fix: move k_lens to GPU in SDPA fallback (tested locally)

model_manager.py

@@ -130,6 +130,7 @@ class ModelManager:
             "        attn_mask = None\n"
             "        is_causal_flag = causal\n"
             "        if k_lens is not None:\n"
+            "            k_lens = k_lens.to(q.device)\n"
             "            valid = torch.arange(lk, device=q.device).unsqueeze(0) < k_lens.unsqueeze(1)\n"
             "            attn_mask = torch.where(valid[:, None, None, :], 0.0, float('-inf')).to(dtype=dtype)\n"
             "            is_causal_flag = False\n"