decode#

Fused FP32 token/replay training recurrence with state checkpoint recomputation.

Functions

fused_recurrence

Run a nonempty activated sequence; encoding keys/gates and carry metadata is external.

fused_recurrence(q, k, v, g, beta, initial, *, state_qdq=False, state_format='fp8_e4m3', state_codec='tile', block_v=16, replay=False, factor_qdq=False, window=8, cursor=0, read_stored=True, scale=None, checkpoint_interval=8)#

Run a nonempty activated sequence; encoding keys/gates and carry metadata is external.