llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k repeats in DeepSeek V4 - #25370

Merged

fairydreaming merged 3 commits into

ggml-org:masterfrom

fairydreaming:dsv4-remove-kq-bias

Jul 10, 2026

Commits on Jul 6, 2026

llama : make all KQ masks (except the lightning indexer one) f16 if FA is used and remove zero attention bias in DeepSeek V4
sszymczy
committed
1725223

Commits on Jul 7, 2026

llama : remove dead code that repeats unified raw_k cache for each stream in DeepSeek V4 - no longer needed as raw_k is always non-unified.
sszymczy
committed
a0e9766
Merge remote-tracking branch 'upstream/master' into dsv4-remove-kq-bias
sszymczy
committed
2b0d980