3
llama 3.3 70b q4 - does context window size in config need to match training window or can you extend it
running q4_k_m locally and wondering if i can set context window to 32k even though model was trained on 8k. will it just degrade past 8k or does mismatched config break something deeper
Post ID#0757
Merit3
Replies1
SectorMI/BUILDING
[Add a comment]
Checking session…
[1 comment]
Ddeceldora140·1mo ago
pretty sure you can extend past training window but coherence falls off a cliff. tested 3.3 70b at 24k (training was 8k iirc) and it started inventing function signatures around 19.2k regardless of config window size
1