added AD perf test

Signed-off-by: Eran Geva <[email protected]>

added memory checks

Signed-off-by: Eran Geva <[email protected]>

changed kv cache test to be hw agnostic

Signed-off-by: Eran Geva <[email protected]>

added test with backend comparison

Signed-off-by: Eran Geva <[email protected]>

both tests pass

Signed-off-by: Eran Geva <[email protected]>

cleanups and refactoring

Signed-off-by: Eran Geva <[email protected]>

fixed llm_root issue

Signed-off-by: Eran Geva <[email protected]>

shrunk the model, and fixes

Signed-off-by: Eran Geva <[email protected]>

fixed trtllm-bench test stability

Signed-off-by: Eran Geva <[email protected]>

preserved the old test

Signed-off-by: Eran Geva <[email protected]>

set mem ratio to 0.3, fixed bug in default params

Signed-off-by: Eran Geva <[email protected]>

Author: MrGeva