[Security] Do not retain Phi-4 MM audio attention matrices (#60614)

Signed-off-by: Juan Pérez de Algaba <jperezde@redhat.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Juan Pérez de Algaba
2026-10-09 22:23:35 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 81c993658e
commit 3d52578e13
3 changed files with 34 additions and 9 deletions
+4 -4
View File
@@ -129,7 +129,7 @@ class ConformerEncoderLayer(nn.Module):
default False.
use_pt_scaled_dot_product_attention: bool, optional
if set to True, use pytorch's scaled dot product attention
implementation in training.
implementation in training. Default: True.
attn_group_sizes: int, optional
the number of groups to use for attention, default 1
(Multi-Head Attention),
@@ -163,7 +163,7 @@ class ConformerEncoderLayer(nn.Module):
attention_glu_type: str = "swish",
activation_checkpointing: str = "",
export: bool = False,
use_pt_scaled_dot_product_attention: bool = False,
use_pt_scaled_dot_product_attention: bool = True,
attn_group_sizes: int = 1,
) -> None:
super().__init__()
@@ -808,7 +808,7 @@ class ConformerEncoder(TransformerEncoderBase):
default 4
use_pt_scaled_dot_product_attention: whether to use pytorch scaled
dot product attention in training.
Default: False
Default: True
nemo_conv_settings: dict, optional
A dictionary of settings for NeMo Subsampling.
default: None
@@ -873,7 +873,7 @@ class ConformerEncoder(TransformerEncoderBase):
activation_checkpointing: str = "",
relative_attention_bias_args: dict[str, Any] | None = None,
time_reduction: int = 4,
use_pt_scaled_dot_product_attention: bool = False,
use_pt_scaled_dot_product_attention: bool = True,
nemo_conv_settings: dict[str, Any] | None = None,
conv2d_extra_padding: Literal["feat", "feat_time", "none", True] = "none",
replication_pad_for_subsample_embedding: bool = False,