frank-medium: checkpoints back to the build's default (0)
Step 2 of the operator's six. --ctx-checkpoints 8 held per-slot KV snapshots; 0 is what llama.cpp ships, and the fleet's clients open conversations fresh.
This commit is contained in:
1 parent
b6d4914fcc
commit
169eba0734
1 file changed
+4
-1
@@ -129,7 +129,10 @@ THREAD_BATCH=8
|
|||||||
# f16 K/V: an agentic checkpoint is the wrong place to spend KV precision.
|
# f16 K/V: an agentic checkpoint is the wrong place to spend KV precision.
|
||||||
CACHE_TYPE_K=f16
|
CACHE_TYPE_K=f16
|
||||||
CACHE_TYPE_V=f16
|
CACHE_TYPE_V=f16
|
||||||
CTX_CHECKPOINTS=8
|
# Context checkpoints: 0 is the build's default -- no per-slot KV snapshots, so
|
||||||
|
# a re-sent or rewound conversation re-prefills instead of rewinding. Stated
|
||||||
|
# explicitly so the tunable stays visible rather than merely absent.
|
||||||
|
CTX_CHECKPOINTS=0
|
||||||
CACHE_RAM=8192
|
CACHE_RAM=8192
|
||||||
|
|
||||||
# Sampling defaults for requests that send none -- the values the vLLM
|
# Sampling defaults for requests that send none -- the values the vLLM
|
||||||
|
|||||||
Reference in new issue
Block a user