[Model] Support EmbeddingGemma2 multimodal pooling architecture (#60254)

Signed-off-by: Luciano Martins <lucianommartins@users.noreply.github.com>
Co-authored-by: Luciano Martins <lucianommartins@users.noreply.github.com>
This commit is contained in:
Luciano Martins
2026-10-06 13:43:38 -04:00
committed by GitHub
co-authored by Luciano Martins
parent 7b665f7a58
commit 02b83919aa
7 changed files with 2122 additions and 2 deletions
@@ -243,9 +243,10 @@ def context_attention_fwd(
b_seq_len: [b]
out: [b * s, head, head_dim]
"""
BLOCK = get_block_size(q.dtype)
Lq, Lk, _ = q.shape[-1], k.shape[-1], v.shape[-1]
BLOCK = get_block_size(q.dtype)
if Lk >= 512:
BLOCK = min(BLOCK, 32)
sm_scale = 1.0 / (Lq**0.5) if softmax_scale is None else softmax_scale
# rescale with 1/ln(2) for triton exp2