fix(2024Q1): fix lammps nlist sort with large sel

.github/workflows/build_wheel.yml

@@ -151,6 +151,8 @@ jobs:
           - variant: "_cu11"
             cuda_version: "11"
     steps:
+      - name: Delete huge unnecessary tools folder
+        run: rm -rf /opt/hostedtoolcache
       - uses: actions/checkout@v4
       - uses: actions/download-artifact@v4
         with:

.github/workflows/test_cuda.yml

@@ -47,7 +47,7 @@ jobs:
          && sudo apt-get -y install cuda-12-3 libcudnn8=8.9.5.*-1+cuda12.3
       if: false  # skip as we use nvidia image
     - run: python -m pip install -U "pip>=21.3.1,!=23.0.0"
-    - run: python -m pip install "tensorflow>=2.15.0rc0" "torch>=2.2.0"
+    - run: python -m pip install "tensorflow>=2.15.0rc0" "torch==2.3.1.*"
     - run: python -m pip install -v -e .[gpu,test,lmp,cu12,torch] mpi4py "ase @ https://gitlab.com/ase/ase/-/archive/8c5aa5fd6448c5cfb517a014dccf2b214a9dfa8f/ase-8c5aa5fd6448c5cfb517a014dccf2b214a9dfa8f.tar.gz"
       env:
         DP_VARIANT: cuda

deepmd/dpmodel/atomic_model/dp_atomic_model.py

@@ -104,6 +104,10 @@ def get_out_bias(self) -> np.ndarray:
         """Return the output bias of the atomic model."""
         return self.fitting["bias_atom_e"]
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the atomic model needs sorted nlist when using `forward_lower`."""
+        return self.descriptor.need_sorted_nlist_for_lower()
+
     def forward_atomic(
         self,
         extended_coord: np.ndarray,

deepmd/dpmodel/atomic_model/linear_atomic_model.py

@@ -111,6 +111,10 @@ def _sort_rcuts_sels(self) -> Tuple[List[float], List[int]]:
         )
         return [p[0] for p in zipped], [p[1] for p in zipped]
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the atomic model needs sorted nlist when using `forward_lower`."""
+        return True
+
     def forward_atomic(
         self,
         extended_coord,

deepmd/dpmodel/atomic_model/make_base_atomic_model.py

@@ -134,6 +134,10 @@ def mixed_types(self) -> bool:
             """
             pass
 
+        @abstractmethod
+        def need_sorted_nlist_for_lower(self) -> bool:
+            """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+
         @abstractmethod
         def fwd(
             self,

deepmd/dpmodel/atomic_model/pairtab_atomic_model.py

@@ -177,6 +177,10 @@ def deserialize(cls, data) -> "PairTabAtomicModel":
         tab_model.tab_data = tab_model.tab.tab_data.reshape(ntypes, ntypes, nspline, 4)
         return tab_model
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the atomic model needs sorted nlist when using `forward_lower`."""
+        return False
+
     def forward_atomic(
         self,
         extended_coord,

deepmd/dpmodel/descriptor/hybrid.py

@@ -140,6 +140,10 @@ def compute_input_stats(self, merged: List[dict], path: Optional[DPPath] = None)
         for descrpt in self.descrpt_list:
             descrpt.compute_input_stats(merged, path)
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return True
+
     def call(
         self,
         coord_ext,

deepmd/dpmodel/descriptor/make_base_descriptor.py

@@ -103,6 +103,10 @@ def compute_input_stats(
             """Update mean and stddev for descriptor elements."""
             raise NotImplementedError
 
+        @abstractmethod
+        def need_sorted_nlist_for_lower(self) -> bool:
+            """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+
         @abstractmethod
         def fwd(
             self,

deepmd/dpmodel/descriptor/se_e2_a.py

@@ -283,6 +283,10 @@ def reinit_exclude(
         self.exclude_types = exclude_types
         self.emask = PairExcludeMask(self.ntypes, exclude_types=exclude_types)
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return False
+
     def call(
         self,
         coord_ext,

deepmd/dpmodel/descriptor/se_r.py

@@ -232,6 +232,10 @@ def cal_g(
         gg = self.embeddings[(ll,)].call(ss)
         return gg
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return False
+
     def call(
         self,
         coord_ext,

deepmd/dpmodel/model/make_model.py

@@ -214,7 +214,12 @@ def call_lower(
             """
             nframes, nall = extended_atype.shape[:2]
             extended_coord = extended_coord.reshape(nframes, -1, 3)
-            nlist = self.format_nlist(extended_coord, extended_atype, nlist)
+            nlist = self.format_nlist(
+                extended_coord,
+                extended_atype,
+                nlist,
+                extra_nlist_sort=self.need_sorted_nlist_for_lower(),
+            )
             cc_ext, _, fp, ap, input_prec = self.input_type_cast(
                 extended_coord, fparam=fparam, aparam=aparam
             )
@@ -309,6 +314,7 @@ def format_nlist(
             extended_coord: np.ndarray,
             extended_atype: np.ndarray,
             nlist: np.ndarray,
+            extra_nlist_sort: bool = False,
         ):
             """Format the neighbor list.
 
@@ -334,6 +340,8 @@ def format_nlist(
                 atomic type in extended region. nf x nall
             nlist
                 neighbor list. nf x nloc x nsel
+            extra_nlist_sort
+                whether to forcibly sort the nlist.
 
             Returns
             -------
@@ -343,7 +351,12 @@ def format_nlist(
             """
             n_nf, n_nloc, n_nnei = nlist.shape
             mixed_types = self.mixed_types()
-            ret = self._format_nlist(extended_coord, nlist, sum(self.get_sel()))
+            ret = self._format_nlist(
+                extended_coord,
+                nlist,
+                sum(self.get_sel()),
+                extra_nlist_sort=extra_nlist_sort,
+            )
             if not mixed_types:
                 ret = nlist_distinguish_types(ret, extended_atype, self.get_sel())
             return ret
@@ -353,6 +366,7 @@ def _format_nlist(
             extended_coord: np.ndarray,
             nlist: np.ndarray,
             nnei: int,
+            extra_nlist_sort: bool = False,
         ):
             n_nf, n_nloc, n_nnei = nlist.shape
             extended_coord = extended_coord.reshape([n_nf, -1, 3])
@@ -368,8 +382,9 @@ def _format_nlist(
                     ],
                     axis=-1,
                 )
-            elif n_nnei > nnei:
-                # make a copy before revise
+
+            if n_nnei > nnei or extra_nlist_sort:
+                n_nf, n_nloc, n_nnei = nlist.shape
                 m_real_nei = nlist >= 0
                 ret = np.where(m_real_nei, nlist, 0)
                 coord0 = extended_coord[:, :n_nloc, :]
@@ -382,9 +397,11 @@ def _format_nlist(
                 ret = np.take_along_axis(ret, ret_mapping, axis=2)
                 ret = np.where(rr > rcut, -1, ret)
                 ret = ret[..., :nnei]
-            else:  # n_nnei == nnei:
-                # copy anyway...
+            # not extra_nlist_sort and n_nnei <= nnei:
+            elif n_nnei == nnei:
                 ret = nlist
+            else:
+                pass
             assert ret.shape[-1] == nnei
             return ret
 
@@ -469,6 +486,10 @@ def mixed_types(self) -> bool:
             """
             return self.atomic_model.mixed_types()
 
+        def need_sorted_nlist_for_lower(self) -> bool:
+            """Returns whether the model needs sorted nlist when using `forward_lower`."""
+            return self.atomic_model.need_sorted_nlist_for_lower()
+
         def atomic_output_def(self) -> FittingOutputDef:
             """Get the output def of the atomic model."""
             return self.atomic_model.atomic_output_def()

deepmd/pt/model/atomic_model/dp_atomic_model.py

@@ -100,6 +100,10 @@ def mixed_types(self) -> bool:
         """
         return self.descriptor.mixed_types()
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the atomic model needs sorted nlist when using `forward_lower`."""
+        return self.descriptor.need_sorted_nlist_for_lower()
+
     def serialize(self) -> dict:
         dd = BaseAtomicModel.serialize(self)
         dd.update(

deepmd/pt/model/atomic_model/linear_atomic_model.py

@@ -132,6 +132,10 @@ def _sort_rcuts_sels(self) -> Tuple[List[float], List[int]]:
         sorted_sels: List[int] = outer_sorted[:, 1].to(torch.int64).tolist()
         return sorted_rcuts, sorted_sels
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the atomic model needs sorted nlist when using `forward_lower`."""
+        return True
+
     def forward_atomic(
         self,
         extended_coord: torch.Tensor,

deepmd/pt/model/atomic_model/pairtab_atomic_model.py

@@ -164,6 +164,10 @@ def mixed_types(self) -> bool:
         # to match DPA1 and DPA2.
         return True
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the atomic model needs sorted nlist when using `forward_lower`."""
+        return False
+
     def serialize(self) -> dict:
         dd = BaseAtomicModel.serialize(self)
         dd.update(

deepmd/pt/model/descriptor/descriptor.py

@@ -159,6 +159,10 @@ def forward(
         """Calculate DescriptorBlock."""
         pass
 
+    @abstractmethod
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor block needs sorted nlist when using `forward_lower`."""
+
 
 def make_default_type_embedding(
     ntypes,

deepmd/pt/model/descriptor/dpa1.py

@@ -150,6 +150,10 @@ def mixed_types(self) -> bool:
         """
         return self.se_atten.mixed_types()
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return self.se_atten.need_sorted_nlist_for_lower()
+
     def share_params(self, base_class, shared_level, resume=False):
         """
         Share the parameters of self to the base_class with shared_level during multitask training.

deepmd/pt/model/descriptor/dpa2.py

@@ -307,6 +307,10 @@ def mixed_types(self) -> bool:
         """
         return True
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return True
+
     def share_params(self, base_class, shared_level, resume=False):
         """
         Share the parameters of self to the base_class with shared_level during multitask training.

deepmd/pt/model/descriptor/hybrid.py

@@ -140,6 +140,10 @@ def mixed_types(self):
         """
         return any(descrpt.mixed_types() for descrpt in self.descrpt_list)
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return True
+
     def share_params(self, base_class, shared_level, resume=False):
         """
         Share the parameters of self to the base_class with shared_level during multitask training.
@@ -369,6 +373,10 @@ def mixed_types(self) -> bool:
         """
         return all(descriptor.mixed_types() for descriptor in self.descriptor_list)
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor block needs sorted nlist when using `forward_lower`."""
+        return True
+
     @property
     def dim_out(self):
         """Returns the output dimension of this descriptor."""

deepmd/pt/model/descriptor/repformers.py

@@ -205,6 +205,10 @@ def mixed_types(self) -> bool:
         """
         return True
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor block needs sorted nlist when using `forward_lower`."""
+        return False
+
     @property
     def dim_out(self):
         """Returns the output dimension of this descriptor."""

deepmd/pt/model/descriptor/se_a.py

@@ -132,6 +132,10 @@ def mixed_types(self):
         """
         return self.sea.mixed_types()
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return self.sea.need_sorted_nlist_for_lower()
+
     def share_params(self, base_class, shared_level, resume=False):
         """
         Share the parameters of self to the base_class with shared_level during multitask training.
@@ -433,6 +437,10 @@ def mixed_types(self) -> bool:
         """
         return False
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor block needs sorted nlist when using `forward_lower`."""
+        return False
+
     @property
     def dim_out(self):
         """Returns the output dimension of this descriptor."""

deepmd/pt/model/descriptor/se_atten.py

@@ -196,6 +196,10 @@ def mixed_types(self) -> bool:
         """
         return True
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor block needs sorted nlist when using `forward_lower`."""
+        return False
+
     @property
     def dim_out(self):
         """Returns the output dimension of this descriptor."""

deepmd/pt/model/descriptor/se_r.py

@@ -160,6 +160,10 @@ def mixed_types(self) -> bool:
         """
         return False
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the descriptor needs sorted nlist when using `forward_lower`."""
+        return False
+
     def share_params(self, base_class, shared_level, resume=False):
         """
         Share the parameters of self to the base_class with shared_level during multitask training.

deepmd/pt/model/model/dipole_model.py

@@ -76,6 +76,7 @@ def forward_lower(
             fparam=fparam,
             aparam=aparam,
             do_atomic_virial=do_atomic_virial,
+            extra_nlist_sort=self.need_sorted_nlist_for_lower(),
         )
         if self.get_fitting_net() is not None:
             model_predict = {}

deepmd/pt/model/model/dos_model.py

@@ -74,6 +74,7 @@ def forward_lower(
             fparam=fparam,
             aparam=aparam,
             do_atomic_virial=do_atomic_virial,
+            extra_nlist_sort=self.need_sorted_nlist_for_lower(),
         )
         if self.get_fitting_net() is not None:
             model_predict = {}

deepmd/pt/model/model/dp_zbl_model.py

@@ -86,6 +86,7 @@ def forward_lower(
             fparam=fparam,
             aparam=aparam,
             do_atomic_virial=do_atomic_virial,
+            extra_nlist_sort=self.need_sorted_nlist_for_lower(),
         )
 
         model_predict = {}

deepmd/pt/model/model/ener_model.py

@@ -78,6 +78,7 @@ def forward_lower(
             fparam=fparam,
             aparam=aparam,
             do_atomic_virial=do_atomic_virial,
+            extra_nlist_sort=self.need_sorted_nlist_for_lower(),
         )
         if self.get_fitting_net() is not None:
             model_predict = {}

deepmd/pt/model/model/frozen.py

@@ -102,6 +102,10 @@ def mixed_types(self) -> bool:
         """
         return self.model.mixed_types()
 
+    def need_sorted_nlist_for_lower(self) -> bool:
+        """Returns whether the model needs sorted nlist when using `forward_lower`."""
+        return self.model.need_sorted_nlist_for_lower()
+
     @torch.jit.export
     def forward(
         self,