Length-normalized mean log-prob over the contiguous input-position span
[s0, s0+n) of h [1, T, hidden] (post-finalNorm), scoring targets
ids[s0+1 .. s0+n]. The span-parameterized generalization of
responseOnlyLogpMean (same ops); used for the prompt span under
sft_scope:"full". Whole-vocab autograd head. Returns [1] (caller owns);
disposes nothing it didn't create.
Length-normalized mean log-prob over the contiguous input-position span [s0, s0+n) of
h [1, T, hidden](post-finalNorm), scoring targets ids[s0+1 .. s0+n]. The span-parameterized generalization of responseOnlyLogpMean (same ops); used for the prompt span under sft_scope:"full". Whole-vocab autograd head. Returns [1] (caller owns); disposes nothing it didn't create.