serving

Modules

modelopt.torch.kernels.quantization.linear_attention.serving.forward	Native vLLM 0.30 Triton forwards; training owns private, differentiable state.
modelopt.torch.kernels.quantization.linear_attention.serving.replay	Functional training launches of the optional quantized-ReplaySSM serving kernels.

Optional forward kernels for linear-attention serving precision profiles.

The concrete modules import vLLM lazily. This package must remain importable in the Megatron training test lane, where adapter integration uses an independent recurrence oracle and vLLM is deliberately not installed.