Maintainer execution helpers¶
These modules are documented for maintainers because they sit near dispatch and quantized execution helpers. They are not the preferred public import surface for user code.
Relation execution¶
- mlx_lattice.ops._relation_exec.sparse_quantized_conv_features_from_relation(feats, weight, relation)[source]¶
Execute quantized sparse convolution over a prebuilt relation.
- Return type:
array- Parameters:
feats (array)
weight (QuantizedWeight)
relation (KernelRelation)
- mlx_lattice.ops._relation_exec.sparse_conv_features_from_relation(feats, weight, relation)[source]¶
Execute sparse convolution over a prebuilt relation.
- Return type:
array- Parameters:
feats (array)
weight (array)
relation (KernelRelation)
- mlx_lattice.ops._relation_exec.sparse_conv_features_sorted_from_relation(feats, weight, relation, *, store_sorted=False)[source]¶
Execute the sorted implicit-GEMM convolution path explicitly.
- Return type:
array- Parameters:
feats (array)
weight (array)
relation (KernelRelation)
store_sorted (bool)
- mlx_lattice.ops._relation_exec.sparse_conv_features_sorted_direct_reference_from_relation(feats, weight, relation, *, store_sorted=False)[source]¶
Execute the sorted direct reference path for diagnostics/tests.
- Return type:
array- Parameters:
feats (array)
weight (array)
relation (KernelRelation)
store_sorted (bool)
- mlx_lattice.ops._relation_exec.sparse_pool_features_from_relation(feats, relation, *, input_exclusive, mode)[source]¶
Execute sparse pooling over a prebuilt kernel relation.
- Return type:
array- Parameters:
feats (array)
relation (KernelRelation)
input_exclusive (bool)
mode (str)
Quantized execution¶
- mlx_lattice.ops._quantized.quantized_matmul(feats, weight)[source]¶
Apply affine packed-weight matrix multiplication to feature rows.
- Return type:
array- Parameters:
feats (array)
weight (QuantizedWeight)