fix(dpmodel): fix precision

.github/workflows/test_python.yml

@@ -62,6 +62,7 @@ jobs:
       env:
         NUM_WORKERS: 0
         DP_TEST_TF2_ONLY: 1
+        DP_DTYPE_PROMOTION_STRICT: 1
       if: matrix.group == 1
     - run: mv .test_durations .test_durations_${{ matrix.group }}
     - name: Upload partial durations

deepmd/dpmodel/atomic_model/base_atomic_model.py

@@ -201,18 +201,19 @@ def forward_common_atomic(
         ret_dict = self.apply_out_stat(ret_dict, atype)
 
         # nf x nloc
-        atom_mask = ext_atom_mask[:, :nloc].astype(xp.int32)
+        atom_mask = ext_atom_mask[:, :nloc]
         if self.atom_excl is not None:
-            atom_mask *= self.atom_excl.build_type_exclude_mask(atype)
+            atom_mask = xp.logical_and(
+                atom_mask, self.atom_excl.build_type_exclude_mask(atype)
+            )
 
         for kk in ret_dict.keys():
             out_shape = ret_dict[kk].shape
             out_shape2 = math.prod(out_shape[2:])
-            ret_dict[kk] = (
-                ret_dict[kk].reshape([out_shape[0], out_shape[1], out_shape2])
-                * atom_mask[:, :, None]
-            ).reshape(out_shape)
-        ret_dict["mask"] = atom_mask
+            tmp_arr = ret_dict[kk].reshape([out_shape[0], out_shape[1], out_shape2])
+            tmp_arr = xp.where(atom_mask[:, :, None], tmp_arr, xp.zeros_like(tmp_arr))
+            ret_dict[kk] = xp.reshape(tmp_arr, out_shape)
+        ret_dict["mask"] = xp.astype(atom_mask, xp.int32)
 
         return ret_dict
 

deepmd/dpmodel/common.py

@@ -3,9 +3,14 @@
     ABC,
     abstractmethod,
 )
+from functools import (
+    wraps,
+)
 from typing import (
     Any,
+    Callable,
     Optional,
+    overload,
 )
 
 import array_api_compat
@@ -116,6 +121,105 @@
         return np.from_dlpack(x)
 
 
+def cast_precision(func: Callable[..., Any]) -> Callable[..., Any]:
+    """A decorator that casts and casts back the input
+    and output tensor of a method.
+
+    The decorator should be used on an instance method.
+
+    The decorator will do the following thing:
+    (1) It casts input arrays from the global precision
+    to precision defined by property `precision`.
+    (2) It casts output arrays from `precision` to
+    the global precision.
+    (3) It checks inputs and outputs and only casts when
+    input or output is an array and its dtype matches
+    the global precision and `precision`, respectively.
+    If it does not match (e.g. it is an integer), the decorator
+    will do nothing on it.
+
+    The decorator supports the array API.
+
+    Returns
+    -------
+    Callable
+        a decorator that casts and casts back the input and
+        output array of a method
+
+    Examples
+    --------
+    >>> class A:
+    ...     def __init__(self):
+    ...         self.precision = "float32"
+    ...
+    ...     @cast_precision
+    ...     def f(x: Array, y: Array) -> Array:
+    ...         return x**2 + y
+    """
+
+    @wraps(func)
+    def wrapper(self, *args, **kwargs):
+        # only convert tensors
+        returned_tensor = func(
+            self,
+            *[safe_cast_array(vv, "global", self.precision) for vv in args],
+            **{
+                kk: safe_cast_array(vv, "global", self.precision)
+                for kk, vv in kwargs.items()
+            },
+        )
+        if isinstance(returned_tensor, tuple):
+            return tuple(
+                safe_cast_array(vv, self.precision, "global") for vv in returned_tensor
+            )
+        elif isinstance(returned_tensor, dict):
+            return {
+                kk: safe_cast_array(vv, self.precision, "global")
+                for kk, vv in returned_tensor.items()
+            }
+        else:
+            return safe_cast_array(returned_tensor, self.precision, "global")
+
+    return wrapper
+
+
+@overload
+def safe_cast_array(
+    input: np.ndarray, from_precision: str, to_precision: str
+) -> np.ndarray: ...
+@overload
+def safe_cast_array(input: None, from_precision: str, to_precision: str) -> None: ...
+def safe_cast_array(
+    input: Optional[np.ndarray], from_precision: str, to_precision: str
+) -> Optional[np.ndarray]:
+    """Convert an array from a precision to another precision.
+
+    If input is not an array or without the specific precision, the method will not
+    cast it.
+
+    Array API is supported.
+
+    Parameters
+    ----------
+    input : np.ndarray or None
+        Input array
+    from_precision : str
+        Array data type that is casted from
+    to_precision : str
+        Array data type that casts to
+
+    Returns
+    -------
+    np.ndarray or None
+        casted array
+    """
+    if array_api_compat.is_array_api_obj(input):
+        xp = array_api_compat.array_namespace(input)
+        if input.dtype == get_xp_precision(xp, from_precision):
+            return xp.astype(input, get_xp_precision(xp, to_precision))
+    return input
+
+
 __all__ = [
     "GLOBAL_NP_FLOAT_PRECISION",
     "GLOBAL_ENER_FLOAT_PRECISION",

deepmd/dpmodel/descriptor/dpa1.py

@@ -19,6 +19,7 @@
     xp_take_along_axis,
 )
 from deepmd.dpmodel.common import (
+    cast_precision,
     to_numpy_array,
 )
 from deepmd.dpmodel.utils import (
@@ -329,6 +330,7 @@ def __init__(
         self.tebd_dim = tebd_dim
         self.concat_output_tebd = concat_output_tebd
         self.trainable = trainable
+        self.precision = precision
 
     def get_rcut(self) -> float:
         """Returns the cut-off radius."""
@@ -448,6 +450,7 @@ def change_type_map(
         obj["davg"] = obj["davg"][remap_index]
         obj["dstd"] = obj["dstd"][remap_index]
 
+    @cast_precision
     def call(
         self,
         coord_ext,

deepmd/dpmodel/descriptor/dpa2.py

@@ -14,6 +14,7 @@
     xp_take_along_axis,
 )
 from deepmd.dpmodel.common import (
+    cast_precision,
     to_numpy_array,
 )
 from deepmd.dpmodel.utils import (
@@ -594,6 +595,7 @@ def init_subclass_params(sub_data, sub_class):
         self.rcut = self.repinit.get_rcut()
         self.ntypes = ntypes
         self.sel = self.repinit.sel
+        self.precision = precision
 
     def get_rcut(self) -> float:
         """Returns the cut-off radius."""
@@ -757,6 +759,7 @@ def get_stat_mean_and_stddev(self) -> tuple[list[np.ndarray], list[np.ndarray]]:
             stddev_list.append(self.repinit_three_body.stddev)
         return mean_list, stddev_list
 
+    @cast_precision
     def call(
         self,
         coord_ext: np.ndarray,

deepmd/dpmodel/descriptor/se_e2_a.py

@@ -15,6 +15,7 @@
     NativeOP,
 )
 from deepmd.dpmodel.common import (
+    cast_precision,
     to_numpy_array,
 )
 from deepmd.dpmodel.utils import (
@@ -29,9 +30,6 @@
 from deepmd.dpmodel.utils.update_sel import (
     UpdateSel,
 )
-from deepmd.env import (
-    GLOBAL_NP_FLOAT_PRECISION,
-)
 from deepmd.utils.data_system import (
     DeepmdDataSystem,
 )
@@ -340,6 +338,7 @@ def reinit_exclude(
         self.exclude_types = exclude_types
         self.emask = PairExcludeMask(self.ntypes, exclude_types=exclude_types)
 
+    @cast_precision
     def call(
         self,
         coord_ext,
@@ -415,9 +414,7 @@ def call(
         # nf x nloc x ng x ng1
         grrg = np.einsum("flid,fljd->flij", gr, gr1)
         # nf x nloc x (ng x ng1)
-        grrg = grrg.reshape(nf, nloc, ng * self.axis_neuron).astype(
-            GLOBAL_NP_FLOAT_PRECISION
-        )
+        grrg = grrg.reshape(nf, nloc, ng * self.axis_neuron)
         return grrg, gr[..., 1:], None, None, ww
 
     def serialize(self) -> dict:
@@ -506,6 +503,7 @@ def update_sel(
 
 
 class DescrptSeAArrayAPI(DescrptSeA):
+    @cast_precision
     def call(
         self,
         coord_ext,
@@ -585,7 +583,5 @@ def call(
         # grrg = xp.einsum("flid,fljd->flij", gr, gr1)
         grrg = xp.sum(gr[:, :, :, None, :] * gr1[:, :, None, :, :], axis=4)
         # nf x nloc x (ng x ng1)
-        grrg = xp.astype(
-            xp.reshape(grrg, (nf, nloc, ng * self.axis_neuron)), input_dtype
-        )
+        grrg = xp.reshape(grrg, (nf, nloc, ng * self.axis_neuron))
         return grrg, gr[..., 1:], None, None, ww

deepmd/dpmodel/descriptor/se_r.py

@@ -14,6 +14,7 @@
     NativeOP,
 )
 from deepmd.dpmodel.common import (
+    cast_precision,
     get_xp_precision,
     to_numpy_array,
 )
@@ -289,6 +290,7 @@ def cal_g(
         gg = self.embeddings[(ll,)].call(ss)
         return gg
 
+    @cast_precision
     def call(
         self,
         coord_ext,
@@ -352,7 +354,6 @@ def call(
         res_rescale = 1.0 / 5.0
         res = xyz_scatter * res_rescale
         res = xp.reshape(res, (nf, nloc, ng))
-        res = xp.astype(res, get_xp_precision(xp, "global"))
         return res, None, None, None, ww
 
     def serialize(self) -> dict:

deepmd/dpmodel/descriptor/se_t.py

@@ -14,6 +14,7 @@
     NativeOP,
 )
 from deepmd.dpmodel.common import (
+    cast_precision,
     get_xp_precision,
     to_numpy_array,
 )
@@ -264,6 +265,7 @@ def reinit_exclude(
         self.exclude_types = exclude_types
         self.emask = PairExcludeMask(self.ntypes, exclude_types=exclude_types)
 
+    @cast_precision
     def call(
         self,
         coord_ext,
@@ -317,7 +319,6 @@ def call(
         # we don't require atype is the same in all frames
         exclude_mask = xp.reshape(exclude_mask, (nf * nloc, nnei))
         rr = xp.reshape(rr, (nf * nloc, nnei, 4))
-        rr = xp.astype(rr, get_xp_precision(xp, self.precision))
 
         for embedding_idx in itertools.product(
             range(self.ntypes), repeat=self.embeddings.ndim
@@ -349,7 +350,6 @@ def call(
                 result += res_ij
         # nf x nloc x ng
         result = xp.reshape(result, (nf, nloc, ng))
-        result = xp.astype(result, get_xp_precision(xp, "global"))
         return result, None, None, None, ww
 
     def serialize(self) -> dict:

deepmd/dpmodel/descriptor/se_t_tebd.py

@@ -16,7 +16,7 @@
     xp_take_along_axis,
 )
 from deepmd.dpmodel.common import (
-    get_xp_precision,
+    cast_precision,
     to_numpy_array,
 )
 from deepmd.dpmodel.utils import (
@@ -168,6 +168,7 @@ def __init__(
         self.tebd_dim = tebd_dim
         self.concat_output_tebd = concat_output_tebd
         self.trainable = trainable
+        self.precision = precision
 
     def get_rcut(self) -> float:
         """Returns the cut-off radius."""
@@ -287,6 +288,7 @@ def change_type_map(
         obj["davg"] = obj["davg"][remap_index]
         obj["dstd"] = obj["dstd"][remap_index]
 
+    @cast_precision
     def call(
         self,
         coord_ext,
@@ -741,7 +743,6 @@ def call(
         res_ij = res_ij * (1.0 / float(self.nnei) / float(self.nnei))
         # nf x nl x ng
         result = xp.reshape(res_ij, (nf, nloc, self.filter_neuron[-1]))
-        result = xp.astype(result, get_xp_precision(xp, "global"))
         return (
             result,
             None,

deepmd/dpmodel/fitting/dipole_fitting.py

@@ -11,6 +11,9 @@
 from deepmd.dpmodel import (
     DEFAULT_PRECISION,
 )
+from deepmd.dpmodel.common import (
+    cast_precision,
+)
 from deepmd.dpmodel.fitting.base_fitting import (
     BaseFitting,
 )
@@ -174,6 +177,7 @@ def output_def(self):
             ]
         )
 
+    @cast_precision
     def call(
         self,
         descriptor: np.ndarray,

deepmd/dpmodel/fitting/general_fitting.py

@@ -439,18 +439,22 @@ def _call_common(
                 ):
                     assert xx_zeros is not None
                     atom_property -= self.nets[(type_i,)](xx_zeros)
-                atom_property = atom_property + self.bias_atom_e[type_i, ...]
-                atom_property = atom_property * xp.astype(mask, atom_property.dtype)
+                atom_property = xp.where(
+                    mask, atom_property, xp.zeros_like(atom_property)
+                )
                 outs = outs + atom_property  # Shape is [nframes, natoms[0], 1]
         else:
-            outs = self.nets[()](xx) + xp.reshape(
-                xp.take(self.bias_atom_e, xp.reshape(atype, [-1]), axis=0),
-                [nf, nloc, net_dim_out],
-            )
+            outs = self.nets[()](xx)
             if xx_zeros is not None:
                 outs -= self.nets[()](xx_zeros)
+        outs += xp.reshape(
+            xp.take(
+                xp.astype(self.bias_atom_e, outs.dtype), xp.reshape(atype, [-1]), axis=0
+            ),
+            [nf, nloc, net_dim_out],
+        )
         # nf x nloc
         exclude_mask = self.emask.build_type_exclude_mask(atype)
         # nf x nloc x nod
-        outs = outs * xp.astype(exclude_mask[:, :, None], outs.dtype)
-        return {self.var_name: xp.astype(outs, get_xp_precision(xp, "global"))}
+        outs = xp.where(exclude_mask[:, :, None], outs, xp.zeros_like(outs))
+        return {self.var_name: outs}