Allow capturing only necessary hidden_states with capture_outputs (#48081)
* [feat] Add selective hidden-state capture via output_hidden_states_layers.
This PR extends the generic output-capturing hook so callers can request
only specific hidden-state layers: outputs = model(..., output_hidden_states_layers=[1, 3])
When this argument is provided, hidden-state capture is enabled automatically, but tensors are
only retained for the requested layer ids. Non-requested positions are kept as None placeholders
so the existing indexing convention remains unchanged
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* update
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* add dflash case
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* format
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fmt
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove unnecessary test
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* use a single kwarg
* add general test
---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co>