cublas

NumPy & SciPy for GPU

Python interface to GPU-powered libraries

CUDA matrix multiplication benchmarking on Jetson Orin Nano. Four implementations, three power modes, five matrix sizes. 99.5% mathematical validation. C++/CUDA and Python.

Tools to run and parse MKL verbose mode

Classification of Hyperspectral Images ( HSIs ) with Principal Component Analysis ( PCA ) in CUDA ( cuBLAS ).

Enhanced GPU throttle diagnostic for DGX Spark (GB10): NVML direct telemetry, throttle cause decoder, PCIe link monitoring, baseline drift detection, timeline capture.

bilibili视频【CUDA 12.x 并行编程入门(Python版)】配套代码

Heterogeneous linear algebra for Python: AVX2 CPU kernels and an optional cuBLAS CUDA backend behind one zero-copy, NumPy-compatible API, with per-machine CPU/GPU routing

A minimal transformer inference engine built from scratch in CUDA C++, targeting GPT-2 scale models. No high-level ML frameworks — just raw CUDA kernels, explicit memory management, and a clean Python API via pybind11.

High-performance CUDA implementation of Muon optimizer for LLM training. Features Newton-Schulz polar decomposition, cuBLAS acceleration, and transpose optimization for 8x FLOP savings on transformer FFN layers. Benchmarked on NVIDIA A100 with Llama 3.1 8B architectures (4096×11008 weights).

BLAS Levels 1–3 for AWS Trainium via NKI (cuBLAS-equivalent) — GEMM with stationary-tile reuse, batched GEMM, TRSM, validated DF-MP2 for quantum chemistry.

NumPy-like API accelerated with CUDA

Mahalanobis scoring stage of PaDiM anomaly detection, implemented four ways - PyTorch, C++17 CPU, hand-written CUDA (naive + PBCR), cuBLAS - behind a fixed artifact boundary, with a 1e-4 numerical agreement gate on every reported timing.

GPU profiling suite & CUDA kernels on A100 80GB — ResNet-50 benchmarks, Nsight Systems profiling, tiled matrix multiplication with shared memory

NVIDIA GPU validation: PCIe transport, Unified Memory prefetch, SGEMM compute, drift detection.