Develop, optimize, and debug TileLang GPU kernels for AI workloads.
npx skills add https://github.com/yzlnew/infra-skills --skill tilelang-developer
Please help me install this Agent Skill. Skill: tilelang-developer Source: https://github.com/yzlnew/infra-skills/tree/main/tilelang-developer Command: npx skills add https://github.com/yzlnew/infra-skills --skill tilelang-developer
Designs, optimizes, and debugs high-performance TileLang GPU kernels for AI workloads.
Design, implement, and benchmark a minimal TileLang kernel by executing a simple 32x32 GEMM on a CUDA device.
High-intent search queries and answers about installing and using this skill.
To write a custom GPU kernel using TileLang, you scaffold the kernel structure, apply shared memory swizzle layouts, implement 3-stage software pipelining, and ensure explicit FP32 accumulation for high-performance AI workloads.
Yes, you can port CUDA kernels to AMD and Ascend backends using TileLang. It provides cross-vendor portability, allowing the same kernel logic to target NVIDIA, AMD, and Ascend hardware.
The best way to optimize GEMM or Attention operators in TileLang is by utilizing shared memory swizzle layouts, implementing 3-stage software pipelining, and enforcing explicit FP32 accumulation to maximize throughput.
Yes, TileLang supports implementing non-standard operators like DeepSeek MLA. It provides the necessary kernel scaffolding and memory management patterns to develop and debug these custom operators.
To start benchmarking a minimal TileLang kernel on a CUDA device, you design, implement, and execute a simple 32x32 GEMM operation to validate performance and establish a baseline for optimization.
Your TileLang kernel requires explicit FP32 accumulation to maintain numerical precision and prevent overflow during high-throughput matrix operations, which is essential for validating correctness and achieving stable performance.
m0atCross-compile Zig for Linux servers and CUDA/HIP GPU systems.
alibabaAnalyzes GPU kernel performance using Roofline models and profiling data.
jstzwjOptimize large-scale deep learning training with ZeRO and parallelism techniques.
xiechyDetect host CPU, GPU, memory, and disk resources into a .claude_resources.json file.
