core.ssm.ops.gdp

Inference-only fork of the Gated Delta Product kernels from flash-linear-attention <https://github.com/fla-org/flash-linear-attention>_ (v0.5.1).

Training and the static-batching inference path call the pip flash-linear-attention kernels, which own the backward pass. Only the dynamic-batching decode and prefill steps route here:

• fused_recurrent_gated_delta_rule_update – decode.

• gdp_decode_prepare – the reshape/gating step feeding decode.

• chunk_gated_delta_product_varlen – prefill.

All entry points are forward-only.

Submodules

• core.ssm.ops.gdp.metadata
• core.ssm.ops.gdp.chunk_o
• core.ssm.ops.gdp.fused_recurrent
• core.ssm.ops.gdp.common
• core.ssm.ops.gdp.decode_prepare
• core.ssm.ops.gdp.wy_fast
• core.ssm.ops.gdp.solve_tril
• core.ssm.ops.gdp.chunk_h
• core.ssm.ops.gdp.chunk
• core.ssm.ops.gdp.scaled_dot_kkt
• core.ssm.ops.gdp.cumsum

Package Contents

Data

__all__

API

core.ssm.ops.gdp.__all__

[‘chunk_gated_delta_product_varlen’, ‘fused_recurrent_gated_delta_rule_update’, ‘gdp_decode_prepare’…