linear_attention

Modules

modelopt.torch.kernels.quantization.linear_attention.replayssm	GDN replay forward corresponding to the ModelOpt two-precision window adjoint.
modelopt.torch.kernels.quantization.linear_attention.replayssm_cache	Indexed two-precision GDN decode using externally owned cache pages.
modelopt.torch.kernels.quantization.linear_attention.replayssm_read	Per-request, per-head NVFP4 read refresh with ModelOpt's E4M3/E2M1 rounding.
modelopt.torch.kernels.quantization.linear_attention.serving	Optional forward kernels for linear-attention serving precision profiles.

Linear-attention quantization adapters for optional serving kernels.

Serving adapters import forward kernels only when selected; this package initializer has no kernel dependencies. The existing FLA chunk kernels remain available for direct calls with fla-core==0.5.2.