From ebd859cf6165788492b1f5ee09b39c43261cb0b5 Mon Sep 17 00:00:00 2001 From: Yechan Kim <161688079+yechank-nvidia@users.noreply.github.com> Date: Tue, 3 Feb 2026 18:59:44 +0900 Subject: [PATCH] [https://nvbugs/5854419][fix] Fix Qwen3-VL-Dense/MoE accuracy drop (#11134) Signed-off-by: yechank <161688079+yechank-nvidia@users.noreply.github.com> Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com> --- tensorrt_llm/_torch/models/modeling_qwen3vl.py | 1 + tensorrt_llm/_torch/models/modeling_qwen3vl_moe.py | 1 + 2 files changed, 2 insertions(+) diff --git a/tensorrt_llm/_torch/models/modeling_qwen3vl.py b/tensorrt_llm/_torch/models/modeling_qwen3vl.py index 312e5ea745..8a5087731e 100644 --- a/tensorrt_llm/_torch/models/modeling_qwen3vl.py +++ b/tensorrt_llm/_torch/models/modeling_qwen3vl.py @@ -1138,6 +1138,7 @@ class Qwen3VLModelBase(PreTrainedModel): "video": "<|vision_start|><|video_pad|><|vision_end|>", }, placeholder_placement=MultimodalPlaceholderPlacement.BEFORE_TEXT, + placeholders_separator="", ), ) class Qwen3VLModel(Qwen3VLModelBase): diff --git a/tensorrt_llm/_torch/models/modeling_qwen3vl_moe.py b/tensorrt_llm/_torch/models/modeling_qwen3vl_moe.py index 74cd16ec69..66da3c153e 100644 --- a/tensorrt_llm/_torch/models/modeling_qwen3vl_moe.py +++ b/tensorrt_llm/_torch/models/modeling_qwen3vl_moe.py @@ -41,6 +41,7 @@ from .modeling_utils import ModelConfig, register_auto_model, register_vision_en "video": "<|vision_start|><|video_pad|><|vision_end|>", }, placeholder_placement=MultimodalPlaceholderPlacement.BEFORE_TEXT, + placeholders_separator="", ), ) class Qwen3MoeVLModel(Qwen3VLModelBase):