From d12b313b93e1547d9b02c3a84426aa88519fc494 Mon Sep 17 00:00:00 2001 From: Cheng Wan <54331508+ch-wan@users.noreply.github.com> Date: Fri, 28 Aug 2026 15:05:08 -0700 Subject: [PATCH] fix: KT's last MoE layer stops deferring experts again (#36921) --- python/sglang/srt/layers/moe/kt_ep_wrapper.py | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-) diff --git a/python/sglang/srt/layers/moe/kt_ep_wrapper.py b/python/sglang/srt/layers/moe/kt_ep_wrapper.py index 9c55d8e7a..4beccf3d2 100644 --- a/python/sglang/srt/layers/moe/kt_ep_wrapper.py +++ b/python/sglang/srt/layers/moe/kt_ep_wrapper.py @@ -76,14 +76,9 @@ def create_kt_config_from_server_args( if server_args.kt_weight_path is None: return None - # Try to get num_layers from model config - num_layers = None - try: - hf_config = server_args.get_hf_config() - num_layers = getattr(hf_config, "num_hidden_layers", None) - except Exception: - # If we can't get the config, num_layers will be None - pass + num_layers = getattr( + server_args.get_model_config().hf_config, "num_hidden_layers", None + ) return KTConfig( layer_idx=layer_idx,