Tag
Explore by tags
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.
Provides code, pretrained weights, and tooling for protein language models and structure prediction — including ESMC, ESMFold2, sparse autoencoders (SAEs), and the ESM Atlas. Includes model checkpoints, tutorials, Hugging Face & Biohub integration, and an MIT license.
A research codebase and model family for vision–language models that experiments with data‑centric post‑training strategies and long‑context multimodal reasoning. Includes model reports, released research weights (non‑commercial), grounding tools (LocateAnything) and integrations for inference/optimization.
Official code companion to the O'Reilly book by Jay Alammar and Maarten Grootendorst: 12 chapters of runnable notebooks on tokens, embeddings, Transformers, text classification, clustering, prompt engineering, semantic search, RAG, and fine-tuning.
Runs huge mixture-of-experts LLMs like DeepSeek-R1/V3 on a single 24GB GPU plus CPU DRAM by keeping attention on the GPU and offloading expert weights to CPU. Reports 3-28x speedups via Intel AMX/AVX512 kernels and fits 139K context in 24GB VRAM.
Chains four swappable open modules — voice activity detection, speech-to-text, an LLM, and text-to-speech — into a local voice agent that needs no proprietary APIs. Runs on CUDA, Apple Silicon, or Docker, with an OpenAI-compatible realtime WebSocket mode.
Trains a sub-100M-parameter LLM from scratch — pretraining, SFT, LoRA, DPO/RLHF, and distillation, sized from ~26M up to ~100M-plus dense and MoE. Headline figure: the ~64M minimind-3 variant's SFT stage runs 1 epoch in ~2h and ~3 RMB on one NVIDIA 3090.
Runs and optimizes ML and generative-AI models on-device across mobile, desktop, web, and IoT. Successor to TensorFlow Lite, it adds automated GPU/NPU accelerator selection and zero-copy buffer interop to cut latency without cloud round-trips.
Trains a 65M-parameter vision-language model from scratch in ~2 hours on one RTX 3090, about 3 RMB (~$0.40) of GPU rental. Connects a frozen SigLIP2 encoder to a small MiniMind LLM via a two-layer MLP projector; full PyTorch code for pretraining and SFT.
GPU‑accelerated framework for training physically simulated humanoid characters and robots using reinforcement learning and motion imitation. Provides a modular multi‑backend simulator stack, large‑scale multi‑GPU training recipes, built‑in motion retargeting and an ONNX deployment pathway to real robots.