A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit and 4-bit floating point (FP8 and FP4) precision on Hopper, Ada and Blackwell GPUs, to provide better performance with lower memory utilization in both training and inference.

cuDNN Frontend is NVIDIA's modern, open-source entry point to the cuDNN library and a growing collection of high-performance open-source kernels.

Microsoft Automatic Mixed Precision Library

An innovative library for efficient LLM inference via low-bit quantization

Kernel Library for Large Headdim Attention (64~1024, BF16/FP8/FP4), 1.5x~15x↑ vs PyTorch SDPA.

Flux diffusion model implementation using quantized fp8 matmul & remaining layers use faster half precision accumulate, which is ~2x faster on consumer devices.

Optimized SGLang runtime for Qwen3.8-27B FP8 with DFlash2 and Qwen3.8 Flash-Next NVFP4 with FR-Spec on one NVIDIA RTX PRO 6000 Blackwell 96 GB GPU (SM120): 524K context, HiCache and NIXL.

Qwen3.8-Flash-Next on 2× RTX 3090: with 128 GB RAM up to 4,191 tok/s prefill · 111.5 tok/s decode (131K prompt), full 256K window at 2,865 tok/s; with 64 GB RAM 3,410 tok/s prefill · 84 tok/s decode. New: opt-in uncensored mode (runtime abliteration, no new weights).

Measured SM121 compatibility recipe for MiniMax H3 FL2VA on one NVIDIA DGX Spark with vLLM-Omni and online FP8.

Bit-exact emulation of NVIDIA tensor-core FP8 GEMMs (torch._scaled_mm, Transformer Engine) as an H100 PCIe or RTX 4060 computes them, on the CPU or another CUDA GPU

Serving Qwen3.8-27B-FP8 on a single DGX Spark (GB10): 7.88 to 58.5 tok/s single-stream from decode strategy alone, weights untouched. Speculative decoding and prefix caching benchmarked, plus DFlash 2 — the only Qwen3.8-27B build that can serve it under vLLM.

Pure Go machine learning framework. Train, run, and serve ML models with go build. Zero CGo.

vLLM 0.32 fork for consumer GPUs: 3 x 262K sessions on 4x RTX 3090, Qwen 3.8 Flash-Next at 184 tok/s, up to 17 on bigger cards. Seven checkpoints incl. Qwen3.8-27B, Nemotron 3.5 Super VL (1M) and Ornith 1.5. One command: merlin serve.

Abliterated DeepSeek-V4.1-Flash-1017 (FP8 dense, FP4 experts) on 4x RTX PRO 6000 with vLLM-Moet r21 and DSpark: 371-376 tok/s single-stream agent decode at 30k-120k context, 1M context, Flex-start GCE VM

Lynn 原生 LLM 推理引擎 · W4A8/NVFP4 量化 · 自写 CUDA/Triton kernel · MoE · 投机解码 | Lynn-native LLM inference engine for NVIDIA Blackwell

FP8 Metal compute kernels for Apple Silicon MPS — fixing what PyTorch doesn't support yet. FLUX/SD3.5/ComfyUI on Mac.

Benchmarks and notes for running modern LLMs with vLLM on 8x Tesla V100-32GB in 2026.

LLM inference engine for AMD RDNA4 — Rust + Vulkan compute shaders, gguf & native FP8.

JAX Scalify: end-to-end scaled arithmetics

CUDA-only Triton W8A8 block-FP8 kernel tuner for vLLM, with Qwen3 shape detection, custom shapes, and multi-GPU tuning.