Tin mới
GitHub CopilotWrite better code with AI
GitHub Copilot appDirect agents from issue to merge
CodespacesInstant dev environments
Code ReviewManage code changes
Code QualityEnforce quality at merge
GitHub Advanced SecurityFind and fix vulnerabilities
Code securitySecure your code as you build
Secret protectionStop leaks before they start
GitHub SponsorsFund open source developers
Copilot for BusinessEnterprise-grade AI features
Premium SupportEnterprise-grade 24/7 support
The server now makes checkpoints of the SWA memory in order to minimize the amount of context reprocessing. A SWA checkpoint represents the state (both the KV cells and KV data) of the cache. Only the SWA part is stored
llama.cpp/tools/server/server.cpp
Learn more about hiding disruptive comments
The reason will be displayed to describe this comment to others. Learn more.
server : handle state restore fails
llama : add extended llama_state_seq_ API
server : do not make checkpoints if --swa-full
llama : remove flags value for NONE
server : configure number of SWA checkpoints with CLI arg
args : fix scope of new argument
changelog : libllama API #9289
Add GPT-OSS from OpenAI - closed in favor of 689 ikawrakow/ik_llama.cpp#683
Eval bug: Nemotron v2 Nano always reprocesses prompt #15677
implement context checkpointing for hybrid and recurrent models #16382
changelog : libllama API COG-GTM/llama.cpp#246
server : add SWA checkpoints (#15293)
Qwen3.5 waiting 3 seconds after generating. JamePeng/llama-cpp-python#77
server: fix multi-turn cache reuse for hybrid/recurrent models #20428
server : add SWA checkpoints (
server : add SWA checkpoints (
server : add SWA checkpoints (
server: merge custom preset aliases into existing HF cached model entries #23952