AI Infra2025
Optimized MLA (Multi-head Latent Attention) decoding kernels powering DeepSeek-V3/V3.2 inference on Hopper and Blackwell GPUs. Dense decoding reaches ~3000 GB/s and 660 TFLOPS on H800; the sparse path stores the KV cache in FP8.
AI Train2025
inclusionAI (AReaL Team), Institute for Interdisciplinary Information Sciences, Tsinghua University +1Ant Group, Institute for Interdisciplinary Information Sciences, Tsinghua University
Trains LLM reasoning and agentic models with fully asynchronous reinforcement learning, decoupling rollout generation from policy updates for a 2.77x speedup over synchronous RL. Covers GRPO, PPO and DAPO across Megatron, FSDP, vLLM and SGLang backends.
Embodied AI2025
A vision-language-action foundation model and reference stack for generalized humanoid and cross-embodiment robot manipulation. Provides pretrained checkpoints, demo datasets, and tooling for fine-tuning, evaluation, and deployment (ONNX/TensorRT); released as Early Access.
AI Train2025
Trains multi-step LLM agents with reinforcement learning (GRPO) on your own tasks, wrapping existing agent code behind an OpenAI-compatible client. Its RULER mode scores trajectories with an LLM judge, so there's no reward function to hand-write.
AI Image2025
Real-time DETR detector on a DINOv2 backbone, covering detection, segmentation, and keypoints. Ships in six sizes (Nano to 2XL), beats YOLO on the COCO speed-accuracy curve, and transfers better to non-COCO real-world domains.
AI Agent2025
Autonomously proposes hypotheses, runs experiments, analyzes results, and drafts workshop-level papers via an agentic tree-search pipeline. Unlike template-driven predecessors, it explores open-ended ML research paths but requires GPU/PyTorch and careful sandboxing due to execution of LLM-written code.
AI Audio2025
Open-source TTS that clones a voice from a short reference clip across 23+ languages, with adjustable emotional intensity via exaggeration/cfg controls and a built-in Perth neural watermark on every output.
AI Video2025
Provides PyTorch code, pretrained checkpoints, and evaluation tooling for V-JEPA 2 — a Meta FAIR family of self-supervised video encoders and an action-conditioned world model. Includes training recipes, HuggingFace checkpoints, evaluation probes, and robot post‑training artifacts.
AI Image2025
Converts document images—scans, photos, born-digital PDFs—into structured text in two stages: first map layout and reading order, then parse each element (text, tables, formulas, figures) in parallel, each guided by its own task prompt.
AI Train2025
A PyTorch DTensor-native SPMD library for training and fine-tuning LLMs, VLMs, diffusion and retrieval models. Integrates with Hugging Face for day-0 model support, provides YAML-driven recipes, DTensor/FSDP2 parallelism and NVIDIA-optimized kernels (Transformer Engine, DeepEP, FlexAttn).
AI Train2025
Enables bidirectional checkpoint conversion between Hugging Face and Megatron formats and provides a PyTorch-native training library with tensor/pipeline parallelism, FP8/BF16 mixed precision, SFT and PEFT (LoRA) support for large and multimodal models.