forward

Native vLLM 0.30 Triton forwards; training owns private, differentiable state.

Functions

prefill	Run native chunk prefill, including the selected profile's raw-gate preparation.
step	Run native fused decode on a private cache, preserving the incoming autograd state.
validate_profile	Validate the selected native serving contract before launching prefill.

prefill(q, k, v, g, beta, state, scale, normalize=False, gate_inputs=None, *, precision='vllm')

Run native chunk prefill, including the selected profile’s raw-gate preparation.

step(q, k, v, g, beta, state, scale, normalize=False, gate_inputs=None, *, precision='vllm')

Run native fused decode on a private cache, preserving the incoming autograd state.

validate_profile(gate_inputs, normalize, *, precision='vllm')

Validate the selected native serving contract before launching prefill.