From e3590e6bdae66a246c3eaa0ff42b56d3b1d9550f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=88=98=E8=88=AA=E5=AE=87?= <3364451258@qq.com> Date: Sun, 14 Jun 2026 12:09:28 +0800 Subject: [PATCH] =?UTF-8?q?perf:=20=E9=99=8D=E4=BD=8E=E5=90=88=E5=B9=B6?= =?UTF-8?q?=E9=98=88=E5=80=BC=200.85=E2=86=920.80=EF=BC=88=E7=BB=A7?= =?UTF-8?q?=E7=BB=AD=E6=8E=A2=E5=BA=95=EF=BC=89?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- 代码/code/infer.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/代码/code/infer.py b/代码/code/infer.py index c24fb58..e32fec8 100644 --- a/代码/code/infer.py +++ b/代码/code/infer.py @@ -502,7 +502,7 @@ def load_model(ckpt_path, device='cuda:0'): print("[INFO] Model converted to FP16 (embedding kept in FP32)") # === 按 Expert 权重相似度合并冗余 expert === - _merge_experts(model, sim_threshold=0.85) + _merge_experts(model, sim_threshold=0.80) else: print(f"[WARNING] Checkpoint {ckpt_path} not found, using random weights")