Seeing that most of the base t2i models were apparently restored unto the Inference API without LoRA support, and briefly running into the sorry sight of quickly amassing unacknowledged ârequestsâ for specific Flux LoRAs (new & widely used ones alike) to be (re-/)APIâd, Iâm left no longer just jarred, but additionally genuinely perplexed by the implication: that to the HF team, model adapters (and, by extension, other forms of fine-tunes) are but an afterthought, rather than a core feature and perhaps the most obvious and all-around âdemocraticâ element of the current open source ecosystem. I mean, just looking at the sheer numbers, and across just about all modalities, most HF repositories contain fine-tunes rather than base models, and most text to image and (lately) text to video repositories are adapters (mostly LoRAs) rather than full checkpoints. If thereâs any chance that LoRAs are to be permanently deprecated from Inference API, due to some ill-thought-out (maybe vaguely news-inspired) policy a la perpetually-warm-models-only, then I think HF may well and truly be sailing into troubled watersâŚ
1 Like
Even if some t2i models are restored on the Inference API, right now there are still 0 image-to-image models available. I guess they are just less popular, hence not being a priority, but it would have been nice to at least get some sort of notice of this upcoming change, since this made HF unusable for my specific use case.
1 Like
I followed your advice and tried with Fireworks, and it was honestly such a breeze. Took me 10 minutes tops to migrate and it works great. Thanks!
2 Likes
hey guys, having the same error 500 request for any of the huggingface models DLed through ollama (after the recent update) - thanks guys really appreciate it.
1 Like
Youâre getting a 500 error (server error) after downloading?
Maybe Ollama is trying to communicate with something.