[None][feat] Enable KVCacheManagerV2 by default for Gemma3 and Gemma4 - #17396

Merged

nvpohanh merged 8 commits into

NVIDIA:mainfrom

erictsai-nv:feat/gemma-v2cpp-default

Aug 18, 2026

Commits on Aug 14, 2026

[None][feat] Enable KVCacheManagerV2 by default for Gemma3 and Gemma4
erictsai-nv
committed
b688572
Add use_kv_cache_manager_v2=True to Gemma4ForConditionalGeneration.get_model_defaults
erictsai-nv
committed
4b83678
[None][fix] Address review: cover Gemma3 VLM, MM transceiver hook, resolution-path tests
erictsai-nv
committed
aae7372
[None][doc] Add Gemma row to the KV cache manager selection table
erictsai-nv
committed
17988fa
[None][fix] Migrate Gemma V2 preference to get_preferred_kv_cache_manager_version
erictsai-nv
committed
8c221e1
[None][fix] Prefer the Python transceiver for DeepSeek-V4 and MiniMax-M3; add NIXL V2-retention test
erictsai-nv
committed
e11ade7
[None][fix] Drop duplicate Gemma3 transceiver hook (added upstream in #16787)
erictsai-nv
committed
f9b4a48
[None][test] Cover GLM 5.2 in the NIXL V2-retention test (#17283 made it prefer the Python transceiver)
erictsai-nv
committed
a5c6e94