feat: torch.compile 单独编译 Expert.forward（fc1→relu→fc2 融合）

- 仅编译 Expert.forward，不碰 MoE 循环和 attention
- 纯静态函数无分支，编译成功率高
- 替代 2:4 稀疏方案

代码/code/infer.py

@@ -505,30 +505,16 @@ def load_model(ckpt_path, device='cuda:0'):
 
     model.to(dev)
 
-    # === 2:4 非结构化稀疏：仅裁剪 Expert FFN 权重，不动 attention/gate ===
+    # === torch.compile 融合 Expert FFN（fc1→relu→fc2），不含动态分支 ===
     try:
-        sp_count = 0
-        for layer in model.seq_encoder.moe:
-            for expert in layer.experts:
-                for attr in ['fc1', 'fc2']:
-                    linear = getattr(expert, attr)
-                    w = linear.weight.data.clone()
-                    shape = w.shape
-                    # 2:4 幅度剪枝：每 4 个连续元素保留 top 2
-                    w_flat = w.reshape(-1, 4)
-                    _, top_idx = torch.topk(w_flat.abs(), k=2, dim=1)
-                    mask = torch.zeros_like(w_flat)
-                    mask.scatter_(1, top_idx, 1.0)
-                    pruned = (w_flat * mask).reshape(shape)
-                    sparse_w = torch.sparse.to_sparse_semi_structured(pruned)
-                    bias = linear.bias
-                    linear.forward = lambda x, sw=sparse_w, b=bias: (
-                        torch.matmul(x, sw.t()) + b if b is not None else torch.matmul(x, sw.t())
-                    )
-                    sp_count += 1
-        print(f"[INFO] 2:4 sparsity applied to {sp_count} Expert Linear layers")
+        cc = 0
+        for moe_layer in model.seq_encoder.moe:
+            for expert in moe_layer.experts:
+                expert.forward = torch.compile(expert.forward, mode="reduce-overhead")
+                cc += 1
+        print(f"[INFO] torch.compile applied to {cc} Expert.forward methods")
     except Exception as e:
-        print(f"[WARNING] 2:4 sparsity failed ({e}), keeping dense weights")
+        print(f"[WARNING] Expert torch.compile failed ({e}), using original forward")
 
     model.eval()
     print(f"[INFO] Model ready. Device: {dev}")