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

cont : clamp fragmentation to 0.0

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2025-02-07 09:50:32 +02:00
parent 32b8ce5b96
commit 861d3b99de
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/llama.cpp
View file

@ -8801,10 +8801,10 @@ static int llama_decode_impl(
//llama_synchronize(&lctx);
// decide if we need to defrag the kv cache
if (cparams.causal_attn && cparams.defrag_thold >= 0.0f) {
if (cparams.causal_attn && cparams.defrag_thold > 0.0f) {
// - do not defrag small contexts (i.e. < 2048 tokens)
// - count the padding towards the number of used tokens
const float fragmentation = kv_self.n >= 2048 ? 1.0f - float(kv_self.used + llama_kv_cache_get_padding(cparams))/float(kv_self.n) : 0.0f;
const float fragmentation = kv_self.n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self.used + llama_kv_cache_get_padding(cparams))/float(kv_self.n)) : 0.0f;
// queue defragmentation for next llama_kv_cache_update
if (fragmentation > cparams.defrag_thold) {