fix(tf): validate flattened multi-device op widths #5827
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -79,8 +79,12 @@ class ProdVirialSeAGradOp : public OpKernel { | |
|
|
||
| int nframes = net_deriv_tensor.shape().dim_size(0); | ||
| int nloc = natoms(0); | ||
| int ndescrpt; | ||
| int nnei; | ||
| OP_REQUIRES_OK(context, deepmd::tf_compat::GetPerAtomWidth( | ||
| &ndescrpt, net_deriv_shape, nloc, "net deriv")); | ||
| OP_REQUIRES_OK(context, deepmd::tf_compat::GetPerAtomWidth( | ||
| &nnei, nlist_shape, nloc, "nlist")); | ||
|
|
||
| // check the sizes | ||
| OP_REQUIRES( | ||
|
|
@@ -109,6 +113,9 @@ class ProdVirialSeAGradOp : public OpKernel { | |
| OP_REQUIRES( | ||
| context, (nnei == n_a_sel + n_r_sel), | ||
| deepmd::tf_compat::InvalidArgument("number of neighbors should match")); | ||
| OP_REQUIRES(context, (static_cast<int64_t>(nnei) * 4 == ndescrpt), | ||
| deepmd::tf_compat::InvalidArgument( | ||
| "descriptor width should be four times neighbor width")); | ||
|
|
||
| // Create an output tensor | ||
| TensorShape grad_net_shape; | ||
|
|
@@ -223,8 +230,12 @@ class ProdVirialSeRGradOp : public OpKernel { | |
|
|
||
| int nframes = net_deriv_tensor.shape().dim_size(0); | ||
| int nloc = natoms(0); | ||
| int ndescrpt; | ||
| int nnei; | ||
| OP_REQUIRES_OK(context, deepmd::tf_compat::GetPerAtomWidth( | ||
| &ndescrpt, net_deriv_shape, nloc, "net deriv")); | ||
| OP_REQUIRES_OK(context, deepmd::tf_compat::GetPerAtomWidth( | ||
| &nnei, nlist_shape, nloc, "nlist")); | ||
|
|
||
| // check the sizes | ||
| OP_REQUIRES( | ||
|
|
@@ -250,6 +261,9 @@ class ProdVirialSeRGradOp : public OpKernel { | |
| OP_REQUIRES( | ||
| context, (int_64(nloc) * nnei * 3 == rij_shape.dim_size(1)), | ||
| deepmd::tf_compat::InvalidArgument("dim of rij should be nnei * 3")); | ||
| OP_REQUIRES(context, (nnei == ndescrpt), | ||
| deepmd::tf_compat::InvalidArgument( | ||
| "descriptor width should equal neighbor width")); | ||
|
|
||
| // Create an output tensor | ||
| TensorShape grad_net_shape; | ||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,134 @@ | ||
| # SPDX-License-Identifier: LGPL-3.0-or-later | ||
| """Regression tests for flattened TensorFlow custom-op input dimensions.""" | ||
|
|
||
| from deepmd.tf.env import ( | ||
| GLOBAL_TF_FLOAT_PRECISION, | ||
| op_grads_module, | ||
| op_module, | ||
| tf, | ||
| ) | ||
|
|
||
|
|
||
| class TestMultiDeviceShapeValidation(tf.test.TestCase): | ||
| """Ensure malformed flattened widths fail before native kernel dispatch.""" | ||
|
|
||
| def setUp(self) -> None: | ||
| self.sess = self.cached_session().__enter__() | ||
| self.nloc = 2 | ||
| self.nnei = 1 | ||
| self.ndescrpt = 4 | ||
| self.natoms = tf.constant([self.nloc, self.nloc, 1], dtype=tf.int32) | ||
|
|
||
| def _floats(self, width: int): | ||
| """Create one frame of flattened floating-point custom-op input.""" | ||
| return tf.zeros([1, width], dtype=GLOBAL_TF_FLOAT_PRECISION) | ||
|
|
||
| def _nlist(self, width: int): | ||
| """Create one frame of flattened neighbor indices.""" | ||
| return tf.zeros([1, width], dtype=tf.int32) | ||
|
|
||
| def test_prod_force_rejects_partial_net_deriv_atom(self) -> None: | ||
| # The old integer division truncated 9 / 2 to four descriptors and | ||
| # allowed the extra value to survive until raw pointer dispatch. | ||
| with self.assertRaisesRegex( | ||
| tf.errors.InvalidArgumentError, | ||
| r"net deriv width 9 should be divisible by nloc 2", | ||
| ): | ||
| self.sess.run( | ||
| op_module.prod_force_se_a( | ||
| self._floats(self.nloc * self.ndescrpt + 1), | ||
| self._floats(self.nloc * self.ndescrpt * 3), | ||
| self._nlist(self.nloc * self.nnei), | ||
| self.natoms, | ||
| n_a_sel=self.nnei, | ||
| n_r_sel=0, | ||
| ) | ||
| ) | ||
|
|
||
| def test_prod_force_rejects_in_deriv_width_mismatch(self) -> None: | ||
| with self.assertRaisesRegex( | ||
| tf.errors.InvalidArgumentError, r"number of descriptors should match" | ||
| ): | ||
| self.sess.run( | ||
| op_module.prod_force_se_a( | ||
| self._floats(self.nloc * self.ndescrpt), | ||
| self._floats(self.nloc * self.ndescrpt * 3 - 1), | ||
| self._nlist(self.nloc * self.nnei), | ||
| self.natoms, | ||
| n_a_sel=self.nnei, | ||
| n_r_sel=0, | ||
| ) | ||
| ) | ||
|
|
||
| def test_prod_force_r_rejects_descriptor_stride_mismatch(self) -> None: | ||
| with self.assertRaisesRegex( | ||
| tf.errors.InvalidArgumentError, | ||
| r"descriptor width should equal neighbor width", | ||
| ): | ||
| self.sess.run( | ||
| op_module.prod_force_se_r( | ||
| self._floats(self.nloc * (self.nnei + 1)), | ||
| self._floats(self.nloc * (self.nnei + 1) * 3), | ||
| self._nlist(self.nloc * self.nnei), | ||
| self.natoms, | ||
| ) | ||
| ) | ||
|
|
||
| def test_prod_force_grad_rejects_partial_nlist_atom(self) -> None: | ||
| # Fixed-width placeholders in the original tests rejected this feed | ||
| # before the custom op ran, leaving its release-build checks untested. | ||
| with self.assertRaisesRegex( | ||
| tf.errors.InvalidArgumentError, | ||
| r"nlist width 3 should be divisible by nloc 2", | ||
| ): | ||
| self.sess.run( | ||
| op_grads_module.prod_force_se_a_grad( | ||
| self._floats(self.nloc * 3), | ||
| self._floats(self.nloc * self.ndescrpt), | ||
| self._floats(self.nloc * self.ndescrpt * 3), | ||
| self._nlist(self.nloc * self.nnei + 1), | ||
| self.natoms, | ||
| n_a_sel=self.nnei, | ||
| n_r_sel=0, | ||
| ) | ||
| ) | ||
|
|
||
| def test_prod_virial_grad_rejects_descriptor_stride_mismatch(self) -> None: | ||
| mismatched_ndescrpt = self.ndescrpt * 2 | ||
| with self.assertRaisesRegex( | ||
| tf.errors.InvalidArgumentError, | ||
| r"descriptor width should be four times neighbor width", | ||
| ): | ||
| self.sess.run( | ||
| op_grads_module.prod_virial_se_a_grad( | ||
| self._floats(9), | ||
| self._floats(self.nloc * mismatched_ndescrpt), | ||
| self._floats(self.nloc * mismatched_ndescrpt * 3), | ||
| self._floats(self.nloc * self.nnei * 3), | ||
| self._nlist(self.nloc * self.nnei), | ||
| self.natoms, | ||
| n_a_sel=self.nnei, | ||
| n_r_sel=0, | ||
| ) | ||
| ) | ||
|
|
||
| def test_prod_virial_grad_rejects_rij_width_mismatch(self) -> None: | ||
| with self.assertRaisesRegex( | ||
| tf.errors.InvalidArgumentError, r"dim of rij should be nnei \* 3" | ||
| ): | ||
| self.sess.run( | ||
| op_grads_module.prod_virial_se_a_grad( | ||
| self._floats(9), | ||
| self._floats(self.nloc * self.ndescrpt), | ||
| self._floats(self.nloc * self.ndescrpt * 3), | ||
| self._floats(self.nloc * self.nnei * 3 - 1), | ||
| self._nlist(self.nloc * self.nnei), | ||
| self.natoms, | ||
| n_a_sel=self.nnei, | ||
| n_r_sel=0, | ||
| ) | ||
| ) | ||
|
|
||
|
|
||
| if __name__ == "__main__": | ||
| tf.test.main() |
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
[P1] Apply this helper and the descriptor-to-neighbor stride checks to the forward
ProdVirialSeA/ProdVirialSeRoperators too. They still use truncatingdim_size(1) / nlocinprod_virial_multi_device.ccand never enforce SE-Andescrpt == 4 * nneior SE-R equality. A malformed input can therefore pass the wrapper while the native kernel derives its own wider descriptor stride fromnneiand reads pastnet_deriv/in_deriv. This is the same raw-buffer vulnerability class fixed here for force and virial-grad, and it needs a matching forward-virial regression test.