• Operator fusion with correctness guarantees — every kernel ships with CPU-testable NumPy reference implementations, not just speed claims.
• Production-ready FP8 GEMM pipeline — explicit scale management and overflow handling, not toy quantization examples.
• Latency-driven autotuner with persistent config cache — TritonAutoTuner + ConfigCache, not one-off benchmark scripts.
• OpenSpec-driven development — every non-trivial change is design-documented before code, not YOLO-coded.

User API (triton_ops.__init__)
    ├── Validation Layer (device, dtype, shape, contiguity)
    ├── Compute Reference Layer (NumPy, CPU-testable)
    ├── Kernel Layer (Triton, GPU)
    └── Tooling Layer (autotuner, benchmark, performance metrics)

See the Architecture Lab and Kernel Families docs for details.

git clone https://github.com/LessUp/triton-fused-ops.git
cd triton-fused-ops
pip install -e ".[dev]"

CPU-only validation (no GPU required):

ruff format --check . && ruff check . && mypy triton_ops/
pytest tests/ -v -k "not cuda and not gpu" --ignore=tests/benchmarks/
python3 -m build

Full GPU benchmark (requires CUDA):

import torch
from triton_ops import fused_rmsnorm_rope, BenchmarkSuite
from triton_ops.reference import fused_rmsnorm_rope as fused_rmsnorm_rope_reference

x = torch.randn(2, 2048, 4096, device="cuda", dtype=torch.float16)
suite = BenchmarkSuite(warmup_runs=10, benchmark_runs=100)
result = suite.benchmark_kernel(
    fused_rmsnorm_rope, fused_rmsnorm_rope_reference,
    "fused_rmsnorm_rope", (2, 2048, 4096), x, ...
)
print(result.metrics.latency_ms)

Representative numbers on NVIDIA A100 SXM4 80GB (CUDA 12.1, PyTorch 2.1, Triton 2.1). Methodology: 10 warmup runs + 100 benchmark runs with torch.cuda.synchronize() before and after timing.

See Benchmark Visualization for interactive charts.

This repository is OpenSpec-driven for non-trivial work. See AGENTS.md, CLAUDE.md, and openspec/README.md.

@software{triton_fused_ops,
  title = {Triton Fused Ops: High-Performance GPU Kernels for Transformer Inference},
  author = {LessUp},
  year = {2025},
  url = {https://github.com/LessUp/triton-fused-ops},
  note = {Built on OpenAI Triton, PyTorch, and CUDA}
}

• OpenAI Triton for the compiler and Python DSL
• PyTorch for the tensor runtime
• NVIDIA for CUDA, FP8 hardware, and performance tooling

MIT.