kernel-fusion

AMD RAD's multi-GPU Triton-based framework for seamless multi-GPU programming

An agent harness that compiles a model into one provably-correct, self-retargeting CUDA megakernel and self-tunes it past cuBLAS at batch-1 LLM decode, paper: https://arxiv.org/abs/2606.09682

Fused Triton kernels for TurboQuant KV cache compression — 2-4 bit quantization with RHT rotation. Drop-in HuggingFace & vLLM integration. Up to 4.9x KV cache compression for Llama, Qwen, Mistral, and more.

Noeris — autonomous kernel fusion discovery + Triton autotuning for LLM kernels and Gemma layer deeper fusion (A100/H100 wins).

面向 Transformer 推理学习的精简 Triton 算子：RMSNorm+RoPE、Gated MLP、FlashAttention

LAMB go brrr

MLX + Metal implementation of mHC: Manifold-Constrained Hyper-Connections by DeepSeek-AI.

A CUDA framework for separating graph-recoverable launch overhead from byte-elimination gains in LLM decode-chain fusion

High-performance CUDA implementation of LayerNorm for PyTorch achieving 1.46x speedup through kernel fusion. Optimized for large language models (4K-8K hidden dims) with vectorized memory access, warp-level primitives, and mixed precision support. Drop-in replacement for nn.LayerNorm with 25% memory reduction.

Write the math. Get the kernel. Fused CUDA kernel generation from mathematical specifications.

A torch.compile backend that pays attention: custom FX fusion passes, a Triton LayerNorm kernel, and a persistent compile cache. Apache-2.0 free core of the g2n platform.

Fused STFT + mel filterbank CUDA kernel — ~3x faster log-mel spectrograms than torchaudio's GPU path, verified on NVIDIA GB10 (DGX Spark)

Experiments with optimized inference of Qwen3-ASR with vllm

Production-grade Triton kernel fusing residual add + RMSNorm + packed QKV projection into a single GPU launch for decoder-only transformer inference (Llama-3, Mistral, Qwen2). +2.4% tok/s, -1.5 GB VRAM on A10G.

GPU-accelerated volatility, market-regime, and drawdown forecasting with fused CUDA features, PyTorch, ONNX, and TensorRT.

Multi-Engine (PyTorch & JAX/XLA) Zero-Branching Geometric Acceleration Core. Enforces 0% Graph Breaks & Real-time Fault-Isolation via hardware-native bitwise MUX operations (torch.where / jax.lax.select) to permanently eliminate 'jmp' instructions and host-device synchronization fences.

A deep learning framework built to be read: lazy tensors, reverse-mode autograd, kernel fusion, a printable IR, conv layers, dtypes from int8 to float64, and JIT backends for C and CUDA.