From b1dc87a0989fd98a614e4e7e6b318a19e8c5c6f9 Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Fri, 17 Apr 2026 13:37:21 +0100 Subject: [PATCH] [Models][Gemma4] Prevent GPU/CPU sync in `embed_input_ids` (#39234) Signed-off-by: Lukas Geiger --- vllm/model_executor/models/gemma4_mm.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/vllm/model_executor/models/gemma4_mm.py b/vllm/model_executor/models/gemma4_mm.py index cd0ca37616..08978ec3f4 100644 --- a/vllm/model_executor/models/gemma4_mm.py +++ b/vllm/model_executor/models/gemma4_mm.py @@ -1254,9 +1254,10 @@ class Gemma4ForConditionalGeneration( # computation (using token_type_ids == 0 as text_mask). # Replicate this: map image token positions to token 0. if is_multimodal is not None: - is_multimodal = is_multimodal.to(input_ids.device) ple_input_ids = torch.where( - is_multimodal, torch.zeros_like(input_ids), input_ids + is_multimodal.to(input_ids.device, non_blocking=True), + torch.zeros_like(input_ids), + input_ids, ) else: ple_input_ids = input_ids