fix: torch.compile mode 改为 default（避免 CUDA Graph 因 N 变化重编译）

2026-06-13 14:20:14 +08:00
parent a74af49456
commit 480a81a033
1 changed files with 1 additions and 1 deletions
@@ -510,7 +510,7 @@ def load_model(ckpt_path, device='cuda:0'):
        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")
+                expert.forward = torch.compile(expert.forward, mode="default")
                cc += 1
        print(f"[INFO] torch.compile applied to {cc} Expert.forward methods")
    except Exception as e: