Hugging Face reposted this
Gradio Workflow is on Product Hunt! Connect nodes to build AI pipelines, powered by Hugging Face 🤗 ⬇️ PH link is in the first comment.
About us
The AI community building the future.
- Website
-
https://huggingface.co
External link for Hugging Face
- Industry
- Software Development
- Company size
- 51-200 employees
- Type
- Privately Held
- Founded
- 2016
- Specialties
- machine learning, natural language processing, and deep learning
Products
![Click here to view Hugging Face]()
Hugging Face
Natural Language Processing (NLP) Software
We’re on a journey to solve and democratize artificial intelligence through natural language.
Locations
-
Primary
Get directions
-
Get directions
Paris, FR
Employees at Hugging Face
Updates
-
-
Hugging Face reposted this
Tokenizers v1's first release candidate is out 🔥 absolutely huge release across the board: Up to 30x faster tokenization, rivaling the best in the ecosystem. Scaling across threads, much better latency, less memory consumed, and microscopic crate size. We're releasing a blogpost going deep into our changes, and benchmarking the library against itself (v1 vs v0.23) and relative to other tools in the ecosystem. Come take a look!
-
-
Hugging Face reposted this
We're open-sourcing hfdask: run #Dask workloads on Hugging Face Jobs. Point it at a small config file describing the machines you need. It starts them, runs your Dask code across them, and tears them down when the work is done. Context, and a genuine thank you to the Hugging Face team. We like working in this ecosystem, and more of our stack keeps moving that way. Jobs gives us CPUs and GPUs on demand. Datasets, models, and buckets are already there. When your bottleneck is data, compute sitting next to it is the whole game. Most synthetic data work ends up looking like the same pipeline: CPU → GPU → CPU. CPUs read and filter source documents, GPUs run the generation, CPUs validate what came back and write the dataset. Generating the text is the easy part. Deciding which of it is actually training data is the work. We ran this in stages, which is where most teams start. Hand-chunk the dataset, launch a GPU job per slice, write results to a bucket, wait, then wake a CPU job to read the bucket back. It worked, but nothing in it knew about anything else. Every stage ended by writing its results to a bucket, and the next stage began by reading them back. Nothing moved until someone started it. Dask has been solving exactly this for years: scaling work across many machines while keeping the code clear and readable. It just had no way to run on Hugging Face Jobs. That's the gap hfdask fills. We built it for our own training data, and we're giving it away. What you get: - Your code stays normal Dask code. Nothing to rewrite, no new framework to learn. - Ask for a GPU on a stage and it runs on a GPU. Everything else runs on CPU. - Every CPU core gets used. Every GPU is dedicated to a single worker. - Nothing is exposed to the public internet. The machines connect to each other over an encrypted link. - Your exact code and dependency versions ship with the run, so every machine runs the same environment. - It shuts the paid machines down when the run ends, and checks that they actually stopped. A forgotten GPU job is a bill. There's a full write-up on the Hugging Face blog that walks the whole pipeline end to end, validation included. Links in the comments. 💅 #HuggingFace #OpenSource #MLOps #SyntheticData
-
Hugging Face reposted this
Helion is now supported in 🤗 Kernels! Write portable kernels in Python. Tune them once. Share on the Hub. Helion attention kernel on H100: 1.2× faster on average than PyTorch’s FlashAttention across 19 tuned shapes. We collaborated with the Helion team at Meta to also write a blog post about this. Please find the link in the first comment of this post. Dunfan Lu it was a pleasure collaborating with you.
-
-
Hugging Face reposted this
1.49 million illustrations from the 1500s to the 1920s, now searchable in one place. I've brought together images from British Library books and the Encyclopaedia Britannica in Historical Illustration Search. Search with a description, pick an image to find similar ones, and download full-resolution images or transparent PNG cutouts. Both collections are searchable together, with filters if you want to explore just one. Source links let you follow an illustration back to its original context, and the underlying datasets are available too. Try it on Hugging Face: https://lnkd.in/emjsptXN
-
Hugging Face reposted this
The Hugging Face Kernels project now has Helion support. This blog walks through how to build, autotune, and ship performant and portable Helion kernels via the Hugging Face Kernels project, allowing users to consume these kernels seamlessly. Learn more here: https://lnkd.in/eSeyJDmf Sayak Paul, Dunfan Lu, Tarindu Jayatilaka, Jongsok Choi
-
-
Hugging Face reposted this
Join the agent swarm that finds the ultimate compression algorithm! Announcing: Hutter Prize challenge 🏆 Hugging Face > join the org, give your agent a write token for the org > click “add your agent”, copy/paste the snippet into your Codex/Claude Code/Hermes Agent > watch it collaborate!
-
Hugging Face reposted this
We finished the Training Agents series. Six live sessions over six months, from evaluating agents to training them inside real environments. All of it is on the Hugging Face YouTube channel and all of the code is open. Here's what we did and who made it happen: 1. Agentic Evaluations WorkshopWhere agent evals actually stand, and why benchmark scores don't match what people see in use. With Avijit Ghosh and Nathan Habib (Hugging Face), Arvind Narayanan (Princeton), Pierre Andrews (Meta), J.J. Allaire (UK AI Security Institute) and Mahesh Sathiamoorthy (Bespoke Labs). 2. RL for Agents Workshop Environments, rollouts, reward design and the inference bottlenecks that appear when you move from RL for LLMs to RL for agents. With Lewis Tunstall (Hugging Face), Will Brown (Prime Intellect), Ofir Press (Princeton) and Alex Zhang (MIT CSAIL). 3. Training Agents 1: SFT on agent traces Public coding-agent traces turned into prompt/completion data, a TRL + LoRA fine-tune on Hugging Face Jobs, metrics in Trackio, and an honest look at what the first eval numbers can and cannot tell you. Joined by Sergio Paniego and Quentin Gallouédec. 4. Training Agents 2: Distillation Off-policy, on-policy and self-distillation for moving capability from a teacher into a smaller coding agent. 5. Training Agents 3: Reinforcement learning GRPO after SFT: group sampling, verifiable reward functions, reading the reward/KL/length curves, and three experiments, one of them with a deliberately gameable reward so we could watch the hacking happen. 6. Training Agents 4: From reward functions to environments The reward stops being a function and becomes a place the agent acts in. We walked the reset()/step() contract from Gym to LLM agents, built an OpenEnv environment and pushed it to the Hub, plugged it into TRL's GRPOTrainer, then trained a real coding agent (OpenCode) through Harbor with AsyncGRPOTrainer on Hugging Face sandboxes. The series has passed 300k views. Thank you to every speaker, to the TRL team, and to everyone who showed up live with questions. Playlist: https://lnkd.in/esUURS7b
-
-
Hugging Face reposted this
Join the agent swarm that finds the ultimate compression algorithm! Announcing: Hutter Prize challenge 🏆 Hugging Face > join the org, give your agent a write token for the org > click “add your agent”, copy/paste the snippet into your Codex/Claude Code/Hermes Agent > watch it collaborate!
-
Hugging Face reposted this
We rebuilt most of AUTOMATIC1111 as a single Gradio Workflow canvas: 11 media pipelines, 73 nodes, no GPU required. But the interesting question isn't what we built. It's who should care. 1. Python devs shipping multi-model AI apps. Every output node becomes a REST endpoint and an MCP tool automatically. 2. ComfyUI builders who hit the sharing wall. You made a great workflow, then had to tell people to install an environment and download 20GB of checkpoints. Here, you just share a Hugging Face URL. 3. Agent builders who need media tools. Point your Claude Code, or Codex at Workflow1111 and use image and video generation tools in one shot. 4. Anyone who misses A1111 but doesn't want to babysit a local install. Txt2img, hires fix, img2img, prompt matrix, interrogate, PNG info, annotators, and img2video, all in your browser. Visitors sign in with their HF account and run it on their own quota. Links to the launch blog post and the Workflow1111 app in the comments. 🔗 ⬇️