We added a custom config, Float8DynamicActivationFloat8SemiSparseWeightConfig, for fp8 + 2:4 sparsity in the past, but following our migration off of AQT (#3361) we now plan to express sparsity as a packing format.
Instead, this functionality will be exposed via packing format instead:
Before:
Float8DynamicActivationFloat8SemiSparseWeightConfig()
After:
Float8DynamicActivationFloat8WeightConfig(
packing_format=Float8PackingFormat.SPARSE_CUTLASS,
granularity=PerRow(),
),