linear_attention

Modules

modelopt.torch.quantization.linear_attention.config	Saved execution policy for GDN/KDA state QAT.
modelopt.torch.quantization.linear_attention.decode	Differentiable per-token state quantization for a training suffix.
modelopt.torch.quantization.linear_attention.gdn	GDN adapter for serving-aligned recurrent-state QAT.
modelopt.torch.quantization.linear_attention.kda	KDA adapter for serving-aligned recurrent-state QAT.
modelopt.torch.quantization.linear_attention.replay	Differentiable checkpoint and update replay for the optional INT8/Hadamard fork.
modelopt.torch.quantization.linear_attention.training	Training forwards with a chunked prefill prefix and recurrent decode suffix.
modelopt.torch.quantization.linear_attention.utils	Shared helpers for linear-attention quantization.

Numerical policies and differentiable kernels for linear attention.