- 20 2月, 2021 24 次提交
-
-
由 Prakalp Srivastava 提交于
`mhlo.dot_general` and `mhlo.convolution` result element type might be different from operand element type. See `preferred_element_type` attribute that allows i8xi8 to i32 dot computation. `mhlo` to HLO exporter should pass the result element type to Xla builder to override the shape inference of XLA. PiperOrigin-RevId: 358580718 Change-Id: If3ad34b6824a52498663f0a1a031a5bdc29a24ee
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358555614 Change-Id: I601a64ef0cdb53564397e9644f529a92132497c2
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358555610 Change-Id: I1ef2f9fc4612d5596a5f3942817298447d9c0258
-
由 Haoliang Zhang 提交于
Use scoped diagnostic handler to capture diagnostics within the `lower_static_tensor_list` pass when `allow_tensorlist_pass_through=true`. PiperOrigin-RevId: 358555572 Change-Id: I3545278631d992e4eb8640d7cbde3d7736c012a3
-
由 Jaesung Chung 提交于
* Sharing TensorFlow Resource and Variant Tensors across multi-subgraphs The existing Flex delegate handles the cases for the multi-partitions in a subgraph well but when sharing tensors across subgraphs, we need to be careful to implement since resource and variant tensors are RAII objects, which means that the memcpy does not work on them. Other value-based tensor formats, numbers and strings, can be easily a format of memory data and easily copied around subgraphs by just copying data. Sharing the pointer of resource and variant tensor should be done instead. * Sharing TensorFlow tensor pointer across multi-subgraphs To do that, Flex resource and variant buffer format is just a container for storing the TensorFlow tensor object, which refers to TensorFlow resource or variant tensor like the below code snippet. struct OpaqueBuffer { // Store a TensorFlow's tensor pointer. The life cycle of the pointer will be // managed by the reference counting in the TensorFlow world and the pointer // will be freed when all the buffer maps, who own it, are gone. const tensorflow::Tensor* tf_tensor; }; * Life cycle of the shared TensorFlow tensor object When TFLite runtime requires those tensors to be transferred to the other subgraphs, TFLite runtime invokes the delegate interface, the CopyFromBufferHandle method. The Flex’s CopyFromBufferHandle method will create the above OpaqueBuffer structure and store the corresponding TensorFlow tensor pointer. Later, the other subgraphs will find the TensorFlow pointer inside and insert it to their buffer map to track the life cycle in the subgraph. Even though the same TensorFlow tensor object can be found in the multiple buffer maps in the Flex delegate. They will be freed when the Flex delegate is finalized. PiperOrigin-RevId: 358553465 Change-Id: I7cb056bf8f216851c771d7b2f26e69821a7cbf6a
-
由 A. Unique TensorFlower 提交于
Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/46551 This PR adds cudaMallocAsync as an option when CUDA 11.2 is used. PiperOrigin-RevId: 358553125 Change-Id: Id7110f54838fafb4107f06ed1d68ce0245010a3a
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 358545703 Change-Id: I46aa9403023db7e1024350593d48d18bd30a2a2e
-
由 Frank Chen 提交于
PiperOrigin-RevId: 358540864 Change-Id: Ie5055e71a7cceffe3534be769cd8d5e23f18cae2
-
由 Tomer Kaftan 提交于
Reduce the number of private symbols Keras relies on, by switching specific usages of Trackable data structures to the generic `wrap_or_unwrap` method that decides automatically according to the input type. PiperOrigin-RevId: 358532293 Change-Id: Id85a1082f1a9bfc9b6e025c57fb078e960e0db8b
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358531745 Change-Id: If7efba60fad4f3b7249408fab01e31571c30e80c
-
由 George Karpenkov 提交于
PiperOrigin-RevId: 358529494 Change-Id: Ida2c931e1f7374711dc200c505e9dd6ff6f96282
-
由 Peng Wang 提交于
PiperOrigin-RevId: 358529149 Change-Id: I533f9bce385480cdbd17492cf21130fab5d00328
-
由 Ruoxin Sang 提交于
Add an `experimental_per_replica_buffer_size` option to `tf.distribute.InputOptions`. This allows users to configure the replica device buffer size. Also make sure InputOptions is propagated correctly in each strategy. PiperOrigin-RevId: 358526640 Change-Id: Ib14a0fc33b60c37fb55983fed87b65fccf83bafd
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358526294 Change-Id: I1915c2629d51cf761ea46bfb6a7b2d01b7e9c125
-
由 Jay Shi 提交于
[tf.data] Allow ParallelBatchDataset to produce elements in non-deterministic order. When the pipeline doesn't require deterministic order of outputs from ParallelBatchDataset, if processing time for different batches varies, setting `deterministic` to be False will fetch the fast produced elements first and may benefit the users. PiperOrigin-RevId: 358526042 Change-Id: I02ae804ac1454377a09545b68bcb5c68ebc10117
-
由 Robert David 提交于
This relies on constant folding of FC to generate the new bias constant. PiperOrigin-RevId: 358524208 Change-Id: If879b81b0e871456df4ba0b59672248e11ae2830
-
由 Jacques Pienaar 提交于
PiperOrigin-RevId: 358524205 Change-Id: I01e1988a837c2d861fe1485c6bf091661b8b42cf
-
由 Jaesung Chung 提交于
PiperOrigin-RevId: 358523401 Change-Id: I8f1ef7500967edec657038bcc3bba15032c4fed4
-
由 A. Unique TensorFlower 提交于
Updates LLVM usage to match [7e54d7304be2](https://github.com/llvm/llvm-project/commit/7e54d7304be2) PiperOrigin-RevId: 358516349 Change-Id: I074a0977852d7c8f0502ad31160fb3fe0957817e
-
由 Yi Situ 提交于
Put definitions that are only used locally within compilation unit behind an anonymous namespace to avoid an ODR. PiperOrigin-RevId: 358516197 Change-Id: I9c8e657d72e618b527b41fdea2b298a31791ccc6
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358447444 Change-Id: I949c08d8aa4da929667b42766a9289a549c87ff8
-
由 Mingming Liu 提交于
PiperOrigin-RevId: 358445685 Change-Id: Idcff7f096dc331f023d6b5d4cce2d9e453091141
-
由 Jiri Simsa 提交于
Making sure that recently introduced `offset` argument for the `InitializeTableFromTextFile` op is handled in a backwards / forwards compatible manner. PiperOrigin-RevId: 358416520 Change-Id: I8b9637c5d9707097738de134e64b2cb72c73022a
-
由 Scott Main 提交于
Update Python guide to install via Debian package to avoid future issues when trying to use Coral libraries. (Debian package installs are incompatible with pip installs.) PiperOrigin-RevId: 358414526 Change-Id: I769d7e178e490fe9d86abe3194e9bfa8778350f6
-
- 19 2月, 2021 16 次提交
-
-
由 A. Unique TensorFlower 提交于
This is necessary in some circumstances when building internally, and is harmless otherwise. PiperOrigin-RevId: 358372796 Change-Id: Ib53a5bc1e45acb7fb3fceee91ea25fe7e69283c4
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358370930 Change-Id: I1890368dec43716c2bbcbdc27b60a13e39903d06
-
由 Stephan Herhut 提交于
PiperOrigin-RevId: 358370348 Change-Id: I034141783fe054ff7b78233dc1fe653abc77b16f
-
由 Sanjoy Das 提交于
PiperOrigin-RevId: 358360541 Change-Id: Ia08a289c1368bfb2c8b726e707c8bb94a208d224
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358355580 Change-Id: I954937e74ee84c8d0105319e954119c4c560e92e
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 358355577 Change-Id: I786507263a90b6fc81c9473d17365b8b8ac737bc
-
由 Yujing Zhang 提交于
PiperOrigin-RevId: 358344741 Change-Id: I41b0536dfbde6168085477306cab6a8b7c7f65a6
-
由 Khanh LeViet 提交于
PiperOrigin-RevId: 358338694 Change-Id: I2f6538530d2cacc9ee56ceba8a662df830e145d2
-
由 Yunlu Li 提交于
PiperOrigin-RevId: 358331167 Change-Id: I8d4e850023c3c2c6784cf3329d19ad3b9c25948d
-
由 A. Unique TensorFlower 提交于
Updates LLVM usage to match [cd8cc00b9e2b](https://github.com/llvm/llvm-project/commit/cd8cc00b9e2b) PiperOrigin-RevId: 358329421 Change-Id: I56c409cb93a70dc0b1643913813ecb935e500603
-
由 David Rim 提交于
PiperOrigin-RevId: 358327761 Change-Id: I1b4c22a0b4b6f01e70f359c2207f925ea77448d6
-
由 David Rim 提交于
PiperOrigin-RevId: 358324561 Change-Id: I2ba23bf11c7b200e49cee1cdff096c3521f12e51
-
由 Taehee Jeong 提交于
PiperOrigin-RevId: 358321694 Change-Id: Ie95ae5701320f5ee7361ca306650603479c92a65
-
由 Peter Hawkins 提交于
Switch tensorflow::Subprocess to use posix_spawnp() instead of fork()/execvp() on non-Android POSIX platforms. The goal of this change is to avoid calling pthread_atfork() handlers. Some libraries, in particular the version of OpenBLAS included in NumPy, have buggy pthread_atfork() handlers. See https://github.com/xianyi/OpenBLAS/pull/3111 and https://github.com/google/jax/issues/5713 for details. Now, while we can and have fixed the buggy atfork handlers, it will take some time for the fix to be deployed in a NumPy release and for users to update to a new NumPy release. So we also take an additional step: avoid running atfork handlers in Subprocess. My copy of the glibc documentation says: " According to POSIX, it unspecified whether fork handlers established with pthread_atfork(3) are called when posix_spawn() is invoked. On glibc, fork handlers are called only if the child is created using fork(2). " It appears glibc 2.24 and newer do not call pthread_atfork() handlers from posix_spawn(). Using posix_spawn() should be at least no worse than an explicit fork()/execvp() pair, and on glibc it should do the right thing. PiperOrigin-RevId: 358317859 Change-Id: Ic1d95446706efa7c0db4e79bf8281f14b2bd99df
-
由 Yi Situ 提交于
PiperOrigin-RevId: 358313521 Change-Id: I9100dbf8d82dc695a896a7b61c433ef38865ec9d
-
由 Yujing Zhang 提交于
- Introduce a CollectiveOpGroupMode enum to describe various modes of interpreting replica groups attached to collective communication operations and GetCollectiveOpGroupMode() function to get the group formation mode implied by an HLO collective op based on whether it has channel_id and use_global_device_ids. - Fix GetParticipatingDevices() to use this mode to correctly find participants in all group formatio... PiperOrigin-RevId: 358313513 Change-Id: I9103d065c0a89149bcad3fa6b275696a257c97ae
-