happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Switch to base cache for position tracking 

Updated position tracking to use base cache instead of SWA cache to avoid kv cache token losses.
This commit is contained in:
John 2026-01-31 05:15:49 +01:00 committed by GitHub
parent 3dd95914d0
commit 2aef9767d9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/llama-kv-cache-iswa.cpp
View File

@ -107,12 +107,14 @@ void llama_kv_cache_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p
}
llama_pos llama_kv_cache_iswa::seq_pos_min(llama_seq_id seq_id) const {
// the base cache is a superset of the SWA cache, so we can just check the SWA cache
return kv_swa->seq_pos_min(seq_id);
// base cache preserves all positions (no SWA cell reuse), making it the ground truth for position tracking
// SWA cache can lose older positions when cells are reused for new tokens outside the SWA window
return kv_base->seq_pos_min(seq_id);
}
llama_pos llama_kv_cache_iswa::seq_pos_max(llama_seq_id seq_id) const {
return kv_swa->seq_pos_max(seq_id);
// base cache preserves all positions (no SWA cell reuse), making it the ground truth for position tracking
return kv_base->seq_pos_max(seq_id);
}
std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_iswa::memory_breakdown() const {