Tag
Explore by tags
Pocket-sized multimodal LLM for efficient image- and video-understanding on mobile and edge devices, featuring mixed 4x/16x visual-token compression (MiniCPM‑V 4.6), compact 1.3B variants, and ready guides for iOS/Android/HarmonyOS deployment.
Runs GPT-4o-class vision, speech, and full-duplex audio-video conversation on a 9B model small enough to deploy on phones and tablets. The 4.5 release scores 77.6 on OpenCompass and adds real-time bilingual voice with voice cloning.
GPU kernel library for LLM inference attention, sampling, and KV-cache, built on block-sparse formats with JIT-compiled customizable templates. Reports 29-69% inter-token-latency cuts vs compiler backends; powers SGLang, vLLM, and MLC-Engine.
Provides a lightweight build platform for HIP and ROCm that supports building ROCm, PyTorch, and JAX from source, multi-architecture nightly releases, and integrated CI/CD and developer tooling for Linux and Windows.
Generates short videos from text, images, or videos and ships a full training/inference pipeline with checkpoints and demos. Key features include multi-stage training (VAE / 3D-VAE), rectified-flow training, video compression modules, and support for 2s–16s clips at up to 720p. Best for researchers and engineers who can provide substantial GPU resources.
End-to-end framework for running and reproducing foundation-model research workflows — from data curation and tokenization to training and evaluation. Emphasizes reproducibility by recording every step (including failed runs) and expressing experiments as dependency-ordered steps.
Hands-on coding tutorial series for large language models with slides and runnable notebooks covering fine-tuning, prompting, RLHF, safety, steganography, watermarking, multimodal models, GUI agents, and deployment. Community-maintained, free course materials for students and researchers.
Applies quantization, pruning, distillation, speculative decoding and sparsity to compress and accelerate deep learning models, producing exportable checkpoints ready for deployment on TensorRT‑LLM, TensorRT, vLLM and similar inference runtimes.
Accelerates video generation with a unified framework for inference, finetuning, LoRA, distillation, sparse attention, and distributed execution for research and demos.
Pretrained foundation model for zero-shot point and quantile forecasting on unseen time series, no per-dataset training. Decoder-only, trained on 100B real-world time-points; v2.5 (200M params) handles up to 16k context and 1k-step horizons.
Stores and reuses LLM key-value caches across GPU, CPU, disk, and remote backends so vLLM and SGLang skip recomputing repeated context. Non-prefix reuse (CacheBlend) and PD disaggregation cut time-to-first-token for long-context and RAG serving.
A community speedrun to train a 124M GPT as fast as possible on 8 H100s, all chasing a fixed 3.28 FineWeb loss. Successive records cut the run from llm.c's 45 minutes to under 1.4, mostly via the new Muon optimizer rather than more hardware.