fix: cuda tests of linear and pair atomic model (#3248)

This PR is to fix LinearAtomicModel GPU compatibility.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

deepmd/pt/model/model/pairtab_atomic_model.py

@@ -7,9 +7,6 @@
 )
 
 import torch
-from torch import (
-    nn,
-)
 
 from deepmd.dpmodel import (
     FittingOutputDef,
@@ -22,9 +19,12 @@
 from .base_atomic_model import (
     BaseAtomicModel,
 )
+from .model import (
+    BaseModel,
+)
 
 
-class PairTabModel(nn.Module, BaseAtomicModel):
+class PairTabModel(BaseModel, BaseAtomicModel):
     """Pairwise tabulation energy model.
 
     This model can be used to tabulate the pairwise energy between atoms for either
@@ -62,11 +62,11 @@ def __init__(
                 tab_info,
                 tab_data,
             ) = self.tab.get()  # this returns -> Tuple[np.array, np.array]
-            self.tab_info = torch.from_numpy(tab_info)
-            self.tab_data = torch.from_numpy(tab_data)
+            self.register_buffer("tab_info", torch.from_numpy(tab_info))
+            self.register_buffer("tab_data", torch.from_numpy(tab_data))
         else:
-            self.tab_info = None
-            self.tab_data = None
+            self.register_buffer("tab_info", None)
+            self.register_buffer("tab_data", None)
 
         # self.model_type = "ener"
         # self.model_version = MODEL_VERSION ## this shoud be in the parent class
@@ -118,8 +118,8 @@ def deserialize(cls, data) -> "PairTabModel":
         tab = PairTab.deserialize(data["tab"])
         tab_model = cls(None, rcut, sel)
         tab_model.tab = tab
-        tab_model.tab_info = torch.from_numpy(tab_model.tab.tab_info)
-        tab_model.tab_data = torch.from_numpy(tab_model.tab.tab_data)
+        tab_model.register_buffer("tab_info", torch.from_numpy(tab_model.tab.tab_info))
+        tab_model.register_buffer("tab_data", torch.from_numpy(tab_model.tab.tab_data))
         return tab_model
 
     def forward_atomic(

source/tests/pt/model/test_linear_atomic_model.py

@@ -56,8 +56,8 @@ def test_pairwise(self, mock_loadtxt):
                 [0.25, 0.0, 0.0, 0.0],
             ]
         )
-        extended_atype = torch.tensor([[0, 0]])
-        nlist = torch.tensor([[[1], [-1]]])
+        extended_atype = torch.tensor([[0, 0]]).to(env.DEVICE)
+        nlist = torch.tensor([[[1], [-1]]]).to(env.DEVICE)
 
         ds = DescrptSeA(
             rcut=0.3,
@@ -82,7 +82,7 @@ def test_pairwise(self, mock_loadtxt):
             zbl_model,
             sw_rmin=0.1,
             sw_rmax=0.25,
-        )
+        ).to(env.DEVICE)
         wgt_res = []
         for dist in np.linspace(0.05, 0.3, 10):
             extended_coord = torch.tensor(
@@ -92,7 +92,7 @@ def test_pairwise(self, mock_loadtxt):
                         [0.0, dist, 0.0],
                     ],
                 ]
-            )
+            ).to(env.DEVICE)
 
             wgt_model.forward_atomic(extended_coord, extended_atype, nlist)
 
@@ -112,7 +112,7 @@ def test_pairwise(self, mock_loadtxt):
                 [0.0, 0.0],
             ],
             dtype=torch.float64,
-        )
+        ).to(env.DEVICE)
         torch.testing.assert_close(results, excepted_res, rtol=0.0001, atol=0.0001)