TileSight

TileSight models GPU kernels from tile shapes, data movement, software pipelines and hardware profiles. It estimates operation costs, cache traffic and execution time on the CPU. CUDA and HIP probes measure the target GPU separately.

Install

Python 3.10+ is required. Modeling examples run on the CPU.

git clone https://github.com/tile-ai/TileSight.git
cd TileSight
python -m pip install -e .

Model a kernel

import tilesight as sight
from tilesight.arch.h200_sxm import H200_SXM
from tilesight.modeling.program import examples as ex

program = ex.gemm_program(m=1024, n=1024, k=1024, dtype="bf16")
result = sight.analyze(
    program, H200_SXM().set_to_microbench(),
    options=sight.Options(cache="fast", ii_mode="periodic_best"),
)
print("Predicted kernel body (s):", result.launches["main"].kernel_body_s)

Build your own programs with KernelBuilder. The example builders and result types document shapes, schedules, traffic and timing. Profiles contain measured, derived or inherited parameters.

python examples/periodic_schedule_basics.py
python examples/fa3_fa4_schedule_modes.py
python examples/flashmla_decode.py --num-splits 4

The GPU measurement guide covers NVIDIA SM and AMD gfx selection, DRAM concurrency scans and optional communication tests.

Layout and tests

Path	Purpose
tilesight/arch/	GPU architecture profiles
tilesight/modeling/	Program frontend, cost/cache models and scheduling
tilesight/util/	Shared utilities
examples/	Runnable modeling examples
micro_benchmark/	CUDA/HIP probes and communication adapters

python -m pip install -e ".[dev]"
python -m pytest -q

Generated measurements and binaries remain local. MIT license.