|
liblloyal 1.0.0
Branched Inference for llama.cpp
|
How a context reads what is placed into it: the one read of its traits that placement decides from. More...
#include <lloyal/decode.hpp>
Public Member Functions | |
| bool | whole () const |
| With no memory, attention that is not causal, or pooling, a dispatch sees only its own entries, so a stream placed in pieces would be read in pieces. | |
| bool | takes_bidirectional () const |
| Whether a bidirectional block can ride inside one mixed batch. | |
Public Attributes | |
| bool | keeps_memory = true |
| It keeps what it was given (a KV cache), so a later dispatch attends to an earlier one. | |
| bool | causal = true |
| Its attention is causal. | |
| bool | pooled = false |
| It pools a sequence into one output. | |
How a context reads what is placed into it: the one read of its traits that placement decides from.
Definition at line 927 of file decode.hpp.
|
inline |
Whether a bidirectional block can ride inside one mixed batch.
llama.cpp encodes a context without memory with causal attention off for the whole batch; a context that keeps memory and attends causally cannot, since its causal mode is context-wide.
Definition at line 945 of file decode.hpp.
|
inline |
With no memory, attention that is not causal, or pooling, a dispatch sees only its own entries, so a stream placed in pieces would be read in pieces.
Any other context (a reasoning model's: memory, causal, no pooling) is placed a segment at a time.
Definition at line 939 of file decode.hpp.
| bool lloyal::decode::StreamReading::causal = true |
Its attention is causal.
Definition at line 931 of file decode.hpp.
| bool lloyal::decode::StreamReading::keeps_memory = true |
It keeps what it was given (a KV cache), so a later dispatch attends to an earlier one.
Definition at line 929 of file decode.hpp.
| bool lloyal::decode::StreamReading::pooled = false |
It pools a sequence into one output.
Definition at line 933 of file decode.hpp.