From 9089310edefb6007ea9d1a5be453bc3676133e0a Mon Sep 17 00:00:00 2001
From: jingyaogong <gongjy.cs@qq.com>
Date: Fri, 27 Mar 2026 16:22:38 +0800
Subject: [PATCH] [update] fp16 inference

---
 eval_vlm.py             | 2 +-
 scripts/web_demo_vlm.py | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/eval_vlm.py b/eval_vlm.py
index 498cf15..cec1512 100644
--- a/eval_vlm.py
+++ b/eval_vlm.py
@@ -26,7 +26,7 @@ def init_model(args):
         model.vision_encoder, model.processor = MiniMindVLM.get_vision_model("./model/siglip2-base-p16-ve")
     get_model_params(model, model.config)
     preprocess = model.processor
-    return model.eval().to(args.device), tokenizer, preprocess
+    return model.half().eval().to(args.device), tokenizer, preprocess
 
 
 def main():
diff --git a/scripts/web_demo_vlm.py b/scripts/web_demo_vlm.py
index 1d679ec..41a7d58 100644
--- a/scripts/web_demo_vlm.py
+++ b/scripts/web_demo_vlm.py
@@ -40,7 +40,7 @@ def load_vlm_model(model_path):
     if model.vision_encoder is None: raise FileNotFoundError(f"视觉编码器未找到: {vision_model_path}")
     preprocess = model.processor
     lm_config = model.config
-    model = model.eval().to(device)
+    model = model.half().eval().to(device)
     model.vision_encoder = model.vision_encoder.to(device)
     current_model_name = os.path.basename(model_path)
     param_str = f'{sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.2f}M'