root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-10 10:41:47 +00:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

cont : no need for special "greedy" logic

Browse Source 
top-k == 1 is the same
This commit is contained in:
Georgi Gerganov 2024-10-17 18:09:57 +03:00
parent cb75bebcad
commit 57fb835e5b
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 37 additions and 50 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
common/sampling.cpp
View File

@ -171,7 +171,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
params.penalize_nl, params.penalize_nl,
params.ignore_eos)); params.ignore_eos));
if (params.temp >= 0.0f) {
if (params.mirostat == 0) { if (params.mirostat == 0) {
for (const auto & cnstr : params.samplers) { for (const auto & cnstr : params.samplers) {
switch (cnstr) { switch (cnstr) {
@ -213,18 +212,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
} else { } else {
GGML_ASSERT(false && "unknown mirostat version"); GGML_ASSERT(false && "unknown mirostat version");
} }
} else {
// negative temperatures will trigger "greedy" sampling: simply take the most likely token each time
if (params.n_probs > 0) {
// some use cases require to sample greedily, but still obtain the probabilities of the top tokens
// ref: https://github.com/ggerganov/llama.cpp/pull/9605
//
// the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
// it is much faster, since we avoid sorting all tokens and should give a good approximation
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
}
llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
}
return result; return result;
} }