opt/kernels.json

{
 "kernels": "onnx",
 "gpu": "NVIDIA A100-SXM4-80GB",
 "torch": "2.9.0+cu126",
 "triton": "3.5.0",
 "attention": [
  {
   "seq": 2048,
   "d": 128,
   "heads": 8,
   "block_m": 64,
   "max_abs_diff_vs_sdpa": 0.80859375,
   "sdpa_ms": 0.149,
   "sdpa_tflops": 115.6,
   "triton_ms": 0.434,
   "triton_tflops": 39.6,
   "triton_vs_sdpa_x": 0.34,
   "compiled_ms": 0.187
  },
  {
   "seq": 8192,
   "d": 128,
   "heads": 8,
   "block_m": 128,
   "max_abs_diff_vs_sdpa": 0.578125,
   "sdpa_ms": 1.429,
   "sdpa_tflops": 192.3,
   "triton_ms": 2.647,
   "triton_tflops": 103.8,
   "triton_vs_sdpa_x": 0.54,
   "compiled_ms": 1.524
  },
  {
   "seq": 32768,
   "d": 128,
   "heads": 8,
   "block_m": 128,
   "max_abs_diff_vs_sdpa": 0.36328125,
   "sdpa_ms": 22.129,
   "sdpa_tflops": 198.7,
   "triton_ms": 39.998,
   "triton_tflops": 110.0,
   "triton_vs_sdpa_x": 0.55,
   "compiled_ms": 22.316
  }
 ],
 "onnx_error": "UnsupportedOperatorError(\"Exporting the operator 'aten::_native_multi_head_attention' to ONNX opset version 17 is not supported\")"
}