From fb8d7eedda98617b743ecd6c5e03bd391e8ee994 Mon Sep 17 00:00:00 2001 From: weireweire Date: Wed, 2 Sep 2026 05:19:14 +0800 Subject: [PATCH] Fix dummy initialization of inverse weight scales (#35491) Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com> Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com> --- python/sglang/srt/model_loader/weight_utils.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/model_loader/weight_utils.py b/python/sglang/srt/model_loader/weight_utils.py index 965123fcc..fb96ff210 100644 --- a/python/sglang/srt/model_loader/weight_utils.py +++ b/python/sglang/srt/model_loader/weight_utils.py @@ -1664,8 +1664,11 @@ def initialize_dummy_weights( is fixed, the random values generated by this function only depends on the parameter's number of elements and its data type. """ - for param in model.state_dict().values(): + for name, param in model.state_dict().items(): if torch.is_floating_point(param): + if name.endswith("weight_scale_inv"): + param.fill_(1.0) + continue generator = torch.Generator(device=param.data.device) generator.manual_seed(seed) # Tensor subclasses such as MXFP8 wrappers expose a low-bit raw