coreai_opt.quantization.spec.resolve_block_sizes

coreai_opt.quantization.spec.resolve_block_sizes(tensor_shape, block_sizes)[source]

Resolve a per-axis block partition against the shape it applies to.

-1 takes the whole axis as one block, which is how per-channel and per-tensor are spelled as a partition. Every other entry blocks its axis and must divide it.

Parameters:

  • tensor_shape (Sequence[int]) – Shape of the tensor being partitioned.

  • block_sizes (Sequence[int]) – Block extent per axis, one entry per dimension.

Returns:

The resolved extents, with each -1 replaced by its full dimension.

Return type:

list[int]

Raises:

  • ValueError – If the two lengths differ, or an entry is neither positive nor -1.

  • _BlockSizeMismatchError – If an entry does not divide its dimension.