replay#
Training adapter for the optional INT8/Hadamard quantized-ReplaySSM fork.
Functions
Return the serving checkpoint's decoded Hadamard-basis values and FP16 scales. |
|
Use serving's FP32 checkpoint transform when returning a dense state. |
|
Launch one token on private cache buffers; never mutate tensors saved for backward. |
- checkpoint(value, enabled)#
Return the serving checkpoint’s decoded Hadamard-basis values and FP16 scales.
- original_basis(value)#
Use serving’s FP32 checkpoint transform when returning a dense state.
- step(q, k, v, gate, beta, carry, window, enabled, scale, normalize, gate_inputs)#
Launch one token on private cache buffers; never mutate tensors saved for backward.