cute-dsl

100 Days Of GPU Programming.

CUDA/Triton/CuTe DSL/CUTLASS kernel optimization agent

Kernel programming environment

Agent Skill for in-kernel profiling of CuTe DSL kernels with IKET and run-iket, summarising NVIDIA's CUTLASS docs.

Eight GPU-kernel projects for modern LLM serving

IKET in-kernel event tracing on flashinfer CuTe-DSL kernels, 8x B300 (SM103)

CUDA, Triton and CuTe DSL kernels for transformer inference: fused RMSNorm, decode GEMV, RoPE/KV append and W4A16, with benchmarks and profiling.

Triton and CuTe GPU kernels for variable-width mixture-of-experts inference and training.

CuTe MoE Communication and Overlap Kernels

Hand-optimised GPU kernels with performance journals, benchmarks, and reference implementations

My solutions for the GPUMODE NVFP4 competition, written in CuTe DSL

Drop-in performance optimization layer for RF-DETR using custom fused CuteDSL & Triton kernels: same API, significantly faster execution.