#cuda
DeepEP is a special communication library for Mixture-of-Experts (MoE) models. It helps these models work faster and more efficiently by improving how data is shared between different parts of the system. DeepEP supports low-precision operations and can handle data transfer between different types of connections, like NVLink and RDMA. This makes it very useful for both training and using AI models, especially when speed is important. Users benefit from faster processing times and better performance overall.
https://github.com/deepseek-ai/DeepEP
Пока весь мир ждет доступа к новой модели со зрением GPT-4V(ision), опенсорс команда (пара азитов со степенью PhD из американских вузов) уже выпустили свой аналог и бесплатную версию #LLaVA (Large Language and Vision Assistant), которая выдает результат (не) хуже GPT4V и может работать локально.
Вот такая скорость развития и конкуренции в этом новом #AI рынке.
🧠LLava - вебсайт
📄WhitePaper
🧬Github code
🔋Demo для потестить на своих дикпиках
🦒Colab (для запуска у себя на серваке)
#python#apple_silicon#florence2#idefics#llava#llm#local_ai#mlx#molmo#paligemma#pixtral#vision_framework#vision_language_model#vision_transformer
MLX-VLM lets you run, chat with, and fine-tune Vision Language Models (VLMs) plus audio/video models on your Mac using MLX—install easily with `pip install -U mlx-vlm`. Use CLI for quick text/image/audio generation (e.g., `mlx_vlm.generate --model ... --image photo.jpg`), Gradio UI for chats, Python scripts, or a FastAPI server with OpenAI-compatible endpoints supporting multi-images/videos. Features like TurboQuant cut KV cache memory by 76%, and LoRA/QLoRA fine-tuning works on consumer hardware. You benefit by experimenting with powerful multimodal AI locally—fast, memory-efficient, no cloud costs, perfect for Mac users tweaking models affordably.
https://github.com/Blaizzy/mlx-vlm