Merge ccb352f901 into cc312c1cc6

fix: eval_llm.py Ctrl+C 退出异常修复

README.md

@@ -230,6 +230,10 @@ minimind2系列旧模型均经过权重映射+（微调训练）QKVO线性层校
 # 克隆仓库、安装依赖
 git clone --depth 1 https://github.com/jingyaogong/minimind
 cd minimind && pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple
+#部分老版本Windows powershell不能识别&&，Windows用户应将&&链式命令改为分行写法
+git clone --depth 1 https://github.com/jingyaogong/minimind
+cd minimind 
+pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple
 ```
 
 ## Ⅰ 🚀 模型推理
@@ -242,6 +246,13 @@ cd minimind && pip install -r requirements.txt -i https://mirrors.aliyun.com/pyp
 modelscope download --model gongjy/minimind-3 --local_dir ./minimind-3
 # 方式2
 git clone https://huggingface.co/jingyaogong/minimind-3
+# 国内网络进入huggingface网站存在困难，可使用镜像站
+# Windows
+$env:HF_ENDPOINT = "https://hf-mirror.com" 
+git clone https://hf-mirror.com/jingyaogong/minimind-3
+#Linux
+export HF_ENDPOINT="https://hf-mirror.com"
+git clone https://hf-mirror.com/jingyaogong/minimind-3
 ```
 
 ### 2' CLI 推理

README_en.md

@@ -229,6 +229,10 @@ After this update, maintenance for the entire `minimind-v1` series will be disco
 # Clone repository and install dependencies
 git clone --depth 1 https://github.com/jingyaogong/minimind
 cd minimind && pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple
+#Tip:Some older versions of Windows PowerShell (v5.1 and below) do not recognize the && operator. 
+git clone --depth 1 https://github.com/jingyaogong/minimind
+cd minimind 
+pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple
 ```
 
 ## Ⅰ 🚀 Model Inference

eval_llm.py

@@ -91,4 +91,7 @@ def main():
         print(f'\n[Speed]: {gen_tokens / (time.time() - st):.2f} tokens/s\n\n') if args.show_speed else print('\n\n')
 
 if __name__ == "__main__":
+    try:
         main()
+    except (EOFError, KeyboardInterrupt):
+        print()

model/model_minimind.py

@@ -158,7 +158,12 @@ class MOEFeedForward(nn.Module):
         x_flat = x.view(-1, hidden_dim)
         scores = F.softmax(self.gate(x_flat), dim=-1)
         topk_weight, topk_idx = torch.topk(scores, k=self.config.num_experts_per_tok, dim=-1, sorted=False)
-        if self.config.norm_topk_prob: topk_weight = topk_weight / (topk_weight.sum(dim=-1, keepdim=True) + 1e-20)
+        if self.config.norm_topk_prob:
+            if self.config.num_experts_per_tok > 1: 
+                topk_weight = topk_weight / (topk_weight.sum(dim=-1, keepdim=True) + 1e-20)
+            else:
+                top1 = torch.topk(F.softmax(self.gate(x_flat.detach()), dim=-1), k=1, dim=-1, sorted=False)[0]
+                topk_weight = top1 - top1.detach() + 1.0 # k=1: 1.0 forward on purpose, gradient via straight-through
         y = torch.zeros_like(x_flat)
         for i, expert in enumerate(self.experts):
             mask = (topk_idx == i)