[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
@@ -0,0 +1,28 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from vllm.model_executor.models.phi4mm_utils import MultiHeadedAttention
def test_eager_attention_does_not_retain_probability_matrix():
module = MultiHeadedAttention(
n_head=2,
n_feat=16,
dropout_rate=0.0,
use_pt_scaled_dot_product_attention=False,
)
module.eval()
batch, time, feat = 2, 8, 16
x = torch.randn(batch, time, feat)
with torch.no_grad():
out = module(x, x, x, None, None, None)
assert out.shape == (batch, time, feat)
assert getattr(module, "attn", None) is None
def test_multiheaded_attention_defaults_to_scaled_dot_product():
module = MultiHeadedAttention(n_head=2, n_feat=16, dropout_rate=0.0)
assert module.use_pt_scaled_dot_product_attention is True