From 0f6ac5e21db0aca05bdb0fce72b344d173cfa8c2 Mon Sep 17 00:00:00 2001 From: Adam Yanxiao Zhao Date: Sat, 6 Sep 2025 05:20:36 +0800 Subject: [PATCH] [Bug Fix] Fix Glm4vVisionBlock norm (#9884) --- python/sglang/srt/models/glm4v.py | 3 +-- python/sglang/srt/models/qwen2_5_vl.py | 5 +++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/python/sglang/srt/models/glm4v.py b/python/sglang/srt/models/glm4v.py index 79eae3946..95c70804f 100644 --- a/python/sglang/srt/models/glm4v.py +++ b/python/sglang/srt/models/glm4v.py @@ -93,9 +93,8 @@ class Glm4vVisionBlock(Qwen2_5_VisionBlock): quant_config=quant_config, prefix=prefix, num_dummy_heads=config.num_dummy_heads, + rms_norm_eps=config.rms_norm_eps, ) - self.norm1 = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) - self.norm2 = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.mlp = Glm4vVisionMLP( config.hidden_size, diff --git a/python/sglang/srt/models/qwen2_5_vl.py b/python/sglang/srt/models/qwen2_5_vl.py index 20165c3c7..82370de54 100644 --- a/python/sglang/srt/models/qwen2_5_vl.py +++ b/python/sglang/srt/models/qwen2_5_vl.py @@ -113,12 +113,13 @@ class Qwen2_5_VisionBlock(nn.Module): quant_config: Optional[QuantizationConfig] = None, prefix: str = "", num_dummy_heads: int = 0, + rms_norm_eps: float = 1e-6, ) -> None: super().__init__() if norm_layer is None: norm_layer = partial(nn.LayerNorm, eps=1e-6) - self.norm1 = RMSNorm(dim, eps=1e-6) - self.norm2 = RMSNorm(dim, eps=1e-6) + self.norm1 = RMSNorm(dim, eps=rms_norm_eps) + self.norm2 = RMSNorm(dim, eps=rms_norm_eps) if attn_implementation is None: softmax_in_single_precision = False