fhelium.backend.triton._tensor
Generate ordinary floating-point elementwise and cyclic-indexing kernels.
TritonTensorFusionImplementation
class View source
python
TritonTensorFusionImplementation(name: str = 'triton-tensor-fused', operation_types: tuple[type[Operation], ...] = (fusion.FusedOp,), supports_in_place: bool = False)1
Fuse supported ordinary Tensor expressions without introducing CKKS.
Automatic matching requires known CUDA layouts. Unknown placement remains unassigned; CPU execution and unsupported operations retain ordinary Torch implementations. Each output has separate storage, and floating-point contraction is disabled to retain the source operation rounding sequence.
Attributes
| Name | Type | Default/value |
|---|---|---|
name | str | 'triton-tensor-fused' |
operation_types | tuple[type[Operation], ...] | (fusion.FusedOp,) |
supports_in_place | bool | False |
match_fusion
method
python
def match_fusion(operations: Sequence[Operation]) -> int | None: ...1
resource_requirements
method
python
def resource_requirements(invocation: OperationInvocation) -> tuple[ResourceRequirement, ...]: ...1
prepare_operation
method
python
def prepare_operation(operation: Operation): ...1
execute
method
python
def execute(invocation, inputs, resources, *, in_place): ...1