:
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", revision="main", commit_message="Add model configuration",)repo_type accepts model, dataset, space, or mcp. MCP repositories use
the same metadata, files, revisions, Community methods, and native Xet
transport; marketplace execution and billing remain separate Hub services.
Job methods#
job = client.run_job( image="python:3.12-slim", command=["python", "-c", "print('hello')"], timeout="10m", labels={"lane": "release"},) for line in client.fetch_job_logs(job.id, follow=True, tail=100): print(line) final = client.wait_for_job(job.id, timeout=900)print(final.status.stage)The Job method surface is:
| Workflow | Methods |
|---|---|
| Dispatch | run_job / create_job |
| Observe | list_jobs, inspect_job / get_job, fetch_job_logs, wait_for_job, list_jobs_hardware, get_jobs_usage, get_compute_billing |
| Control | cancel_job |
| Schedules | create_scheduled_job, list_scheduled_jobs, inspect_scheduled_job, suspend_scheduled_job, resume_scheduled_job, trigger_scheduled_job, delete_scheduled_job |
See Jobs for limits, state transitions, namespaces, and Web API payloads.
Inference client#
InferenceClient targets the independent OpenAI-compatible Router rather than the Hub repository API:
from megatensors import InferenceClient client = InferenceClient(provider="auto")result = client.chat.completions.create( model="mega/gpt-5.4-mini", messages=[{"role": "user", "content": "Hello"}],)Provider values fastest, cheapest, and preferred map to the corresponding model-selection suffix. A concrete Provider slug pins routed requests to that Provider when the client uses a MEGA token. Passing that Provider's own API key calls its public endpoint directly instead. feature_extraction uses the Router's OpenAI Embeddings endpoint.
Set MEGA_INFERENCE_ROUTER_ENDPOINT only for a staging or self-hosted Router. Production defaults to https://inference.tensorplay.cn. See Your First Inference Provider Call for Python, OpenAI SDK, CLI, and curl examples.
Space methods#
The compatibility client exposes repository metadata and runtime actions:
from megatensors._hub import MegaApi api = MegaApi()space = api.space_info("mega/openapi")runtime = api.get_space_runtime("mega/openapi")api.restart_space("mega/openapi")Use request_space_hardware, set_space_sleep_time, pause_space, restart_space, variable and secret methods only with one of MEGA's configured fixed VPS flavors. See Spaces and the live OpenAPI Explorer.
Webhook methods#
created = client.create_webhook( name="release-verifier", url="https://ci.example/hooks/mega", events=["repo.updated"],) delivery = client.test_webhook(created.webhook.webhook_id)print(delivery.delivery_id, delivery.state)The lifecycle methods are list_webhooks, get_webhook, create_webhook, update_webhook, test_webhook, list_webhook_deliveries, and delete_webhook.
Account public keys#
from pathlib import Path public_key = Path("~/.ssh/id_ed25519.pub").expanduser().read_text()key = client.add_account_key( key_type="ssh", name="Work laptop", public_key=public_key,)Only ssh and gpg public keys are accepted. The client rejects private-key material locally.
Errors#
Service failures raise MegaHubError:
from megatensors.hub import MegaHubError try: client.repo_info("private/missing")except MegaHubError as error: print(error.status_code, error.method, error.url) print(str(error))Validation performed before a request usually raises ValueError. File methods may also raise normal filesystem exceptions such as FileNotFoundError.
Client selection#
| Need | Use |
|---|---|
| Shell and CI commands | MEGA CLI |
| Typed Python Hub automation | MegaHubClient on this page |
| Routed model inference | InferenceClient and the Inference Provider guides |
| Framework tensor loading | Tensor Runtime API |
| Another language or custom transport | OpenAPI Explorer |