[Model] Gemma4: add bidirectional vision attention for sliding layers with window guard (#40534)

Signed-off-by: Luciano Martins <lucianommartins@users.noreply.github.com>
Signed-off-by: Luciano Martins <lucianomartins@google.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Luciano Martins <lucianommartins@users.noreply.github.com>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Luciano Martins
2026-04-24 08:27:46 +00:00
committed by GitHub
co-authored by Luciano Martins Isotr0py Isotr0py
parent 4c34b2f6fc
commit 512f522192
2 changed files with 73 additions and 1 deletions
+14 -1
View File
@@ -2314,13 +2314,26 @@ class GPUModelRunner(
if self.is_mm_prefix_lm:
req_doc_ranges = {}
# Gemma4 bidi: skip ranges that exceed the sliding
# window. When image tokens > sliding_window, bidi causes
# early image tokens to attend to the entire image
# (e.g. 6 → 1092 targets), degrading spatial precision.
# Per-range filtering keeps bidi for small images/video
# frames while skipping oversized images.
hf_text_config = self.model_config.hf_text_config
_bidi_sw = getattr(hf_text_config, "sliding_window", None)
for req_id in self.input_batch.req_ids:
image_doc_ranges = []
req_state = self.requests[req_id]
for mm_feature in req_state.mm_features:
pos_info = mm_feature.mm_position
img_doc_range = pos_info.extract_embeds_range()
image_doc_ranges.extend(img_doc_range)
for r in img_doc_range:
if _bidi_sw is not None and (r[1] - r[0] + 1) > _bidi_sw:
continue
image_doc_ranges.append(r)
req_idx = self.input_batch.req_id_to_index[req_id]
req_doc_ranges[req_idx] = image_doc_ranges