FlyDSL Examples

Unofficial GEMM kernel examples built with FlyDSL for AMD GPUs.

The goal of this repository is to show how to write high-performance GEMM kernels from scratch in Python, while keeping the implementation close enough to the hardware to reason about tiling, memory movement, and WMMA/MFMA execution. The programming style is similar in spirit to CUDA/CuteDSL, but targets AMD GPUs through FlyDSL.

Kernels

• FP16/BF16 GEMM WMMA for MI355
• FP8 PTPC GEMM WMMA for MI355

Results

Run Tests And Benchmarks

rm -rf ~/.flydsl/ ; pytest -sv test_hgemm.py

Use code agent to scale the kernel to your hardware

python agent/agent.py --input=agent/instructions/flydsl_gemm_mi308_support.txt

References

• FlyDSL
• MLIR docs
• ROCm blog: Accelerating LLM inference on AMD GPUs with low-latency GEMMs

Contact: [email protected]