A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
Retrieve `sliding_window` from text config in Gemma3 MM (#25085)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Sep 18, 2025 at 07:29 UTC
f4cd80f94404787859ba72dcddb5e818d8f0c9e7
2 files changed
+2
-4
vllm/model_executor/models/gemma3_mm.py
+2
-1
@@ -688,7 +688,8 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
688
global_attn_mask = torch.where(img_mask == 2, 0, global_attn_mask)
689
global_attn_masks.append(global_attn_mask)
690
691
- if (sliding_window := self.config.sliding_window) is not None:
691
+ sliding_window = self.config.text_config.sliding_window
692
+ if sliding_window is not None:
693
# Create a local causal mask with sliding window (1024).
694
local_attn_mask = torch.ones_like(global_attn_mask)
695
local_attn_mask = torch.tril(local_attn_mask,
vllm/model_executor/models/gemma3n_mm.py
-3
@@ -461,9 +461,6 @@ class Gemma3nForConditionalGeneration(nn.Module, SupportsMultiModal,
461
self.multimodal_config = multimodal_config
462
self.vocab_size = config.text_config.vocab_size
463
464
- self.sliding_window = getattr(config.text_config,
465
- "interleaved_sliding_window", None)
466
-
464
self.vision_tower = AutoModel.from_config(config=config.vision_config)
465
self.audio_tower = AutoModel.from_config(config=config.audio_config)
466
self.embed_vision = Gemma3nMultimodalEmbedder(config.vision_config,