linear_attention

Modules

modelopt.torch.quantization.linear_attention.config	Saved execution policy for GDN training-time numerical emulation.
modelopt.torch.quantization.linear_attention.decode	Explicit token-state and encoded-update replay references for QAT.
modelopt.torch.quantization.linear_attention.decode_prefill	Differentiable explicit per-sequence prefill/decode phase handoff.
modelopt.torch.quantization.linear_attention.kda	Differentiable KDA prefill with stable per-channel decay interactions.
modelopt.torch.quantization.linear_attention.matmul	Operand QDQ and explicit arithmetic schedules for linear-attention matmuls.
modelopt.torch.quantization.linear_attention.prefill	Batched differentiable GDN prefill with materialized numerical boundaries.
modelopt.torch.quantization.linear_attention.reference	Small differentiable oracles, independent of FLA and Triton.
modelopt.torch.quantization.linear_attention.solve	Explicit triangular inverse policies for differentiable prefill emulation.
modelopt.torch.quantization.linear_attention.validation	Capabilities of the initial fused GDN fake-quant path.

Numerical policies and differentiable references for linear attention.