Add 4 pt descriptor compression (#4227)

se_a, se_atten(DPA1), se_t, se_r

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

## Release Notes

- **New Features**
- Introduced a model compression feature across multiple descriptor
classes, enhancing performance and efficiency.
- Added `enable_compression` methods to various classes, allowing users
to enable and configure compression settings.

- **Bug Fixes**
- Improved error handling for unsupported compression scenarios and
parameter validation.

- **Tests**
- Added comprehensive unit tests for new compression functionalities
across multiple descriptor classes to ensure accuracy and reliability.

- **Documentation**
- Enhanced documentation for new methods and classes to clarify usage
and parameters related to compression.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Jinzhe Zeng <[email protected]>
Signed-off-by: Yan Wang <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Jinzhe Zeng <[email protected]>

Author: 3 people

deepmd/dpmodel/descriptor/make_base_descriptor.py

@@ -147,6 +147,31 @@ def compute_input_stats(
             """Update mean and stddev for descriptor elements."""
             raise NotImplementedError
 
+        def enable_compression(
+            self,
+            min_nbor_dist: float,
+            table_extrapolate: float = 5,
+            table_stride_1: float = 0.01,
+            table_stride_2: float = 0.1,
+            check_frequency: int = -1,
+        ) -> None:
+            """Receive the statisitcs (distance, max_nbor_size and env_mat_range) of the training data.
+
+            Parameters
+            ----------
+            min_nbor_dist
+                The nearest distance between atoms
+            table_extrapolate
+                The scale of model extrapolation
+            table_stride_1
+                The uniform stride of the first table
+            table_stride_2
+                The uniform stride of the second table
+            check_frequency
+                The overflow check frequency
+            """
+            raise NotImplementedError("This descriptor doesn't support compression!")
+
         @abstractmethod
         def fwd(
             self,

deepmd/pt/model/descriptor/dpa1.py

@@ -24,9 +24,15 @@
 from deepmd.pt.utils.env import (
     RESERVED_PRECISON_DICT,
 )
+from deepmd.pt.utils.tabulate import (
+    DPTabulate,
+)
 from deepmd.pt.utils.update_sel import (
     UpdateSel,
 )
+from deepmd.pt.utils.utils import (
+    ActivationFn,
+)
 from deepmd.utils.data_system import (
     DeepmdDataSystem,
 )
@@ -261,6 +267,8 @@ def __init__(
         if ln_eps is None:
             ln_eps = 1e-5
 
+        self.tebd_input_mode = tebd_input_mode
+
         del type, spin, attn_mask
         self.se_atten = DescrptBlockSeAtten(
             rcut,
@@ -293,6 +301,7 @@ def __init__(
         self.use_econf_tebd = use_econf_tebd
         self.use_tebd_bias = use_tebd_bias
         self.type_map = type_map
+        self.compress = False
         self.type_embedding = TypeEmbedNet(
             ntypes,
             tebd_dim,
@@ -551,6 +560,84 @@ def t_cvt(xx):
         )
         return obj
 
+    def enable_compression(
+        self,
+        min_nbor_dist: float,
+        table_extrapolate: float = 5,
+        table_stride_1: float = 0.01,
+        table_stride_2: float = 0.1,
+        check_frequency: int = -1,
+    ) -> None:
+        """Receive the statisitcs (distance, max_nbor_size and env_mat_range) of the training data.
+
+        Parameters
+        ----------
+        min_nbor_dist
+            The nearest distance between atoms
+        table_extrapolate
+            The scale of model extrapolation
+        table_stride_1
+            The uniform stride of the first table
+        table_stride_2
+            The uniform stride of the second table
+        check_frequency
+            The overflow check frequency
+        """
+        # do some checks before the mocel compression process
+        if self.compress:
+            raise ValueError("Compression is already enabled.")
+        assert (
+            not self.se_atten.resnet_dt
+        ), "Model compression error: descriptor resnet_dt must be false!"
+        for tt in self.se_atten.exclude_types:
+            if (tt[0] not in range(self.se_atten.ntypes)) or (
+                tt[1] not in range(self.se_atten.ntypes)
+            ):
+                raise RuntimeError(
+                    "exclude types"
+                    + str(tt)
+                    + " must within the number of atomic types "
+                    + str(self.se_atten.ntypes)
+                    + "!"
+                )
+        if (
+            self.se_atten.ntypes * self.se_atten.ntypes
+            - len(self.se_atten.exclude_types)
+            == 0
+        ):
+            raise RuntimeError(
+                "Empty embedding-nets are not supported in model compression!"
+            )
+
+        if self.se_atten.attn_layer != 0:
+            raise RuntimeError("Cannot compress model when attention layer is not 0.")
+
+        if self.tebd_input_mode != "strip":
+            raise RuntimeError("Cannot compress model when tebd_input_mode == 'concat'")
+
+        data = self.serialize()
+        self.table = DPTabulate(
+            self,
+            data["neuron"],
+            data["type_one_side"],
+            data["exclude_types"],
+            ActivationFn(data["activation_function"]),
+        )
+        self.table_config = [
+            table_extrapolate,
+            table_stride_1,
+            table_stride_2,
+            check_frequency,
+        ]
+        self.lower, self.upper = self.table.build(
+            min_nbor_dist, table_extrapolate, table_stride_1, table_stride_2
+        )
+
+        self.se_atten.enable_compression(
+            self.table.data, self.table_config, self.lower, self.upper
+        )
+        self.compress = True
+
     def forward(
         self,
         extended_coord: torch.Tensor,

deepmd/pt/model/descriptor/se_a.py

@@ -58,11 +58,34 @@
 from deepmd.pt.utils.exclude_mask import (
     PairExcludeMask,
 )
+from deepmd.pt.utils.tabulate import (
+    DPTabulate,
+)
+from deepmd.pt.utils.utils import (
+    ActivationFn,
+)
 
 from .base_descriptor import (
     BaseDescriptor,
 )
 
+if not hasattr(torch.ops.deepmd, "tabulate_fusion_se_a"):
+
+    def tabulate_fusion_se_a(
+        argument0,
+        argument1,
+        argument2,
+        argument3,
+        argument4,
+    ) -> list[torch.Tensor]:
+        raise NotImplementedError(
+            "tabulate_fusion_se_a is not available since customized PyTorch OP library is not built when freezing the model. "
+            "See documentation for model compression for details."
+        )
+
+    # Note: this hack cannot actually save a model that can be runned using LAMMPS.
+    torch.ops.deepmd.tabulate_fusion_se_a = tabulate_fusion_se_a
+
 
 @BaseDescriptor.register("se_e2_a")
 @BaseDescriptor.register("se_a")
@@ -93,6 +116,7 @@ def __init__(
             raise NotImplementedError("old implementation of spin is not supported.")
         super().__init__()
         self.type_map = type_map
+        self.compress = False
         self.sea = DescrptBlockSeA(
             rcut,
             rcut_smth,
@@ -225,6 +249,53 @@ def reinit_exclude(
         """Update the type exclusions."""
         self.sea.reinit_exclude(exclude_types)
 
+    def enable_compression(
+        self,
+        min_nbor_dist: float,
+        table_extrapolate: float = 5,
+        table_stride_1: float = 0.01,
+        table_stride_2: float = 0.1,
+        check_frequency: int = -1,
+    ) -> None:
+        """Receive the statisitcs (distance, max_nbor_size and env_mat_range) of the training data.
+
+        Parameters
+        ----------
+        min_nbor_dist
+            The nearest distance between atoms
+        table_extrapolate
+            The scale of model extrapolation
+        table_stride_1
+            The uniform stride of the first table
+        table_stride_2
+            The uniform stride of the second table
+        check_frequency
+            The overflow check frequency
+        """
+        if self.compress:
+            raise ValueError("Compression is already enabled.")
+        data = self.serialize()
+        self.table = DPTabulate(
+            self,
+            data["neuron"],
+            data["type_one_side"],
+            data["exclude_types"],
+            ActivationFn(data["activation_function"]),
+        )
+        self.table_config = [
+            table_extrapolate,
+            table_stride_1,
+            table_stride_2,
+            check_frequency,
+        ]
+        self.lower, self.upper = self.table.build(
+            min_nbor_dist, table_extrapolate, table_stride_1, table_stride_2
+        )
+        self.sea.enable_compression(
+            self.table.data, self.table_config, self.lower, self.upper
+        )
+        self.compress = True
+
     def forward(
         self,
         coord_ext: torch.Tensor,
@@ -366,6 +437,10 @@ def update_sel(
 class DescrptBlockSeA(DescriptorBlock):
     ndescrpt: Final[int]
     __constants__: ClassVar[list] = ["ndescrpt"]
+    lower: dict[str, int]
+    upper: dict[str, int]
+    table_data: dict[str, torch.Tensor]
+    table_config: list[Union[int, float]]
 
     def __init__(
         self,
@@ -425,6 +500,13 @@ def __init__(
         self.register_buffer("mean", mean)
         self.register_buffer("stddev", stddev)
 
+        # add for compression
+        self.compress = False
+        self.lower = {}
+        self.upper = {}
+        self.table_data = {}
+        self.table_config = []
+
         ndim = 1 if self.type_one_side else 2
         filter_layers = NetworkCollection(
             ndim=ndim, ntypes=len(sel), network_type="embedding_network"
@@ -443,6 +525,7 @@ def __init__(
         self.filter_layers = filter_layers
         self.stats = None
         # set trainable
+        self.trainable = trainable
         for param in self.parameters():
             param.requires_grad = trainable
 
@@ -470,6 +553,10 @@ def get_dim_out(self) -> int:
         """Returns the output dimension."""
         return self.dim_out
 
+    def get_dim_rot_mat_1(self) -> int:
+        """Returns the first dimension of the rotation matrix. The rotation is of shape dim_1 x 3."""
+        return self.filter_neuron[-1]
+
     def get_dim_emb(self) -> int:
         """Returns the output dimension."""
         return self.neuron[-1]
@@ -578,6 +665,19 @@ def reinit_exclude(
         self.exclude_types = exclude_types
         self.emask = PairExcludeMask(self.ntypes, exclude_types=exclude_types)
 
+    def enable_compression(
+        self,
+        table_data,
+        table_config,
+        lower,
+        upper,
+    ) -> None:
+        self.compress = True
+        self.table_data = table_data
+        self.table_config = table_config
+        self.lower = lower
+        self.upper = upper
+
     def forward(
         self,
         nlist: torch.Tensor,
@@ -627,6 +727,7 @@ def forward(
         for embedding_idx, ll in enumerate(self.filter_layers.networks):
             if self.type_one_side:
                 ii = embedding_idx
+                ti = -1
                 # torch.jit is not happy with slice(None)
                 # ti_mask = torch.ones(nfnl, dtype=torch.bool, device=dmatrix.device)
                 # applying a mask seems to cause performance degradation
@@ -648,10 +749,35 @@ def forward(
                 rr = dmatrix[:, self.sec[ii] : self.sec[ii + 1], :]
             rr = rr * mm[:, :, None]
             ss = rr[:, :, :1]
-            # nfnl x nt x ng
-            gg = ll.forward(ss)
-            # nfnl x 4 x ng
-            gr = torch.matmul(rr.permute(0, 2, 1), gg)
+
+            if self.compress:
+                if self.type_one_side:
+                    net = "filter_-1_net_" + str(ii)
+                else:
+                    net = "filter_" + str(ti) + "_net_" + str(ii)
+                info = [
+                    self.lower[net],
+                    self.upper[net],
+                    self.upper[net] * self.table_config[0],
+                    self.table_config[1],
+                    self.table_config[2],
+                    self.table_config[3],
+                ]
+                ss = ss.reshape(-1, 1)  # xyz_scatter_tensor in tf
+                tensor_data = self.table_data[net].to(ss.device).to(dtype=self.prec)
+                gr = torch.ops.deepmd.tabulate_fusion_se_a(
+                    tensor_data.contiguous(),
+                    torch.tensor(info, dtype=self.prec, device="cpu").contiguous(),
+                    ss.contiguous(),
+                    rr.contiguous(),
+                    self.filter_neuron[-1],
+                )[0]
+            else:
+                # nfnl x nt x ng
+                gg = ll.forward(ss)
+                # nfnl x 4 x ng
+                gr = torch.matmul(rr.permute(0, 2, 1), gg)
+
             if ti_mask is not None:
                 xyz_scatter[ti_mask] += gr
             else: