### Details: - Allows to avoid kv cache rearrange in cases when beam search is used. Instead of that beam table is maintained to track the history of required cache modifications which is passed to subsequent gemm primitive - Beam table is a part of the model state, so GPU `VariableState` was extended to multi-tensor case - Improves 2nd+ token latency and reduces memory consumption with beam size > 1 ### Tickets: - *124119* --------- Co-authored-by: Sergey Shlyapnikov <sergey.shlyapnikov@intel.com> |
||
|---|---|---|
| .. | ||
| auto | ||
| auto_batch | ||
| hetero | ||
| intel_cpu | ||
| intel_gpu | ||
| proxy | ||
| template | ||
| CMakeLists.txt | ||
| README.md | ||
README.md
OpenVINO Plugins
OpenVINO Plugins provide support for hardware devices.
The list of supported plugins: