Spaces:
Running on CPU Upgrade
ML Intern usage is now billed to user account
Hello folks, ML Intern is coming out of beta and part of that means the usage will now be billed to your HF account. There are three main axes where billing occurs:
- Every call made by the agent is routed to an inference provider
- Every sandbox created by the agent (which requires user approval unless YOLO mode is activated)
- Every HF Job launched (which requires user approval unless YOLO mode is activated)
Your usage can be tracked under the Usage icon in the UI and your monthly credits for HF subscriptions can be used for making calls to the various models we support (Opus, Kimi K2.6 etc):
If you wish to use local models and compute, you can run ML Intern locally via the CLI: https://github.com/huggingface/ml-intern#local-models
Hello HF team,
Thank you for the communication. Will HF allow alternative inference providers for ml-intern? ie FAL is currently very pricy. A market place of providers, with much better harness can be useful. For example, in the attached image, these four bash calls are same commands to fire up sandboxes. While one is spinning out, there's no need for others to fire up. If these redundant requests will be billed to users, I suspect this will make people upset.
Besides, additional cost breakdown will be immensely useful. For example, if "Every HF Job launched (which requires user approval unless YOLO mode is activated)" will incur a cost, will that be fixed? or will that differ between on L4 to A100?
Hello @vedatonuryilmaz thank you for the feedback!
Right now, the web app supports models which we've found to work well with the scaffold: Opus 4.8 is generally the best fit, but you can switch to e.g. Kimi K2.6 or Minimax which are ~4x cheaper.
Regarding the Jobs usage, we show this in aggregate per session and the cost before approval is estimated based on hardware pricing:
Thank you so much @lewtun ,
How about cases where ml-intern just throws hf jobs that fail fast? I attach a breakdown of my hf-jobs use tru ml-intern. After ml-intern was out, I immediately became a pro user, then spent ~$100 usd in a single month on jobs, which many of them got botched. It is just unsettling to see HF trying to profit $$ this way. While we pay for HF jobs already, I really don't get the logic why ml-intern would cost to post those jobs, which majority would fail.
Please see my breakdown. Out of the 269 total jobs, ml-intern only had ~30% success rate. May be not worth paying until there's more clarification and transparency.
| Status | Total (all) |
|---|---|
| Completed | 80 |
| Error | 113 |
| Canceled | 72 |
| Insufficient credits | 4 |
| Total | 269 |
Hopefully @clem can reverse this decision.