Tag

DeepEP

Provides high-throughput, low-latency GPU communication kernels for Mixture-of-Experts (MoE) and expert-parallel workloads, with NVLink↔RDMA-aware forwarding, FP8/BF16 support, and low-latency RDMA hooks for inference decoding.

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.

MLX LM

Runs and fine-tunes LLMs locally on Apple silicon via the MLX framework, pulling thousands of Hugging Face models with one command. Adds 4- and 8-bit quantization, LoRA and full fine-tuning, prompt caching, and distributed inference across Macs.

NVIDIA Dynamo

Splits LLM inference into separate prefill and decode GPU pools, then routes requests with KV-cache awareness to cut redundant recomputation. Reports up to 30x throughput on DeepSeek-R1 (GB200 NVL72) and works across TensorRT-LLM, vLLM, and SGLang.

RF-DETR

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.

LiteRT-LM

Run large-language and multimodal models locally on edge devices (Android, iOS, desktop, web, Raspberry Pi) with hardware acceleration, function-calling, and multi-language SDKs—designed for low-latency, privacy-sensitive on-device inference.

AIPerf

Measures generative AI inference performance with token-level metrics (TTFT, inter-token latency), latency, and throughput under realistic traffic patterns. Provides a multiprocess engine, real-time TUI dashboard, extensible plugins, and integrations for telemetry and result uploads, aimed at inference benchmarking and capacity planning.

Dolphin

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.

Parses the local JSONL logs that coding-agent CLIs write and turns them into token and cost reports, no API keys or telemetry. Breaks spend down by day, month, session, and Claude's 5-hour billing windows across Claude Code, Codex, Gemini CLI and more.

Turns commodity WiFi Channel State Information into spatial sensing: 17-keypoint pose estimation, presence detection, and contactless breathing/heart-rate monitoring through walls, with no camera. Runs on a mesh of ESP32-S3 nodes (~$9 each).

Nano-vLLM

Reimplements the vLLM inference engine from scratch in ~1,200 lines of readable Python, matching its offline throughput on small models. Prefix caching, tensor parallelism, torch.compile, and CUDA graphs are all kept legible.

Magenta RealTime 2

A toolkit and open-weights system for real-time streaming music generation — offers two model sizes (230M / 2.4B), a Python inference library (JAX/MLX), and a C++ engine optimized for Apple Silicon for embedding into DAWs and apps; real-time streaming requires M‑series chips.