Errors

All custom exceptions in this repository inherit from TritonKernelError.

Hierarchy 

TritonKernelError
├── ShapeMismatchError
├── UnsupportedDtypeError
├── NumericalOverflowError
├── TuningFailedError
└── DeviceError

ShapeMismatchError

Attached attributes may include:

  • expected
  • actual
  • tensor_name

Typical use: incompatible tensor dimensions for kernel inputs.

UnsupportedDtypeError

Attached attributes may include:

  • dtype
  • supported_dtypes
  • tensor_name

Typical use: integer tensor passed into a floating-only kernel path.

NumericalOverflowError

Attached attributes may include:

  • max_value
  • scale
  • attempts

Typical use: FP8 quantization still overflows after repeated scale reduction.

TuningFailedError

Attached attributes may include:

  • problem_size
  • configs_tried
  • last_error

Typical use: every candidate configuration fails during autotuning.

DeviceError

Attached attributes may include:

  • expected_device
  • actual_device
  • tensor_name

Typical use: calling a Triton kernel path with CPU tensors or mixed-device inputs.

Catching patterns 

from triton_ops import fused_rmsnorm_rope
from triton_ops import DeviceError, ShapeMismatchError, UnsupportedDtypeError

try:
    y = fused_rmsnorm_rope(x, weight, cos, sin)
except DeviceError as exc:
    print(exc.expected_device, exc.actual_device)
except ShapeMismatchError as exc:
    print(exc.tensor_name, exc.expected, exc.actual)
except UnsupportedDtypeError as exc:
    print(exc.tensor_name, exc.dtype)