Le meilleur du dev et de l'IA, scoré chaque jour par un agent. Filtré, résumé, classé. Aucune couleur, aucun bruit — juste la matière.
Sentence Transformers can finetune sparse encoder/embedding models for retrieval, hybrid search, and reranking. The post outlines training components (model, datasets, losses, trainer, evaluators) and provides practical examples, including using naver/splade-v3 and decoding sparse embeddings for interpretability.
Le NVIDIA Llama Nemotron Nano VL est un modèle Vision-Language 8B optimisé pour l’identification et l’OCR de documents, offrant une haute précision OCRBench v2 et une extraction fiable de texte, tableaux et éléments visuels. Disponible sur Hugging Face, il s’appuie sur Llama-3.1-8B-Instruct et C-RADIOv2-VLM-H et peut être post-entraîné via NVIDIA NeMo; un tutoriel Notebook permet de construire rapidement des solutions d’automatisation pour factures, contrats et documents de santé.
Gemma 3n is now available in major open-source libraries (transformers, timm, llama.cpp) and runs on-device with multimodal support (text, image, audio, video). It ships two sizes (5B/8B real; 2B/4B VRAM), a flexible MatFormer architecture, and mix-and-match sub-models. It includes a MobileNet-V5 vision encoder and a USM audio encoder, supports 140 languages, and offers a Hugging Face Space demo.
SGLang now supports Hugging Face transformers as a backend, enabling high-throughput, low-latency inference for any transformers-compatible model. It can auto-fall back to transformers or be forced with impl='transformers', and adds features like RadixAttention for efficiency. The article includes practical code snippets and deployment patterns (offline engine, server, and OpenAI-compatible API).
New post shows efficient fine-tuning of FLUX.1-dev on consumer hardware via QLoRA with the diffusers library, targeting peak VRAM under ~10 GB on a single GPU. It explains loading a quantized 4-bit base model, training FP16/BF16 LoRA adapters, uses an 8-bit AdamW optimizer, and discusses options to load or merge LoRA adapters with results demonstrated on an RTX 4090.
Groq is now a supported Inference Provider on the Hugging Face Hub, enabling serverless inference on model pages and tight integration with Python and JavaScript SDKs. The LPUs promise lower latency and higher throughput for LLMs, with support for models like Meta's Llama 4 and Qwen's QWQ-32B, and two billing modes: direct provider key or routing through HF.
The article analyzes how long prefill prompts can block the prefill queue in a multi-request setting, and explains that decoding steps are light but must be sequential. It discusses two patterns - chunked prefill and request-parallel prefills - and why long prompts undermine throughput, with implications for vLLM scheduling.
Hugging Face's Kernel Hub lets Python apps load pre-compiled, optimized kernels directly from the Hub, avoiding local builds. It includes a quick code example to fetch a kernel (e.g., activation) and apply it, and discusses integrating kernels into models like RMSNorm and FlashAttention. The article also covers performance benchmarking and real-world use cases.
Featherless AI est désormais supporté comme Inference Provider sur Hugging Face Hub, permettant l'inférence serverless directement sur les pages des modèles et accessible via les SDK JS et Python. Il prend en charge un large éventail de modèles open-source et offre deux modes d’appel (clé personnalisée ou routée par HF) avec une tarification directe sur le compte utilisateur. Des exemples Python et JS montrent comment l'utiliser avec Featherless AI.
Cet article présente GR00T N1.5 et explique comment réaliser un post-entraînement sur le bras LeRobot SO-101. Il propose un tutoriel pas-à-pas couvrant l'installation, la préparation du dataset et le fine-tuning, puis l'évaluation et le déploiement. Des commandes et configurations (modality.json, scripts/gr00t_finetune.py) permettent une adaptation rapide du modèle à votre robot.
Hugging Face and NVIDIA announce Training Cluster as a Service to give researchers easy access to large GPU clusters for training foundational models, paying only for training durations. The solution combines NVIDIA DGX Cloud Lepton, regional capacity, and Hugging Face tooling to provision, price, and manage clusters, with user requests via hf.co/training-cluster.
ScreenSuite est une suite de benchmarking complète pour évaluer les GUI Agents, couvrant perception, grounding, actions simples et multi‑étapes à travers 13 benchmarks. Elle propose des environnements Docker (Ubuntu/Android) et des sandboxes distants pour tester des agents GUI en conditions reproductibles, en restant vision‑only.
L’article décrit la mise en œuvre du KV Caching dans nanoVLM (PyTorch), qui réduit la redondance de calcul lors de la génération autoregressive en réutilisant les K et V précédents, aboutissant à ~38% de gain en vitesse. Il clarifie où la redondance apparaît dans l’attention et propose un exemple PyTorch minimal pour illustrer l’approche.
An on-device, Arm-powered sound generator lets you describe a sound and obtain a .wav within seconds, without cloud or GPU. Built around Stability AI’s Stable Audio Open model, it runs via PyTorch and TorchAudio with optimized multithreading, and integrates seamlessly with Ableton Live.
Surfer-H is a web-native agent powered by the Holo1 family of Action Vision Language Models (VLMs), enabling GUI automation and precise UI localization. Holo1-3B and Holo1-7B achieve up to 76.2% UI localization accuracy on benchmarks, with open-source releases on Hugging Face and a WebClick benchmark of 1,639 tasks. Surfer-H operates entirely in-browser with a three-component architecture (Policy, Localizer, Validator) and claims cost-efficient performance.
SmolVLA-450M is an open-source, compact Vision-Language-Action model for robotics that runs on consumer hardware. It is pretrained on publicly licensed community data and trainable on a single consumer GPU, with asynchronous inference boosting throughput and strong performance on simulation and real-world tasks.
TRL now runs vLLM co-located with training, sharing GPUs to eliminate idle time and HTTP overhead. By embedding vLLM in the same process group, training and inference take turns on the same devices, with torchrun compatibility, TP/DP support, and GRPO-enabled workflows. The article covers design, implementation notes, and benchmarks (1.5B, 7B, 72B) plus a train_grpo_colocate.py script to try.
Cet article démontre que forcer CodeAgents à générer pensées et code dans un JSON structuré peut surpasser les approches traditionnelles sur des benchmarks comme SmolBench. Il explique pourquoi la structure améliore l’exécution des actions et propose des conseils d’implémentation et des cas d’usage pratiques pour les développeurs IA.
Report on debugging Liger GRPO loss with DeepSpeed ZeRO-3 using Qwen/Qwen2.5-0.5B-Instruct in bf16, highlighting a shape mismatch during training. The traceback traces through grpo_loss and fused_linear_ppo, pointing to a forward pass issue in the Liger kernel. No fix is shown, but it identifies the code paths to inspect (grpo_loss, fused_linear_ppo).
Dell Enterprise Hub now ships a complete suite of on‑prem AI models and ready‑to‑deploy applications (e.g., Llama 4 Maverick, OpenWebUI), optimized for Dell AI Server platforms via Docker/Kubernetes. It adds an Application Catalog and a dell-ai CLI/Python SDK to run everything from development to on‑prem deployment, plus on‑device models for Dell AI PCs powered by Intel/Qualcomm NPUs.