feat(jax): build nlist in the SavedModel & fix nopbc for StableHLO and SavedModel

.github/workflows/test_python.yml

@@ -58,7 +58,7 @@ jobs:
       env:
         NUM_WORKERS: 0
     - name: Test TF2 eager mode
-      run: pytest --cov=deepmd source/tests/consistent/io/test_io.py --durations=0
+      run: pytest --cov=deepmd --cov-append source/tests/consistent/io/test_io.py source/tests/jax/jax2tf --durations=0
       env:
         NUM_WORKERS: 0
         DP_TEST_TF2_ONLY: 1

deepmd/jax/infer/deep_eval.py

@@ -97,6 +97,12 @@ def __init__(
                 stablehlo_atomic_virial=model_data["@variables"][
                     "stablehlo_atomic_virial"
                 ].tobytes(),
+                stablehlo_no_ghost=model_data["@variables"][
+                    "stablehlo_no_ghost"
+                ].tobytes(),
+                stablehlo_atomic_virial_no_ghost=model_data["@variables"][
+                    "stablehlo_atomic_virial_no_ghost"
+                ].tobytes(),
                 model_def_script=model_data["model_def_script"],
                 **model_data["constants"],
             )

deepmd/jax/jax2tf/__init__.py

@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: LGPL-3.0-or-later
 import tensorflow as tf
+import tensorflow.experimental.numpy as tnp
 
 if not tf.executing_eagerly():
     # TF disallow temporary eager execution
@@ -9,3 +10,5 @@
         "If you are converting a model between different backends, "
         "considering converting to the `.dp` format first."
     )
+
+tnp.experimental_enable_numpy_behavior()

deepmd/jax/jax2tf/make_model.py

@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: LGPL-3.0-or-later
+from typing import (
+    Callable,
+)
+
+import tensorflow as tf
+import tensorflow.experimental.numpy as tnp
+
+from deepmd.dpmodel.output_def import (
+    ModelOutputDef,
+)
+from deepmd.jax.jax2tf.nlist import (
+    build_neighbor_list,
+    extend_coord_with_ghosts,
+)
+from deepmd.jax.jax2tf.region import (
+    normalize_coord,
+)
+from deepmd.jax.jax2tf.transform_output import (
+    communicate_extended_output,
+)
+
+
+def model_call_from_call_lower(
+    *,  # enforce keyword-only arguments
+    call_lower: Callable[
+        [
+            tnp.ndarray,
+            tnp.ndarray,
+            tnp.ndarray,
+            tnp.ndarray,
+            tnp.ndarray,
+            bool,
+        ],
+        dict[str, tnp.ndarray],
+    ],
+    rcut: float,
+    sel: list[int],
+    mixed_types: bool,
+    model_output_def: ModelOutputDef,
+    coord: tnp.ndarray,
+    atype: tnp.ndarray,
+    box: tnp.ndarray,
+    fparam: tnp.ndarray,
+    aparam: tnp.ndarray,
+    do_atomic_virial: bool = False,
+):
+    """Return model prediction from lower interface.
+
+    Parameters
+    ----------
+    coord
+        The coordinates of the atoms.
+        shape: nf x (nloc x 3)
+    atype
+        The type of atoms. shape: nf x nloc
+    box
+        The simulation box. shape: nf x 9
+    fparam
+        frame parameter. nf x ndf
+    aparam
+        atomic parameter. nf x nloc x nda
+    do_atomic_virial
+        If calculate the atomic virial.
+
+    Returns
+    -------
+    ret_dict
+        The result dict of type dict[str,tnp.ndarray].
+        The keys are defined by the `ModelOutputDef`.
+
+    """
+    atype_shape = tf.shape(atype)
+    nframes, nloc = atype_shape[0], atype_shape[1]
+    cc, bb, fp, ap = coord, box, fparam, aparam
+    del coord, box, fparam, aparam
+    if tf.shape(bb)[-1] != 0:
+        coord_normalized = normalize_coord(
+            cc.reshape(nframes, nloc, 3),
+            bb.reshape(nframes, 3, 3),
+        )
+    else:
+        coord_normalized = cc
+    extended_coord, extended_atype, mapping = extend_coord_with_ghosts(
+        coord_normalized, atype, bb, rcut
+    )
+    nlist = build_neighbor_list(
+        extended_coord,
+        extended_atype,
+        nloc,
+        rcut,
+        sel,
+        distinguish_types=not mixed_types,
+    )
+    extended_coord = extended_coord.reshape(nframes, -1, 3)
+    model_predict_lower = call_lower(
+        extended_coord,
+        extended_atype,
+        nlist,
+        mapping,
+        fparam=fp,
+        aparam=ap,
+    )
+    model_predict = communicate_extended_output(
+        model_predict_lower,
+        model_output_def,
+        mapping,
+        do_atomic_virial=do_atomic_virial,
+    )
+    return model_predict

deepmd/jax/jax2tf/nlist.py

@@ -0,0 +1,217 @@
+# SPDX-License-Identifier: LGPL-3.0-or-later
+from typing import (
+    Union,
+)
+
+import tensorflow as tf
+import tensorflow.experimental.numpy as tnp
+
+from .region import (
+    to_face_distance,
+)
+
+
+## translated from torch implementation by chatgpt
+def build_neighbor_list(
+    coord: tnp.ndarray,
+    atype: tnp.ndarray,
+    nloc: int,
+    rcut: float,
+    sel: Union[int, list[int]],
+    distinguish_types: bool = True,
+) -> tnp.ndarray:
+    """Build neighbor list for a single frame. keeps nsel neighbors.
+
+    Parameters
+    ----------
+    coord : tnp.ndarray
+        exptended coordinates of shape [batch_size, nall x 3]
+    atype : tnp.ndarray
+        extended atomic types of shape [batch_size, nall]
+        type < 0 the atom is treat as virtual atoms.
+    nloc : int
+        number of local atoms.
+    rcut : float
+        cut-off radius
+    sel : int or list[int]
+        maximal number of neighbors (of each type).
+        if distinguish_types==True, nsel should be list and
+        the length of nsel should be equal to number of
+        types.
+    distinguish_types : bool
+        distinguish different types.
+
+    Returns
+    -------
+    neighbor_list : tnp.ndarray
+        Neighbor list of shape [batch_size, nloc, nsel], the neighbors
+        are stored in an ascending order. If the number of
+        neighbors is less than nsel, the positions are masked
+        with -1. The neighbor list of an atom looks like
+        |------ nsel ------|
+        xx xx xx xx -1 -1 -1
+        if distinguish_types==True and we have two types
+        |---- nsel[0] -----| |---- nsel[1] -----|
+        xx xx xx xx -1 -1 -1 xx xx xx -1 -1 -1 -1
+        For virtual atoms all neighboring positions are filled with -1.
+
+    """
+    batch_size = tf.shape(coord)[0]
+    coord = tnp.reshape(coord, (batch_size, -1))
+    nall = tf.shape(coord)[1] // 3
+    # fill virtual atoms with large coords so they are not neighbors of any
+    # real atom.
+    if tf.size(coord) > 0:
+        xmax = tnp.max(coord) + 2.0 * rcut
+    else:
+        xmax = tf.cast(2.0 * rcut, coord.dtype)
+    # nf x nall
+    is_vir = atype < 0
+    coord1 = tnp.where(
+        is_vir[:, :, None], xmax, tnp.reshape(coord, (batch_size, nall, 3))
+    )
+    coord1 = tnp.reshape(coord1, (batch_size, nall * 3))
+    if isinstance(sel, int):
+        sel = [sel]
+    nsel = sum(sel)
+    coord0 = coord1[:, : nloc * 3]
+    diff = (
+        tnp.reshape(coord1, [batch_size, -1, 3])[:, None, :, :]
+        - tnp.reshape(coord0, [batch_size, -1, 3])[:, :, None, :]
+    )
+    rr = tf.linalg.norm(diff, axis=-1)
+    # if central atom has two zero distances, sorting sometimes can not exclude itself
+    rr -= tf.eye(nloc, nall, dtype=diff.dtype)[tnp.newaxis, :, :]
+    nlist = tnp.argsort(rr, axis=-1)
+    rr = tnp.sort(rr, axis=-1)
+    rr = rr[:, :, 1:]
+    nlist = nlist[:, :, 1:]
+    nnei = tf.shape(rr)[2]
+    if nsel <= nnei:
+        rr = rr[:, :, :nsel]
+        nlist = nlist[:, :, :nsel]
+    else:
+        rr = tnp.concatenate(
+            [rr, tnp.ones([batch_size, nloc, nsel - nnei], dtype=rr.dtype) + rcut],
+            axis=-1,
+        )
+        nlist = tnp.concatenate(
+            [nlist, tnp.ones([batch_size, nloc, nsel - nnei], dtype=nlist.dtype)],
+            axis=-1,
+        )
+    nlist = tnp.where(
+        tnp.logical_or((rr > rcut), is_vir[:, :nloc, None]),
+        tnp.full_like(nlist, -1),
+        nlist,
+    )
+
+    if distinguish_types:
+        return nlist_distinguish_types(nlist, atype, sel)
+    else:
+        return nlist
+
+
+def nlist_distinguish_types(
+    nlist: tnp.ndarray,
+    atype: tnp.ndarray,
+    sel: list[int],
+):
+    """Given a nlist that does not distinguish atom types, return a nlist that
+    distinguish atom types.
+
+    """
+    nloc = tf.shape(nlist)[1]
+    ret_nlist = []
+    tmp_atype = tnp.tile(atype[:, None, :], (1, nloc, 1))
+    mask = nlist == -1
+    tnlist_0 = tnp.where(mask, tnp.zeros_like(nlist), nlist)
+    tnlist = tnp.take_along_axis(tmp_atype, tnlist_0, axis=2)
+    tnlist = tnp.where(mask, tnp.full_like(tnlist, -1), tnlist)
+    for ii, ss in enumerate(sel):
+        pick_mask = tf.cast(tnlist == ii, tnp.int32)
+        sorted_indices = tnp.argsort(-pick_mask, kind="stable", axis=-1)
+        pick_mask_sorted = -tnp.sort(-pick_mask, axis=-1)
+        inlist = tnp.take_along_axis(nlist, sorted_indices, axis=2)
+        inlist = tnp.where(
+            ~tf.cast(pick_mask_sorted, tf.bool), tnp.full_like(inlist, -1), inlist
+        )
+        ret_nlist.append(inlist[..., :ss])
+    ret = tf.concat(ret_nlist, axis=-1)
+    return ret
+
+
+def tf_outer(a, b):
+    return tf.einsum("i,j->ij", a, b)
+
+
+## translated from torch implementation by chatgpt
+def extend_coord_with_ghosts(
+    coord: tnp.ndarray,
+    atype: tnp.ndarray,
+    cell: tnp.ndarray,
+    rcut: float,
+):
+    """Extend the coordinates of the atoms by appending peridoc images.
+    The number of images is large enough to ensure all the neighbors
+    within rcut are appended.
+
+    Parameters
+    ----------
+    coord : tnp.ndarray
+        original coordinates of shape [-1, nloc*3].
+    atype : tnp.ndarray
+        atom type of shape [-1, nloc].
+    cell : tnp.ndarray
+        simulation cell tensor of shape [-1, 9].
+    rcut : float
+        the cutoff radius
+
+    Returns
+    -------
+    extended_coord: tnp.ndarray
+        extended coordinates of shape [-1, nall*3].
+    extended_atype: tnp.ndarray
+        extended atom type of shape [-1, nall].
+    index_mapping: tnp.ndarray
+        mapping extended index to the local index
+
+    """
+    atype_shape = tf.shape(atype)
+    nf, nloc = atype_shape[0], atype_shape[1]
+    # int64 for index
+    aidx = tf.range(nloc, dtype=tnp.int64)
+    aidx = tnp.tile(aidx[tnp.newaxis, :], (nf, 1))
+    if tf.shape(cell)[-1] == 0:
+        nall = nloc
+        extend_coord = coord
+        extend_atype = atype
+        extend_aidx = aidx
+    else:
+        coord = tnp.reshape(coord, (nf, nloc, 3))
+        cell = tnp.reshape(cell, (nf, 3, 3))
+        to_face = to_face_distance(cell)
+        nbuff = tf.cast(tnp.ceil(rcut / to_face), tnp.int64)
+        nbuff = tnp.max(nbuff, axis=0)
+        xi = tf.range(-nbuff[0], nbuff[0] + 1, 1, dtype=tnp.int64)
+        yi = tf.range(-nbuff[1], nbuff[1] + 1, 1, dtype=tnp.int64)
+        zi = tf.range(-nbuff[2], nbuff[2] + 1, 1, dtype=tnp.int64)
+        xyz = tf_outer(xi, tnp.asarray([1, 0, 0]))[:, tnp.newaxis, tnp.newaxis, :]
+        xyz = xyz + tf_outer(yi, tnp.asarray([0, 1, 0]))[tnp.newaxis, :, tnp.newaxis, :]
+        xyz = xyz + tf_outer(zi, tnp.asarray([0, 0, 1]))[tnp.newaxis, tnp.newaxis, :, :]
+        xyz = tnp.reshape(xyz, (-1, 3))
+        xyz = tf.cast(xyz, coord.dtype)
+        shift_idx = tnp.take(xyz, tnp.argsort(tf.linalg.norm(xyz, axis=1)), axis=0)
+        ns = tf.shape(shift_idx)[0]
+        nall = ns * nloc
+        shift_vec = tnp.einsum("sd,fdk->fsk", shift_idx, cell)
+        # shift_vec = tnp.tensordot(shift_idx, cell, axes=([1], [1]))
+        # shift_vec = tnp.transpose(shift_vec, (1, 0, 2))
+        extend_coord = coord[:, None, :, :] + shift_vec[:, :, None, :]
+        extend_atype = tnp.tile(atype[:, :, tnp.newaxis], (1, ns, 1))
+        extend_aidx = tnp.tile(aidx[:, :, tnp.newaxis], (1, ns, 1))
+
+    return (
+        tnp.reshape(extend_coord, (nf, nall * 3)),
+        tnp.reshape(extend_atype, (nf, nall)),
+        tnp.reshape(extend_aidx, (nf, nall)),
+    )