Skip to content

llama : fix defrag logic #11707

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Feb 7, 2025
Merged

llama : fix defrag logic #11707

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
04c01e9
llama : fix defrag logic
ggerganov Feb 6, 2025
32b8ce5
cont : better logic
ggerganov Feb 6, 2025
861d3b9
cont : clamp fragmentation to 0.0
ggerganov Feb 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8801,12 +8801,14 @@ 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) {
const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f;
if (cparams.causal_attn && cparams.defrag_thold > 0.0f) {
// - do not defrag small contexts (i.e. < 2048 tokens)
Copy link
Contributor

@MoonRide303 MoonRide303 Feb 8, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@ggerganov I am sometimes running benchmarks that require only 256 or 512 tokens per slot, with total context size like 512 or 1024 (for big models that don't fully fit into my VRAM). Will it work properly in cases like that?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The defragmentation for such small context is not really worth it, so my expectation is that with this change you should get better performance overall.

// - count the padding towards the number of used tokens
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) {
//LLAMA_LOG_INFO("fragmentation: %.2f\n", fragmentation);
LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);

llama_kv_cache_defrag(kv_self);
}
Expand Down
Loading