vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

main : option to disable context shift (#9484)

Browse source 
* added cli arg to disable context shift

* reverted precommit

* updated README.md for main

* white space

* allow disabling context shift in the server

* Update common/arg.cpp

no-context-shift only works for main example

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* added server example to --no-context-shift args

* removed server changes

* white space

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Vinesh Janarthanan 2024-09-16 01:20:01 -05:00 • committed by GitHub
parent c4965a64f7
commit 441b72b91f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 36 additions and 21 deletions
Download patch file Download diff file Expand all files Collapse all files

46
examples/main/main.cpp
View file

@ -559,29 +559,35 @@ int main(int argc, char ** argv) {
// if we run out of context:
// - take the n_keep first tokens from the original prompt (via n_past)
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() >= n_ctx) {
if (params.n_predict == -2) {
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
if (!params.ctx_shift){
LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
break;
} else {
if (params.n_predict == -2) {
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard;
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
LOG_DBG("clear session path\n");
path_session.clear();
}
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard;
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
LOG_DBG("clear session path\n");
path_session.clear();
}
} else {
// context extension via Self-Extend