Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
victorΒ 
posted an update Aug 21
Post
5341
πŸ™‹ Calling all Hugging Face users! We want to hear from YOU!

What feature or improvement would make the biggest impact on Hugging Face?

Whether it's the Hub, better documentation, new integrations, or something completely different – we're all ears!

Your feedback shapes the future of Hugging Face. Drop your ideas in the comments below! πŸ‘‡

A brand new Ios/android app for the Hchat, the actual is..... Ahm ..... I have no words man , but My younger students need a little bit more apealing

There are a lot of different high quality recipes, notebooks, educational resources… and it’s sometimes difficult to find the right pointer. I’d love to help on that btw😎

Β·

yes there are many resouorces in the documents but very hard to navigate ! << especially between transformers and gradio !
I also fid the gradio documentation very bad ! ( as they are often changing thier syntax )

  • Bookmarking blogposts into our collections (would love to reread some of these blogposts or store them somewhere)
  • Give a person editing access to one of my repos without creating an organization
  • Add search by custom tags [optional + low priority]
  • Huggingface workflows + maybe create a blogpost about it, it would be interesting what the community will create
  • Implement @cfahlgren1 's https://huggingface.co/spaces/cfahlgren1/my-heatmap in all user accounts
  • Maybe allow people to add a README to their own user accounts' homepage
  • dms
Β·
  • Bookmarking blogposts into our collections (would love to reread some of these blogposts or store them somewhere)

I like this one, I would also like to add bookmarking collections as well.

Improved LFS stability. It can't be that files of models randomly become permanently inaccessible due to LFS errors. Try downloading the following and you will get 403 for no reason: https://huggingface.co/jukofyork/miquplus-aurelian-70b/blob/main/model-00003-of-00015.safetensors making the entire model unusable. Uploading models got so unstable that it now requires around 3 tries due to uploads randomly failing with LFS errors like Your proposed upload is smaller than the minimum allowed size. Please remove the 24-hour limit to complete an upload to LFS.

Β·

Interesting, have you seen this? https://huggingface.co/blog/xethub-joins-hf (will be a game changer πŸ”₯)

Screenshot 2024-08-22 003056.png

Please contact me on Instagram or Telegram @KhoroshN, and I'll tell you where the bug is.

Β·

the bug is in Line 404

from Lain


^ These are great
I have a few ideas to throw in the pot ~

Serverless Inference

  • Allow image gen models running on serverless inference to specify width and height parameters.
  • I'd love more transparency as to what it takes for a model to be switched from Cold to Warm
  • Expedite development of the Inference Playground!

QOL

  • Search by Organization but more than 3 results at a time.
  • Add a "New Folder" or "Upload Folder" button in the "Files" page of a space. It's a bit cumbersome to work with folders in the GUI, and I don't like using CLI
  • There should be a way to easily isolate or find your active ZeroGPU spaces. (I had a situation once where an old Zero space had been sleeping for ~ 2 months and got buried in my other spaces, taking up one of my 10 Zero spaces. I had to find a needle in my invoices but since it was sleeping for so long it was tough to identify)
Β·

Allow image gen models running on serverless inference to specify width and height parameters.

Actually, it is possible in its current state, but you will have to read the documentation and try-and-error to find out.
It is not in the UI on the web, and Space generation is not easy with Gradio, though.

Also, I think it would be a good idea to review all the UI-related things on the web of the serverless API, not just image generation.
I think those are stuck in the model situation as it was when they were created, even before SD1.5?

client = InferenceClient("Yntec/epiCPhotoGasm", timeout=600)
client.text_to_image(prompt="1girl", negative_prompt="bad anatomy", height=768, width=512, num_inference_steps=28)

If you want to see the above code in action, try using this space: there are width and height settings in the "Advanced" accordion.
https://huggingface.co/spaces/John6666/Diffusion80XX4sg

Usability Related:

  • (I'm used to that style of community for personal reasons.) The general lack of announcements on HF leaves users confused whenever a server-side error occurs, with no way of knowing when it will be resolved. Please make announcements when possible.
    Space is Building… permanently
    Inference API turned off ? Why?
  • Is it possible to have personnel at HF to respond to error reports in the forum as well as in Discord?
  • There are quite a few substantial hidden features that are not listed even if you read the documentation from cover to cover. If you read the source code, you will find them.
    Is it right that there is no other option but to read it from cover to cover in the first place...
    I don't know how to provide it correctly, but something is wrong anyway.
  • (I'm a geek, fine) HF is too hard to use for anyone but geeks, academics and programmers. Can't someone hire an advisor to run the site for the majority?
  • The searchability of HF is not bad (except for the glitches), but the listing is phenomenally bad; I don't want it to be all pictures like Civitai, but can't we do something more? I like the current UI the way it is, so maybe we can have some kind of separate UI that can be switched?

Paid Services Related:

  • I understand that ZeroGPU service is a special inexpensive service, but that being said, I would like to see a more extensive flat rate service.
    The pay-as-you-go system is heartbreaking for individuals.
    Something like $20 gets you 15 frames instead of 10, or Quota from 30 seconds to 45 seconds, etc.

HF Spaces Related:

  • I would like to do something about the problem of server connection errored out in the middle of executing a time-consuming process in python.
    This may be a problem with Gradio or httpx, but it may actually be a problem with HF's server configuration, so I'm writing it here just in case.
    The space that some organizations have released seems to work around the problem by launching the CLI tool from within python, but that should be a more beautiful solution if it can be completed in python.
    https://huggingface.co/spaces/ggml-org/gguf-my-repo
    https://huggingface.co/spaces/arcee-ai/mergekit-gui

QOL:

  • When I am watching an organization's repos, notifications are sent out by default when there is PR or community activity for any of the repos, and I seem to have to mute notifications for each repo individually if I want to stop them.
    Maybe that would be more convenient if it were an organization I was a member of...
    I would like to have an option to turn off all notifications at once by default, just like when I am following an individual.
  • I'm not very familiar with social networking sites, so I don't know if this trend is normal or not, but I think HF residents tend to be less doing β€œlikeable” or any other feedback.
    I don't know if many of them are not interested in anything other than the subject they are working on, or what, but if they are just shy, could this be improved by having a permanent form like a suggestion box where people can throw in their complaints that can be used semi-anonymously?
    I know the people in charge of scrutinizing the suggestion box would have a mental breakdown...
Β·

yes i would like to see some type of improvement to the gradio spaces and deployments and usages across trhe website :
once you switch to a task on the site using gradio its often 50/50 and some long timeperiod before it loads ! (Very annoying ! ) Perhaps also in spaces we should have a color tto distinguish sleeping spaces from awake spaces and the none working spaces should not feature in the search ? or at least feature at the end : so we should have working spaces , sleeping spaces , running spaces : Paid spaces , free spaces , private spaces , public spaces : all features on the search for a space : same with the models we should have the ablity of these fil;ters!

The Spaces and the models used in the Chat with the assistants are very cool, but you need your own LLM, your own AI, and it needs to be unique - an all-in-one solution for text, video, voice, and image.

Β·

yes im not sure gradio is really doing it well ! for a UI provider for a model !
in fact thier multimoel textboxes are badly documented !

Educational resource in Hindi Video Lectures. I really want it

Β·

interesting ! >>> perhaps AI generated Dialogue would be the anser for this :
by creating the resource in any languguage and using the whisper model to translate the tutorial give you the option to choose your narrator for a tutorial !
hence in fact any tutorial can be any language as long as you have a clean Transcript !

I would like to see an expansion on the HF Agents! - And HF Tools ! << Im not sure as i have not used the inference endpoints : but i would like to be able to use the api to hit the spaces without cost !!

I would like to see and expansion on the model endpoints on your profile : ( not for extensive use But you should be able to test every model (at least once !) Especially your own !

I would like to see a better handling of the datasets : ( perhaps a way to edit them on the site? )

I would like to a lot more documentation on Creating Custom Configs and Instantiating models ! ie generating a model from scratch Especially the mulltimodals : Which Pretrained Components can be Slotted in as Some Archetectures are actually ONLY and Archtecture and Your pretrained models can juist plug in ! ( such as a VisionEncoderDecoder model : ( can be created from an existing pretrained ( llm ) and (vit ) model : ) < << /hece we would see more But these models are not correctly documented ? How for form them : Even the lllava ad Bakllava and these models are not pure models and can be created from other models : hence we should be seeing a lot of these model but we are not :

I would also like to see muych more Abitlity to manage your models : ( i cannot select them all and delete ? or Private ? ) Same with the datasets :

Also when ther transformer Library is updated the site should also Announce the Changes ?

to run a 7b model ::: The free endpoint should be able to manage this simple feat !

I would pay for a service on hugging face if they had a COLAB !<<< this is missing ! what is the point of spaces ? as its like a mini webhosting service just for your models ? so why not a llama factory ( a real training center ) ?as this is the most of the reason to use the colab in general ( as most programming is done local and only intensive project are required in the cloud services : SO in truth a GOOD ! trainer which is Gradio Based ? In a Spaces ? ( i was going to make this for my profile but ? the price for a training session is not comparitive with google colab: In truth Hface is a Alternative to runpod and all these services ? and people use these to spin up a machine to hit via endpoint ! <(noy neccasary to have running all day - but for a purpose : I would be good if we could send a signal via the endpoint to wake it up ? , as well as caching the space locally to speed up the use of the space ? so when using the spaces via endpoint i could be an option to run on endpoint or clone locally ( if it fits a certain criteria ) ( ie does not exceed specific cache size limits )....

Also :
Perhaps a messaging system is a basic thing ! <<< in such a large website with locked spaces and timezones etc : i think messaging is very important : as how do you connect to people ? leave a message on thier model ? so i think this aspect is missing ! ? MostForums Have this and hface have Posts ! ( is this simular to a forum ? ( it seems that posts element is not correctly organised ! ) i would love to share my adeventure and discuss with people thier projects and potential colaberations : hece the abil;tiy to colaberate is what CoLab is all about and many people use it singlular ! <
the hugging face icon - is a very important imogi i denotes together ness and colaberations and lots of love ! --->>> Where is the love ? we need the co-lab mentality !

there are varioius taks associated with a model , which is genrally covered here : But agai the documentatioin espite being Quite good is very (Lots of elements missing ) Important aspect of the models ! ( often they are missing from the documentation : from a model is added it shouod have a way to run the model and a way to create the model or eit the model : fully explained ! many aspects of the usage and runnnign of a model is universal to the hugging face library hence these often do not need expansion but should be in a prime position in the docs !
and the other elements heavily featured in the specific model DOCS:

I would hope Some or all of these features would be solved :

Β·

I would pay for a service on hugging face if they had a COLAB !<<< this is missing !

If you ask me, why wouldn't it be? How likely it is to be.
Google locked Stable Diffusion out of the free version of Colab because they couldn't handle the load, but if we could do the same with HF (even if for a small fee), we would make a lot of progress in training our models.
Best of all, more HF specialties.πŸ€—
https://x.com/thechrisperry/status/1649189902079381505

meta-llama/Meta-Llama-3.1-405B-Instruct-FP8 is not available right now.... it was an improvement... please bring it back if it is withdrawn not due to temporarily technical issues

one of my favourite features of Hugging Face is the daily papers page. It is great, but I think it could be improved in a number of ways:

  • it would be nice for papers to be tagged with topic. [LLM] or [Multimodal] and so on. Then the user could filter by topics relevant to them when looking through daily papers.

  • would be nice to have the ability to search over "influential" papers. Those papers with high upvotes, which are key reading.

  • I find that the current search is a bit temperamental - better ability to search and find relevant papers would be beneficial in my opinion.

Β·

YEs daily papers and annoucements need improvememnt

I would really like to see the expansion of the NVIDIA NIM Serverless Inference collaboration. Currently, it does support major models like LLaMA 3.1 and Mixtral, however, the real power of Hugging Face, in my honest opinion, is the data and customization of models with data. If this collaboration somehow included the deployment of fine-tunes and adapters of these major models, I would switch every single inference workload to here in a heartbeat.

  1. More Ways to Interact with other peoples.
  2. Suggest/Recommends of Models/Datasets/Spaces. (Just like Suggested for you posts )
Β·

i think private message system and a public/private chat system could be good

  • I frequently see random people duplicating top-trending spaces and promoting illegal ads and unethical activities with links to sites. It would be beneficial to restrict these actions by issuing warnings when they attempt to commit or upload files. [ PS: I still come across it. ]

  • Additionally, implementing chat support within Hugging Face would be valuable. This feature could provide knowledge and guidance for those who are just starting to build today, helping them navigate the platform and use its tools effectively.

  • [ Activity Overview ] for Users.

Β·

illegal ads and unethical activities with links to sites.

I think a concrete example would help HF deal with this. (I hope you don't mean me...🀒)

It's kind of a hobby of mine to go through various repos, but the percentage of repos that post URLs to unfamiliar sites is very low.
So it should be easy to create a system that automatically notifies HF staff when there are strange URLs in the README.md or .py. (Slightly more difficult if they are dynamically generated)

Forum trolls, individuals and vendors for profit (like X's impression zombies...) will spring up endlessly and will never give up, so we need to take a stance to either manned countermeasures in the end or ignore them from the beginning. It would be possible to filter them efficiently in the preliminary stages, though.

Clear understating of required CPU / GPU for spaces that a non-program would understand. Min requirements on copy / clone spaces.

Β·

Do you mean when cloning to a PC in the local environment?
If it's free CPU space on HF, it's a virtual environment, so (theoretically) it's all the same environment.
I'd like that for paid space. Even the creator doesn't know exactly what specs are needed, so the server will need to have the ability to automatically benchmark.

@victor Thank you for your desire to hear feddback! :) I love HF and I would love to provde you with my thoughts to make it better :)

I have couple of things in mind that I would love to see implemented:

  1. I would really love to see a paid tier for ZeroGPU where I could pay to have unlimited on-demand access to ZeroGPUs. I think a lot of devs may be interested in such arrangement and it would make ZeroGPU profitable and useful for larger/commercial projects/deployments.

  2. Auto-recovery/auto-reboot for spaces. In many instances my spaces become inactive (with run-time errors) for whatever reasons, so I would love to see an option to allow HF to auto-reboot/auto-repair my spaces if needed.

  3. Auto-upgrade for spaces. I have a lot of spaces and its very difficult to constantly having to update to latest version of gradio. My code/implementations would tolerate auto-upgrade fine so I would love to see an automated option to update gradio and other desired libraries in settings somewhere.

  4. More flexible models/datasets download counter. My models are not integrated with HF so I do not see number of downloads even though people use my models/datasets. It would be very nice if I could specify which files from the repo I want to track for download counts.

  5. Aggregated use/downloads statistics for the HF account. I would like to know what is being used and what is being downloaded in details, so it would be very nice and handy if you would create a dashboard for aggregated stats from all repos of the user account with options to break it down as needed.

  6. Custom order of spaces on the user account page (similar to how GitHub allows it). I want to be able to present my pinned spaces in a specific order.

  7. An account README.md which would show up on top of the main account page, similar to GitHub user account README.md. I want to proved visitors with nice intro to my account and projects and currently there is no way to do so in a nice and easily visible way.

I think this is all for now, but I will update this post with more ideas/suggestions when I will recall what else may be useful and nice to see :)

Thank you for your time to hear my feedback.

Sincerely,

Alex.

Β·

Auto-recovery/auto-reboot for spaces.

I think this needs to be realized as a priority. For example, the space below crashes often because people use it too much, rather than because of a flaw in the code. This time it seems we ran out of disk space.
The individual author could set up a cron with code to knock it up using the HF API, but that would be roundabout and not essential, and would probably end up overloading the server.

In consideration of the server load, why not just prohibit frequent reboots in the auto-reboot function?
If the space does not get the Running state after 3 auto-reboots, then auto-reboot should be disabled for 24 hours or something like that. Without affecting manual operation.
https://huggingface.co/spaces/multimodalart/flux-lora-the-explorer

HF could have a real About page where it explained what it actually does, rather than just a pithy mission statement and no actual information.

Noobs arriving should be welcomed, but this org/co/collective is opaque to outsiders. I literally had to go to Wikipedia to figure what this webpage's primary function was. That's wild in 2024.

Β·

https://en.wikipedia.org/wiki/Hugging_Face
I wandered in when I saw a fellow forum member playing with HF's animagine repo, and I haphazardly learned to use various HF features over time, but this is more like the way ancient underground sites were, isn't it?πŸ˜…
Nowadays, the top page of a website has use cases, tutorials, and visual samples that everyone (including non-programmers) can understand.

I would love better filtering

First I think sort by created is broken, but haven't checked on desktop recently

Second, I would love date filtering, like show me trending models that were only posted or updated in the past 7 days and such

Β·

I was reminded of this when you pointed it out.
Sometimes there are repos that are not caught in the search regardless of the option in the first place. It's not that they have been erased or hidden because they are inappropriate.
In such cases, when I remember whose model it is, I go to the author's page from another appropriate model of the author and follow it from there, but this too may be a malfunction in the search database.

There is a lot of useful data even if it's old (it's a little over 2 years old, even if it's the oldest), and it would be nice to be able to do a time-specific search.

we need a space for flux lora , where we can input lora id and run it

Β·

I've modified multimodalart's space so that you can specify LoRA directly.
It seems to work for now.

Original
https://huggingface.co/spaces/multimodalart/flux-lora-the-explorer
My Mod
https://huggingface.co/spaces/John6666/flux-lora-the-explorer

P.S.

HF's official Flux LoRA service now supports use of LoRA within HF.
https://huggingface.co/spaces/multimodalart/flux-lora-the-explorer

@victor When using the Huggingface Inference API in the format of OpenAI's API, it differs in terms of required parameters and response structures.

image.png

I test this on this
https://app.nextchat.dev
https://chat-preview.lobehub.com

I'd like to see a proper branch management in the web interface. Maybe it's cool for people who are heavily involved in software development and comfortable with git but it's horrible for others. I copied a few python commands to manage branches and deleting files to use these whenever I need, so I kinda can get by. However I see not one or two people who just throw models in subfolders because they couldn't figure how to do it, and it breaks automatic downloading for such models.

A more user-friendly downloading tool would be nice as well. Huggingface-cli requires quite lengthy commands to just download models somewhere that isn't huggingface cache dir. You have specify a folder, not making symlinks and it's a console command, quite a hassle.

I'd like to be able to filter models by name on user's pages. Like you can throw something like "gemma exl2 9b" on the normal model search page and it finds matching results but on user's pages there's more strict search that would fail with such a query.

and obvioulsy there should be a BOT !@ for the docs highly fit on just the huggingface librayr and associations !
which people can ask information from aboiut the docs etc : it should also point you to a source or even give example of imlementation !

Β·

i means thats basic ! we are an AI site lol ! with no online bot !

hello john,

  1. as a general rule, i support john requests on suggested hf development on the forum and threads

  2. Let the API inference compute up for all models like before

  3. give possibility to use Flux model like SDXL models on API inference compute webpage or space webpage.

Β·

Let the API inference compute up for all models like before

What kind of models where you running?

i woud like a lot also see more ai video interface and sound ai interface developped like we see some website where you can do remix of song, separate audio from lyrics etc...

delete

Compatibility:

Sorry for the extremely long post.

As anyone familiar with the Stable Diffusion community can somewhat understand, image generation in HF once rose to prominence and then fell into disuse, in part because the A1111 WebUI became the de facto standard.
The single safetensors file became the norm, making it somewhat incompatible with Diffusers.
The rise of Civitai was probably not the only reason.
(I wasn't yet playing with AI at that time, so I can only speculate...)

As far as common architecture models are concerned, it is possible to convert them to each other by giving up metadata and some information loss, and conversion scripts are also available. In fact, there are probably people who do so.
The internal physical differences are at best just the names of the keys, not the substance...
But it's a pain in the ass and enough of a problem to abandon the site.

As you know, FLUX.1 is currently very popular and is always relatively high on the trend list.
So what is the format used for the distribution and exchange of this model, and it is not even a stand-alone five-body safetensors. Even within HF, where there is plenty of storage, but even more so on Civitai and other sites where capacity is tight.

Specifically, it could be a torch.float8_e4m3fn, or it could be the distribution of only Unets (Transfomers) quantized by NF4 in bitsandbytes, and so on.
But anyway, there is rarely anyone distributing in torch.bfloat16 or torch.float32, which are supported by HF's Serverless Inference API.
Not everyone has the bandwidth to upload 30GB at a time.

We can use both NF4 and torch.float8_e4m3fn from our Zero GPU space, but that's all.
We even have to write cumbersome code to use NF4. (I'll give you an example below, but it's tedious, right?)
Even though it was Mr. lllyasviel and sayakpaul at HF who started NF4 quantization.
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981

What I have described at length is a preamble for those who do not know the circumstances of image generation.

If this continues, it will mean that HF will be even further removed from the de facto standard in the image generation community, which is sad.
As long as the source is open, it is open, but that is not the case.
Seamless integration into the ecosystem is what is most important.
Please be the HF that explores what is possible to achieve this. The key is compatibility.

Unless you are Apple or a 16-bit era computer company...

P.S.

Today, the situation has become even more chaotic with the announcement that the next major version of the Pony model, the current de facto standard within SDXL, will support AuraFlow over FLUX.1.
In any case, we can expect more diversification of exchange methods.

Towards Pony Diffusion V7, going with the flow.
https://civitai.com/articles/6309/towards-pony-diffusion-v7-going-with-the-flow

How to use NF4 quantized FLUX.1 from Diffusers in Zero GPU space:

https://huggingface.co/spaces/nyanko7/flux1-dev-nf4/blob/main/app.py
https://huggingface.co/spaces/nyanko7/flux1-dev-nf4

Β·

AuraFlow was chosen over FLUX.1-dev because of the latter's license that doesn't allow commercialization of any kind. In my opinion FLUX.1-schnell that allows it would have been a better choice because it's significantly superior to AuraFlow, and, I don't know what's Kolors's license, but beside prompt adherence and text coherency, I like Kolors's outputs better.

I want to start by expressing my appreciation for the incredible work Hugging Face has done for the open-source community. Your contributions have been invaluable, and I’m grateful for the tools and resources you've provided.

Please take the following as constructive feedback. I wouldn’t have mentioned these points if you hadn’t asked, and I hope they can be seen as suggestions for further improvement.

  • Software quality: When I first started using transformers, I was thoroughly impressed. The basic "hello world" examples work wonderfully, making the initial experience smooth and enjoyable. However, nowadays I am am regularly diving deeper into the library, and I am regularly facing challenges such as long-time standing bugs, undocumented issues, lack of API documentation, and occasionally broken functionality. I am only guessing here, but I think the majority of these repos is written by research engineers or researchers, whose focus might be more on the methodological correctness (which is of course crucial as well). That said, it might be helpful to include someone who is stronger in software development and less knowledgeable in ML. This would be the first person to complain about "clean code" issues, and also would be the first to notice problems with the software.

  • Posts: Great feature! However, it could be enhanced by adding basic text formatting options. This would make posts more visually appealing and easier to read.

  • Papers: Restricting this to arXiv is too limiting. While I understand the rationale in terms of implementation effort, if the goal is to be the "Github of ML/AI," it might be worth considering support for at least the high-ranking conferences (or a subset thereof). In many cases, the conference version of a paper supersedes the arXiv version, and this restriction may inadvertently encourage the use of preprints over the finalized versions.

Again, these are just my personal pain points, and I’m sharing them with the intention of helping Hugging Face continue to improve.

Miscellaneous:

In any forum or community, the composition of those who are active changes from time to time, and those who check the Post or Forum are part of it, and those who speak up are even more so. The majority are always ROM.

So I think it would be better to conduct a survey like this not once, but several times, preferably through the Forum and other channels (if they exist), to get a better opinion.

Oh another big pain point: notifications

I would love to be able to subscribe to be notified of new models posted by people or organizations, but it's near impossible as is

Β·

I would love to be able to subscribe to be notified of new models posted by people or organizations, but it's near impossible as is

I agree.
You (not bartowski) may not know this if the number of people you follow (not your number of followers) is small, but the top page updates about the people you follow are actually mostly random pickings. Not all of them!
https://huggingface.co/

Maybe we can solve this problem by building our own crawler like people who run news sites do, but I'm not sure I want to go that far...

P.S.

There has been progress on this matter in another post.
https://huggingface.co/posts/bartowski/524900219749834

Notifications when users / orgs release new models / datasets. 🀞

SNS-like feature:

  • I would like the ability to block posts from certain users or organizations to my repo.
    So far I have not been victimized myself, but I occasionally see people being attacked for no good reason, and that seems painful.
  • As others have pointed out, there are many imperfections for a social network, given the situation in 2024.
  • Also, this may be a personality issue of the people who congregate on HF, but I think that in general, people tend to be less in the habit of liking repos. However, perhaps one of the reasons may be the lack of HF's searchability, listing, and social networking functions. This may hinder smooth feedback to model and dataset authors.

πŸ₯ΉπŸ’•πŸ’žπŸ’—I need Original version of "Serverless Inference" - Let ANY users can generate pictures on model page (login users have have higher numbers of generated images, not login users fewer amount ) I think it's better to attract new customers and retain old usersπŸ™‚πŸ₯°

(Recent versions of Serverless Inference limit new/small model grow...I think it will affect the willingness of some model developers..πŸ₯²πŸ˜­πŸ˜­

Β·

Definitively! The most popular models have many spaces to use them and they don't even need the Serverless Inference API because people could use them in the spaces, while the new models don't have anything and it's impossible to use them and make them popular!
This all is backwards!

Image Generation AI Related:

Β·

I need all models back for free, including those models that do not have enough activity to be deployed to Inference API (serverless) yet

QOL:

QOL:

  • I would like to see HF have the ability to inform everyone of these survey-type events, or the habit of informing everyone if it is already there.
    Even Yntec up there, who is a very old and active member of HF, was not aware of this survey until he happened to talk to me.
    Too many notices may sound like SPAM, but it is better to send what is needed.

would be great if we had a space for animatediff + controlnet space , with much functionality of https://www.youtube.com/watch?v=qczh3caLZ8o

HF Public Services:

I would like to see a powerful public space provided by the HF side. For example, the following spaces are available for unlimited use up to a total of 3 hours of all spaces per account per week.
(though we would have to deal with multiple accounts)

  • ComfyUI GPU Space
  • WebUI Forge GPU Space
  • LLM GPU Space
  • Various Converter Space
  • Various Merger Space
  • etc.

With current CPU or Zero GPUs, we can literally only do a moment of processing as a demo, and the range of what we (the poor) can do is limited.
And/Or I would like a metered plan with a maximum amount.

Is it possible to make wasm support all models in huggingface?

Wasm have below advantages

  1. run locally, no server or extran machine needed, which can help user to try new model without any cost.
  2. no any privacy concerns

Now, there are several model that support wasm.

  1. whisper.wasm, which can convert audio to text.
  2. secret-llama: text-to-text LLM, realize by web-gpu
  3. web-llm: text-to-text LLM
    If it is possible to make a project to support wasm for all huggingface models, this can definitely accelerate the development of LLM.

Forum link: https://discuss.huggingface.co/t/is-it-possible-to-make-wasm-support-all-models-in-huggingface/105185

Β·

We have Transformers.js, the JavaScript/WASM/WebGPU port of the python library, which supports ~100 different architectures.
Docs: https://huggingface.co/docs/transformers.js
Repo: http://github.com/xenova/transformers.js

Is that the kind of thing you're looking for? :)

I find it extremely frustrating on Spaces, that many of the apps donΒ΄t provide a short description already on the overview nor on the interface. itΒ΄s too often guesswork. I find this should be a requirement , there is soo much time lost by scrolling through these - no doubt - treasures of apps with their cryptic names. gives an extremely ex-clusive appearance. very discouraging, if you are a beginner, visitor or the like. I feel HF should definitely care for understandable inclusive environment. depends too much on coincidence and random findings to run into the next treasure.

also there should be categories offered with tags to easily navigate to what you search or give orientation.

thanks. lots of waste of time and missed opportunities in this great space that HF is

Hello @victor , thank you very much for this call, I have a few ideas, some of which overlap with other members of the community:

  • the possibility of sending DMs to other users or organizations;
  • the ability to disable update notifications in a space when you're the producer. This would be particularly useful for daily tasks, so as not to pollute the feeds of people who follow users who edit datasets every day...
  • image-text-to-text model support in the serverless API;
  • json mode in line with the OpenAI API for endpoint inference for text generation

I would like to suggest a compact view or something similar on the main page.
Right now, if someone you follow releases a lot of content in the same day, most of the news on the main page is taken by it. What I would like to see is some sort of compact view in which if someone has more, then say 3 posts, everything will be shown if you press a more button or an arrow that expands the view to all the releases.

Miscellaneous:

I have been at HF for a few months now, and I am writing about what I have noticed and what I was having trouble with or feeled weird.

  • Many spaces cannot be updated due to library compatibility issues, either because Gradio frequently truncates backward compatibility in version upgrades and crashes with syntax errors, or because the libraries required for the space are extremely old. (even 2020)
    • As for Gradio, simply ignoring version-derived syntax errors in further new versions would greatly improve the situation. I don't think it is necessary to be compatible with all past features, but right now it is a bit too much. https://github.com/gradio-app/gradio/issues/6339
    • This is not a job that HF should normally do, but since there are some libraries that have stopped being updated even though they are often used in the space, it would be more progressive if HF staff went around fixing the even out-of-date dependencies part alone or fork them to HF dedicated. Many of the libraries are particularly outdated, especially those related to sound and video. Text and images have transformers and diffusers, so these problems rarely occur.
  • There are countless entrances in the Diffusers file format, but only practically three python scripts for exits, and the HF-trained model can't get off the island.
  • In many cases, even if there is a script, the GUI space is not created and it goes unnoticed. (Not only text-to-image related) Or it does not work in the CPU space due to lack of performance. For SD1.5 it is fast enough, and SDXL is manageable. But what about the future?
  • In general, HF seems to have a high hurdle for people who can't write code, and few of the types of people who think it is a problem.
  • HF in general is easy to use once you know it, but if you don't know where it is, who is there, and what it can be used for, you really don't know what to do.
  • The company building is large and has an unexpected variety of things to explore, but there is no receptionist and no information desk, making it worthwhile to explore.
  • It is in my nature, but in any case, everyone does what they want to do by themselves at their own pace, so they are generally indifferent about contact with the outside world, and sharing of know-how is not progressing. As a result, HF relies on outside communities probably more than it should. Just a little bit of introduction elsewhere can spike the number of downloads. Usually, this means that there is no way to find them.
  • HF appears to be somewhat command-less. No, I think it would be depressing if there were...
  • I enjoy communicating in English and it is unrelated to the Tower of Babel, but shouldn't there be an optional field for each person to write about the languages they speak?
  • Can we have an alias in HF for mentions so that if you send a mentions, it will reach all staff and someone will respond? The only biggest problem is the trolling issue...
  • Well, I'm having fun.

QoL:

  • I would like to see an X(Twitter)-like translation function. Basically, Forum, Post, and Discussion are assumed, but it would be better if it could be applied to README.md, HF UI in general, Spaces, etc. Translation, especially of the UI, need not be very precise. The actual processing can be done by HF ourselves or by a link using the services of an external site. Since we can use markdown, it would be good to quote and fold the original text to avoid fatal misunderstandings.
    I don't know about people in other countries, but Japanese people have a terrible allergy to foreign language sites. They tend to run away when they see a foreign language only site.
    It would be nice to install a browser extension such as Google Translate, but in Japan, smartphones are more mainstream than PCs, which lack a wide range of extensions. Furthermore, there is almost no tendency to dare to read foreign-language sites, except for programmers, scholars, stockbrokers, pornographic videos, and piracy. There are probably other countries like that, and I think it is a lost opportunity.
    The prerequisite for this feature is the setting of the main native language. If public and private settings can be set, shy people will be at ease. If you don't know what language to translate to, there is nothing you can do. You can use the browser's code page, but it may not be accurate.
  • I heard that some HF spaces are experiencing a phenomenon where you have to log in to see them, even though they are not NSFW. I have never encountered this problem myself. I don't know if it is a bug or a specification.
  • I've written before, I would like to see a permanent space for soliciting opinions. It would be nice to have one both open and closed, and to explicitly allow submissions in the native language to make it easier to submit. One-way communication would be no problem with translation. The challenge is to prevent pranks. It would be easier to use if there were simple categories such as bug reports, complaints, consultations, etc.

Proprietary format:

  • HF is generally very easy to use as long as the HF-specific formats (README.md, config.json, and unquantized safetensors files) are available. If not, it is generally terrible.
    Transformers' handling of GGUF without config.json was buggy and nearly unusable when I was messing with it just now, and Diffusers seems to be too busy to deal with it in the first place.
    This would not be a problem if HF was intentionally using its own format for its enclosure strategy, but I suspect that they either don't have the capacity to deal with it or are simply unaware of the rapid changes in the surrounding environment.

Serverless Inference API:

  • Come to think of it, why can't we use some of the features that are normally supported by Diffusers save_pretrained, such as StableCascade and the officially adopted community pipeline? I used to think it was due to server load, but I don't see how it could be heavier than Flux.
  • If it is not too difficult in terms of specifications, it would be useful to be able to specify specific files in the repo in the YAML of the README.md, to specify LoRA strength other than 1.0, and to have more licenses that can be specified. It's about time that HF supports FairAI and Flux dev licenses as standard.
    In addition, it would be better if the editing function of README.md could be expanded.
  • It would be useful to be able to specify the scheduler and sampler for image generation, VAE, embedding, chat template for text generation, and various parameters for VLM. (I know it's easy to do locally, but trying to do it server-side is a lot of work, but it should be useful.)
  • I would like to do runtime overrides of the parameters specified in README.md. For example, if I can override base model parameters, it will make it easier to use LoRA.
  • If detailed metadata is not written when generating images, it is recommended that it be written, as it is done in the Animagine space, for example.

Bring back open source hackathons / challenges. You don't even have to handle things. I would love to host a hackathon about AI in the medical domain with the help of Huggingface.

Β·

Was there some unavoidable reason why it was there in the past and then disappeared?
If it is held again, I would like to participate, even though I can only do duct-tape-like coding.
Also, someone in the Post said he was bored, so I'm sure such people would be happy with the event.

I don't use comfyui so I would like to see comfyui models in their own section. I don't want to have to look dig through them for find usable huggingface. Frankly I dont think they belong on this site anyway, as huggingface seems to be more code-oriented. Let civitai have all the non-coders.

Even better, I would like to see models automatically tested to see if they will even run with huggingface code before they are posted.

Β·

https://huggingface.co/docs/diffusers/v0.30.2/api/loaders/single_file
I disagree, because with very few exceptions (broken files, single unets cut out, etc.), ComfyUI and WebUI files are available from Diffusers, and in fact many people have put them there for that purpose.

Even coders don't use only ComfyUI or only Diffusers. Anyone who is concerned about only such things is just a madman. It is similar to the mouse-keyboard controversy or the GUI/CUI controversy, maybe a little different from the Mac, Windows controversy.
Incidentally, coders can convert and use these single files, which are a small exception, and there are some who do so.
https://huggingface.co/spaces/nyanko7/flux1-dev-nf4
https://github.com/huggingface/diffusers/issues/9165#issue-2462431761
However, I would like to see a solution to the problem of the files being scattered around and difficult to find, as well as a server-side verification of whether or not they work with a single shot from_pretrained.
Now that the Inference API has become virtually irrelevant to individual users, I would like to see the benefits of using the HF file format for more than just researchers and corporate (Endpoint API) users.
If it works, it would be even better if the server could generate some safe sample images.

Hub:
Either;

  • New NLP task filter: function-calling (models that are compatible with function-calling feature)
  • New Category: Agents (with sub-tasks: function-calling, structured-outputs, etc), so that, even VLMs or other models could be filtered based on their agentic abilities.

I want to create a pipeline signature and provide a few examples, then have the hub go and test a bunch of models against that scenario and pick the best ones for me. For example, let's say I want to build a username matching function. I should be able to provide some example positive and negative matches, expand those examples into a small eval dataset using one-click synthetic data expansion, and then have the hub go and try hundreds of models against that eval dataset and find the top performers in several parameter size classes.

Β·

can you try https://hf.co/playground and tell me if this helps for your use case?

Bug?:

From the HF Forum.
There seem to be frequent cases where the proper load distribution to each GPU in a multi-GPU environment does not work.
It is difficult to distinguish whether this is a bug or not, but if it is a bug, it is difficult to know which library the bug is from and difficult to report it, so I am writing about it here.
https://discuss.huggingface.co/t/cuda-memory-issue-for-model-generate-in-automodelforcausallm/74216/1

I'm going to write a crappy poem because it just popped into my head. I'm not a forum troll, although I do post too much. I'm just a guy with time on his hands.
Let me say in advance that I really love the OSS community for its worldliness, its looseness, its nice people, and its pace.
But that doesn't mean I don't see the problem.

I came up with an analogy that makes the current problems with HF easy to understand for those who like video games.
In a nutshell, the current HF is like a β€œDried squd (Japanese traditional food) game” or β€œKusoge” or incomplete "Minecraft".
"Dried squad” is a mainly Japanese food that is hard and messy, but the more you chew it, the more the flavor seeps out and the tastier it gets. It has a certain number of lovers.

Think of all the games that have been popular over the past 40 years. They were mostly good at tutorials, level design, visuals, music, and above all, how to comfortably cripple the user. Or they were lucky to get a lot of users at the start.
It's not how free you make the game, but how you create a stress-free yet crippling situation that is important to attract consumers. Why do we need to attract consumers? Every model author wants feedback. For that, we need a population. There will be exceptions. Some people don't like the noise, and I don't either, but the absence of an audience is more of a problem.
Even in open-world games with a high degree of freedom, there is a tutorial and you are given a set of initial equipment that is weak but easy to understand. The first enemies look weak, and the battle background music is also kind of weak mood. Even when unknown enemies appear, hints are usually provided before they appear. The game is designed to keep you hooked.
There's not much you can do in stage 1 of Super Mario, right? That's actually the important thing.

The current HF is the exact opposite. It is designed to do as much as possible, and to avoid limiting use cases as much as possible. What happens as a result is that, analogous to an open world game, you are not given initial equipment and have to self-serve to find it. You are not even told what the clear conditions are. Or you need to get strategy information on an outside forum and then come here. Tutorials are either non-existent or the starting point is hard to find out how to look for them. The enemies (Spaces and Models, Datasets) you see wandering around are at first glance unrecognizable, and you can't tell if they are just mooks or demon kings. Who should you engage in combat with?
One of the worst game designs is that you can do everything but don't know what to do.
But if we reduce what we can do, the HF itself will lose its meaning.

If HF has a marketing guy (who thinks about and improves user demand and experience, not sales pitches), he maybe good to learn about the basics of game design, even if it's just on YouTube. In Japanese, the one by Sakurai, the creator of Smash Bros. is excellent.
If only people could understand that an HF UI designed to do just about anything is synonymous with not being able to do anything for an outsider.
The game balance can be adjusted later.

Doing anything is the flip side of the coin of not being able to do anything. Except for those with hackerish personalities.
Simple demos are more popular, right?

That said, give me a $20 personal Zero GPU space plan so I can build a community tool, and I can build a converter in a high performance space that doesn't have access to a GPU, but 10 spaces is too few for a permanent installation anyway, and the longer I stay at HF, the more inconvenient it becomes, How funny is that?
I don't mind if you strengthen the Enterpsise plan, but the custom of one person calling himself an organization is not in Japan, it's very uncomfortable. Is it a major practice in other countries?

Thanks.

Post:

  • The Forum has a log search function, but I don't think it's in the Post. At least not installed on the screen.

Notifications:

  • I would like to see Notifications come when a Forum Message (Mail) is received.
  • I would like to see an extraction feature like a model or dataset search screen. Just a word search would be fine.
  • I would like Notifications to be able to be subdivided and structured in a semi-automatic manner. I don't need as much as Gmail or Thunderbird, so something simple. It may be smart to augment the tabs that are currently already in place.
  • I would like to see more variations in the color change of the icon when Notifications arrive. Currently only blue and yellow. I can't even tell the difference between an emergency notification and a Parquest Bot, which is a nice guy, but that's not the point.

Collections:

  • I would like to be able to do Collections for Collections.
  • It would be better if other HF resources could be added to the Collections, e.g. URLs of Posts, Forums and Discussions.
  • Structured or hierarchical Collections would be nice

Forum:

  • We're probably in a situation where the forum trolls are spiking, and HF better get ready to hire more response personnel. There is no point in trying to figure out why there is an increase, but at any rate, this type of activity has increased dramatically in many areas over the last year or so, and many communities have been irreparably damaged. The forum vandalism is not likely to end up being a forum problem, as it has also made its way into the Discussion section.
  • It seems that the invitation link from the forum to the HF Discord expired a long time ago and has never been used again. Virtually no new people can join. I don't have a Discord account, so I don't really care.
  • There is a critical shortage of respondents on the forum, but there are far more decent and well-informed question posts than one might expect. It could be a resource if only there were enough respondents.
  • It seems to me that about half of the questions raised in the forum that seem to be error-related can be resolved with a proper search on the error content. However, this seems to be more difficult than expected for beginners. How about having a chatbot with a search function handle the initial response? It would be confusing, so it would be separate from regular submissions.
  • There are often submissions of library glitches, specifications that are not good, repo glitches, etc. Wouldn't it be easier to make the initial response if we could have a chatbot summarize these and deliver them to the appropriate parties where each person can view them at their discretion?

Spaces:

  • There seem to be a few things that are impossible to install when Spaces starts up, even with pre-requirements.txt or packages.txt, and calling subprocess directly in Python. Some VMs have an environment variable that enhances permissions only at startup, and HF seems to have one as well, but using it does not seem to yield satisfactory results.

QOL:

  • I think there are some Organizations that are supposed to be official but have no maintainers. Specifically, there is no response to commits made to the Diffusers community; I could send mentions to multimodalart or sayakpaul, but this would be a bad situation in the long run.
  • The situation that the HF staff themselves do not use the built-in community very much may be an important clue to improve the community function.
  • I don't know if the concept of being in charge of something exists on the HF staff itself, but there is no list of staff in charge of each issue or section, so it's impossible to even send a mentions, except to a few people I happen to know. I can try to trace them through Organizations, but there are too many people and too many non-staff members.
  • It would be better to have a prominent information board about what can be called official infrastructure, such as Spaces belonging to Organization and Utilities. I don't think anyone but the heavy users who have been around since the beginning would be able to figure out where everything is.
  • Propose an inspection of the official's infrastructure. To use an urban function as an analogy, tourists would be horrified to see a dilapidated train station or an unmanned government office, even though in real life everyone has found their own loopholes and is living without any problems.
  • Regarding the management of libraries, as in the case of free software in ancient times, it would be better to leave them alone if it is software or functions that programmers themselves create for themselves as users, or if it is work to support a new AI model, or simply bug fixes. However, without some guidelines apart from those, the developer usually cannot understand the needs of the users and tends to continue to make modifications that miss the mark and benefit no one. This is not limited to libraries.
    Many OSS developers don't like to be told what to do, and I don't either, but can't HF put someone or some mechanism in place to loosely direct the direction of library development? It's easy to understand the problem if you imagine an Apple without Jobs or an unwanted new feature in Windows.
  • It is good to use this kind of opinion poll to set guidelines, but it is better to use it as a help to understand the needs of many developers and consumers in the world, not to adopt the opinions of a noisy minority like me, and to look over the situation again on the HF side and devise features that would improve the situation if they existed, I think it would be easier to achieve better results. In general, it is out of the question to not listen to what customers say, but it is also not good to just take their word for it.
  • The same could be said of library development with regard to overall HF retailing.

Great post! I think improving the documentation would be really helpful for new users.

QoL:

Let's add system requirements to models.
It’d be super helpful to require or at least suggest that AI models include system requirements, just like other software. Minimum and recommended specs, especially for inference with Hugging Face libraries, would make things easier. Hardware info is often hard to find, and not everyone has access to H100 clusters. Setting this as a standard would make models way more accessible.
discussion here

CDNs updates page

CDN URLs are changed from time to time. Example in the past there was http://cdn-lfs.huggingface.co/ which is now cdn-lfs.hf.co.
It is causing many issues especially when we need to add URLs to whitelist.

It would be useful keep a page somewhere under huggingface.co updated with the available CDNs to allow users use it as a reference and get to know about the changes in advance.

Please provide more transparency to PRO account ZeroGPU quota limit.

My spaces run fine using the interface, but they won't run through API at all. Does that mean there are two different pools of quota for API & interface?

The error message was: The upstream Gradio app has raised an exception: You have exceeded your GPU quota (60s requested vs. 52s left). Create a free account to get more usage quota.

Even if I wait myself for a few days, the GPU quota available to me is still lower than 60s. Is there something wrong behind the stage?

For now, I am forced to convert all my Gradio API repo into the front end repo, which makes it gigantic and slow to rebuild and debug.

Β·

and now I have consolidated all my api spaces to one single space in front end... but then the new problem is:

Traceback (most recent call last):
File "/home/user/app/app.py", line 399, in
demo.queue(max_size=16).launch(
File "/usr/local/lib/python3.10/site-packages/spaces/zero/gradio.py", line 142, in launch
task(*task_args, **task_kwargs)
File "/usr/local/lib/python3.10/site-packages/spaces/zero/torch/patching.py", line 348, in pack
_pack(Config.zerogpu_offload_dir)
File "/usr/local/lib/python3.10/site-packages/spaces/zero/torch/patching.py", line 340, in _pack
pack = pack_tensors(originals, fakes, offload_dir, callback=update)
File "/usr/local/lib/python3.10/site-packages/spaces/zero/torch/packing.py", line 114, in pack_tensors
os.posix_fallocate(fd, 0, total_asize)
OSError: [Errno 28] No space left on device

Please make the quota limit shared between interactive usage & api usage under a same account!