|
liblloyal 1.0.0
Branched Inference for llama.cpp
|
Options controlling what state fork() clones from source to child. More...
#include <lloyal/branch.hpp>
Public Attributes | |
| bool | clone_logits = true |
| If true (default), copy src->logits_snapshot to child (~n_vocab*4 bytes, ~600KB for 150k-vocab models). | |
Options controlling what state fork() clones from source to child.
Default {} preserves fork's full "sample same distribution" contract. Set clone_logits=false for prefill-overwrite consumers (rerank leaves, embedding probes) that fork → prefill → read → prune and never sample the parent's distribution.
Definition at line 1352 of file branch.hpp.
| bool lloyal::branch::ForkOpts::clone_logits = true |
If true (default), copy src->logits_snapshot to child (~n_vocab*4 bytes, ~600KB for 150k-vocab models).
If false, leave child has_logits=false until subsequent prefill()/step(); calling sample() on the child before that returns -1 (the kernel's no-logits sentinel; the SDK Branch.sample() converts this to BranchSampleError).
Definition at line 1358 of file branch.hpp.