hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
63c478552e153e12791d87d4ef5ae95513caca99 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/bf279def1a8ccbedfff4e139cfbe5894b65222cc.
PiperOrigin-RevId: 566574665 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,118,361,000 |
2f6c00fd5df210918d79ef8bf177f69fdcf6f038 | Internal Code Change
PiperOrigin-RevId: 566575065 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,118,484,000 |
ebb5828ec7a4abefff02e4413b5b8b2a2a4a783e | Make the MHLO dialect legal in the MHLO -> TFLite convertion pass.
PiperOrigin-RevId: 566593832 | Quentin Khan | qkhan@google.com | 1,695,124,327,000 |
ab6bbb417f5e336f5f0f4164ade5ac0c7555874c | No public description
PiperOrigin-RevId: 566595405 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,124,782,000 |
b9209839ea7d74a771bcb4dd1a70e2092c39cb55 | No public description
PiperOrigin-RevId: 566612185 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,129,821,000 |
f4529e80ab30a51207901b74b438980ac8b3ceaf | [XLA:GPU] Move buffer sharing logic into a separate target (NFC).
This will allow to use the same CanShareBuffer function in gpu_compiler and
hlo_to_llvm_ir testing tool.
PiperOrigin-RevId: 566615614 | Adrian Kuegel | akuegel@google.com | 1,695,130,818,000 |
cf3edc5aa450c8e7a95a93338e06a16e6d207e58 | nicer link to OpenSSF scorecard | Moritz Firsching | firsching@google.com | 1,695,132,664,000 |
e47a51b0d946ae6feb80f3cf027a565b5471de6a | [XLA:GPU] Use CanShareBufferHint() instead of dummy function.
CompileModuleToLlvmIr() is used by hlo_to_llvm_ir testing tool. Now it will
use the same buffer sharing logic as nvptx_compiler.cc. In an upcoming
change, this will make a difference.
PiperOrigin-RevId: 566622318 | Adrian Kuegel | akuegel@google.com | 1,695,132,719,000 |
d6261f100f2452b2a999dc5538f5e025b8b7364a | Remove hostedtoolcache | Michael Hudgins | michaelhudgins@google.com | 1,695,133,998,000 |
4da9ed4267c7afbc597af2e02b9f142cbbfc278e | Add a df-h to see how much space reclaimed | Michael Hudgins | michaelhudgins@google.com | 1,695,134,246,000 |
a90eb068e805fbe39ccfd7f4bfc2e33dd8a592a0 | Fix the missing tracing edge for profiling
PiperOrigin-RevId: 566639901 | Feng Wang | wffw@google.com | 1,695,137,168,000 |
8210e9a67b29bd7b5afd398a906fabdc23f272f8 | [XLA] Fix bug in collective-pipeliner pipeline tree of uses feature.
When multiple collectives feed each other computation we need to merge them.
PiperOrigin-RevId: 566651983 | Marcello Maggioni | maggioni@google.com | 1,695,140,027,000 |
60717e99312165ba09da1809980590e350a0660d | Do not eliminate reductions with dynamic axes.
Otherwise, the following example compiles incorrectly:
```
@tf.function(input_signature=[
tf.TensorSpec([None], tf.float32), tf.TensorSpec([None], tf.int32)])
def f(xs, axes):
return tf.reduce_sum(xs, axes, keepdims=True)
```
PiperOrigin-RevId: 5666538... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,140,483,000 |
536661702f4ffd5b2af4fc74b356acaa89d92107 | Add missing header.
PiperOrigin-RevId: 566654559 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,140,660,000 |
19deadfbcf1da6177fa6a087feeaa4bcb8768917 | PR #5635: [ROCm] FFI GpuStream for runtime
Imported from GitHub PR https://github.com/openxla/xla/pull/5635
This PR has enable more ROCm GpuContext for EnablePeerAccess and FFI GpuStream on runtime/executable.cc
Thanks in advance!
Copybara import of the project:
--
1d05f5c48e9d8f9bc13a513837c324a31e8e0e67 by Chao C... | Chao | cchen104@amd.com | 1,695,141,402,000 |
0d9abe289fcef137279ec79b02907dd94f18b7ed | [XLA:GPU] Use SM model instead of GPU name in HLO op profiler.
PiperOrigin-RevId: 566660219 | Ilia Sergachev | sergachev@google.com | 1,695,141,967,000 |
96f61e5378020cb2b62656293a11514f3a9fb3a2 | [XLA:GPU] Add H100 HLO operation profiles.
PiperOrigin-RevId: 566660291 | Ilia Sergachev | sergachev@google.com | 1,695,141,988,000 |
d3ba6d765a63a59d5e9808cfeb70a77483c17d61 | Change step_id propagation in MLRT batch kernel so that batch tasks can be linked to the `ProcessBatch` event.
PiperOrigin-RevId: 566665137 | Kuangyuan Chen | chky@google.com | 1,695,143,018,000 |
0da215bbc5dbc19ae0d04c965dcee5e99d85b36f | Integrate LLVM at llvm/llvm-project@71c83fb8b618
Updates LLVM usage to match
[71c83fb8b618](https://github.com/llvm/llvm-project/commit/71c83fb8b618)
PiperOrigin-RevId: 566667517 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,143,454,000 |
84a6eaa496e09b3783a7168cf904cf7a9695cc64 | [xla:gpu] Add a flag to enable pipelining point-to-point Send-Recv chains
Add xla_gpu_enable_pipelined_p2p_chain and use it to allow pipelining
point-to-point Send-Recv chains.
PiperOrigin-RevId: 566669028 | Bixia Zheng | bixia@google.com | 1,695,143,719,000 |
e32af092bb0f2eaf5dbbf0d2c7d806bda8f4ef8e | Simplify HostTransferManager lifespan management
PiperOrigin-RevId: 566669421 | Haibo Huang | hhb@google.com | 1,695,143,804,000 |
19d9294f62577c8796ebf4995a4c551f7f478010 | Use HloInstruction::ToShortString when printing very long HLOs for debugging in auto-sharding. Using the default HloInstruction::ToString can lead to log lines being truncated for such instructions.
PiperOrigin-RevId: 566669870 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,143,900,000 |
94cd985d08486157f534f1daa53a60d1968f9e5f | Remove unnecessary PluginIds -- everything was using kDefault anyway.
PiperOrigin-RevId: 566673062 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,144,538,000 |
ebd19138436b56fa0b2401334468c845ffd4bdbc | Update pod_viewer display.
PiperOrigin-RevId: 566676394 | Clive Verghese | cliveverghese@google.com | 1,695,145,138,000 |
31d691f48ee9d390089d5d66e9d743685af65ecc | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/28bf346777aa9ef8ac807ea29a14d0efa288743b.
PiperOrigin-RevId: 566679801 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,145,803,000 |
ad8a58ca1c827d92900edc7c7daf3375c9c9e8c3 | Empty the parallel_for/__init__.py file to eliminate code cycles.
PiperOrigin-RevId: 566682523 | Fiona Lang | flang@google.com | 1,695,146,298,000 |
4767c020dd366d887b054c1828513f9ed8444c52 | [XLA:LatencyHidingScheduler] Add kSerial to ResourceHazardType.
PiperOrigin-RevId: 566688116 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,147,330,000 |
f9fa232be12650504265f83ea1eb167564e18622 | [stream_executor] Add command buffer abstraction to StreamExecutor
No functional changes. Just plumbs CommandBuffer through StreamExecutor APIs to CUDA executor.
PiperOrigin-RevId: 566695040 | Eugene Zhulenev | ezhulenev@google.com | 1,695,148,646,000 |
c58f61a46cee27dd6238d467d8999649d8255751 | Merge pull request #61919 from MichaelHudgins:free-space-docker
PiperOrigin-RevId: 566696282 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,149,975,000 |
b81c2689d0e3175d48fe007b76fb731021b9d712 | Fix class names for custer_wrapper and item_wrapper.
PiperOrigin-RevId: 566700073 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,149,661,000 |
46252df296fbe45fb0d5c6854b0fbd5b0626efd9 | First iteration of the single custom op that can call our ml adjacent library.
PiperOrigin-RevId: 566705416 | Luke Boyer | lukeboyer@google.com | 1,695,150,682,000 |
305aad0c6cff5ed0ac93f1a09572dc03cc58183f | [XLA:Python] Initialize ABSL logging with absl::InitializeLog() when importing the python extension.
Otherwise code in XLA that uses ABSL logging will complain about ABSL logging not being initialized.
Fixes https://github.com/google/jax/issues/17664
PiperOrigin-RevId: 566710511 | Peter Hawkins | phawkins@google.com | 1,695,151,867,000 |
c33f3fe62ae246468dccdac19ab1f2943de2f3d5 | Integrate StableHLO at openxla/stablehlo@887e0692
PiperOrigin-RevId: 566712466 | Kevin Gleason | gleasonk@google.com | 1,695,152,309,000 |
b84ae5f43a86df1f114b12e80719b7bfb8918fa1 | Add a mock `Memory` type
PiperOrigin-RevId: 566726493 | Junwhan Ahn | junwhan@google.com | 1,695,155,312,000 |
e5c25301c3db12e55529617435907f657007365f | Use tsl logging instead of absl logging.
Otherwise we get log spam to stderr because absl::InitializeLog() isn't called.
PiperOrigin-RevId: 566726834 | Jieying Luo | jieying@google.com | 1,695,155,373,000 |
eae93023f9ec1075f22db38ff335904831971eab | [xla:gpu] Fallback when graph update result=kParametersChanged
the memset graph node can cause graph update to fail with result kParameterChanged
PiperOrigin-RevId: 566728346 | Anlun Xu | anlunx@google.com | 1,695,155,667,000 |
a0c477829a24e5be718537de45117afd1e68f798 | Skip overwriting host_indepdent_job_info if it contains zero profile_duration_ms to avoid synthesized host_indepedent_job_info.
PiperOrigin-RevId: 566736378 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,157,300,000 |
a0f3afb85185a86cf67d88ba35b39462290c71c9 | [graph_partition] Add VLOG(1) to show when we add a send/recv pair for a constant value.
PiperOrigin-RevId: 566747663 | Derek Murray | mrry@google.com | 1,695,159,668,000 |
f7ca233ae8051394e32d08fc20e0037a67744a8c | Change 5/6 for making MSA repacking slice aware.
Add optional data structures to repacking AllocationBlocks to export and import slicing data. Also, change MSA to populate the slicing data when it exports repacking AllocationBlocks. Repackers are free to ignore the information, which remains the case with this change.... | Ryan M. Lefever | lefever@google.com | 1,695,161,250,000 |
ab878a01f6599a4f2c212ad1a273b747b314f757 | added clang configs to .bazelrc | Raunak | mayank.kumar.raunak@intel.com | 1,695,162,963,000 |
b395fd3a892d69e62623cce123d9978bd178565d | Replicate constants to CPUs, not TPUs. A small constant with successors on TPUs should be replicated to its TPU's hosts, not the TPU devices themselves.
PiperOrigin-RevId: 566761093 | Michael Delorimier | mdel@google.com | 1,695,162,511,000 |
f91625e9541f6624714ffec5c8fae778ea08083c | [XLA:GPU][NFC] Move ProjectedShmemUsageBytes into gpu_fusible
PiperOrigin-RevId: 566763146 | Son Tuan Vu | vuson@google.com | 1,695,162,857,000 |
0001fa0537e672e493977152858e324355022bb2 | Global rename of tf2xla/api/v0 to tf2xla/api/v1 to line up with tf1/tf2
PiperOrigin-RevId: 566764049 | Mason Chang | masonchang@google.com | 1,695,163,026,000 |
81ff6bb9ca0f89a25d4ee6a1ec79b2b3d6d30356 | Integrate LLVM at llvm/llvm-project@bbe3ee061f6d
Updates LLVM usage to match
[bbe3ee061f6d](https://github.com/llvm/llvm-project/commit/bbe3ee061f6d)
PiperOrigin-RevId: 566766975 | Fangrui Song | maskray@google.com | 1,695,163,610,000 |
99bbebf542588c8eca8d773d6ae04e7d4b362878 | [PJRT C API] Implement PjRtCApiClient::platform_id().
Set PjRtCApiClient::platform_id to be tsl::Fingerprint64(platform_name). Cache both the platform_name and platform_id.
PiperOrigin-RevId: 566767022 | Jieying Luo | jieying@google.com | 1,695,163,622,000 |
e7b01b7c9b18d45d57f6ebd7032a118259be12b9 | Replicate small constants so they don't need to be sent to their successors. A small constant is replicated to each of its successors' devices. The maximum size of a constant to be replicated is 16 elements.
This pass is disabled by default and can be enabled with the flag replicate_small_constants.
PiperOrigin-RevId... | Fiona Lang | flang@google.com | 1,695,166,509,000 |
ebbd3daf5bbd5aee71cc62abc99873bf728f22dd | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/e20500b4e4cb784965a31f67b062cf4982ab1136.
PiperOrigin-RevId: 566787295 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,167,816,000 |
32e97277fd431a05b242056c8c9cb6b923c08ab3 | add id dropping to the sparse core preprocess ops.
PiperOrigin-RevId: 566792706 | Ziyin Huang | ziyinh@google.com | 1,695,169,036,000 |
8f4ef087b7a816a50f89e4fd7db919e2b1322e10 | Add CacheDatasetV2Op to the list of supported TF dataset ops.
PiperOrigin-RevId: 566796779 | Jared Junyoung Lim | jaredlim@google.com | 1,695,170,001,000 |
a9075649062815e7859544728a0c3e0af9311a0c | Pass along the `shape` param from `tf.compat.v1.get_variable` to its underlying variable creator. At this point, we have already checked that either `shape` is compatible w/ `initial_value`, or that it is not specified (`None`). Therefore it should always be safe to pass it. Having it can help subsequent nested variabl... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,172,530,000 |
4d11751715c4a55f3ad2f3a245f0b1a738db0baf | Remove xnnpack_for_tflite and xnnpack_for_tfjs
With the changes to the way we do config, these 2 targets are not useful anymore, all callers should transition to use :XNNPACK directly.
PiperOrigin-RevId: 566813566 | Zhi An Ng | zhin@google.com | 1,695,174,916,000 |
c3410229ff4078ba8f8b8573dc0bcffb2f95dbb2 | Add chlo passes for decomposing chlo::broadcast ops
ConvertTFQuantOpsToMHLO pass may create some chlo::broadcastxxx ops with UQ types. Need to handle them in the lowering passes.
PiperOrigin-RevId: 566814097 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,175,114,000 |
3e5894abb3f2d1918d8cbcbbf974fcb11fd91921 | Replace jakeharmon8's LDAP with GitHub username
PiperOrigin-RevId: 566817403 | Jake Harmon | jakeharmon@google.com | 1,695,176,270,000 |
76ae1b611bdb3c3c9c7bdbd5de0b3c65d265809c | [stream_executor] NFC: Clean up xla/stream_executor:event build dependencies
PiperOrigin-RevId: 566830890 | Eugene Zhulenev | ezhulenev@google.com | 1,695,181,248,000 |
f19e9afc30cbc03b3152a1e11d45a62fa2ebb30a | [stream_executor] NFC: Clean up stream executor headers
PiperOrigin-RevId: 566835062 | Eugene Zhulenev | ezhulenev@google.com | 1,695,182,901,000 |
e784f7d9278dabdb541362c6ba82680b6ff42d4b | lite: disable serialize stablehlo ops by default
PiperOrigin-RevId: 566850779 | Zichuan Wei | zichuanwei@google.com | 1,695,189,129,000 |
917c9e33a40cf07725f86700ab3960681a98180c | Add memory kind check in `PjRtArray::Create`.
PiperOrigin-RevId: 566851924 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,189,479,000 |
9cdd66e6b75e8cf1720fd4e3d54161e25ea223aa | [XLA:GPU] Skip slicing to scalars in Triton GEMM rewriter.
PiperOrigin-RevId: 566890235 | Ilia Sergachev | sergachev@google.com | 1,695,199,989,000 |
4e101ec87686515d10830da52b735c87ecbaab95 | compat: Update forward compatibility horizon to 2023-09-20
PiperOrigin-RevId: 566892239 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,200,525,000 |
79be3fa6129625e695913c99ad0d7592c733fd27 | Update GraphDef version to 1625.
PiperOrigin-RevId: 566892275 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,200,530,000 |
f12ce2113e361dac536d95d6f1b396048259a3c3 | Conv3D doesn't have to transpose weights
PiperOrigin-RevId: 566894491 | Alan Kelly | alankelly@google.com | 1,695,201,121,000 |
fea0c082192a4063faf3b2f54a48b52cbb602487 | [XLA:GPU] Fix boundary checks for split-K GEMM fusions
Without this fix, we get CUDA_ERROR_ILLEGAL_ADDRESS on the added test case.
PiperOrigin-RevId: 566902695 | Tamás Danyluk | tdanyluk@google.com | 1,695,203,350,000 |
879ffa3aca8e6bdb3eb9ae1a3081dc7d37ce0d23 | BF16 capability detection for Ubuntu 20.04
This adds a patch to oneDNN for BF16 capability detection
for Ubuntu 20.04 on aarch64.
The contents in the patch are fully authored by @kawakami-k,
the source used for this patch is available here:
oneapi-src/oneDNN#1670 | Fadi Arafeh | fadi.arafeh@arm.com | 1,695,058,830,000 |
f3578fae9a3195dd6ba58cd896a08869c5783ff2 | Initialize creation_pass_id in GPU Compiler
This sets the creation pass id and the logical creation pass id to -1
for all ops in the input HLO. -1 is the special value which identifies
ops present in the input HLO.
By setting the -1 for input ops we will be able to differentiate ops
originating from the input from op... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,211,912,000 |
1b7a86988e88103ba5567f478f6c9425b3766e01 | [XLA:GPU] Priority fusion: don't allow partial fusions.
Only consider cases when a producer can be fused will all the consumers. Otherwise fusion is not beneficial.
PiperOrigin-RevId: 566932055 | Oleg Shyshkov | shyshkov@google.com | 1,695,212,574,000 |
40567bcecebfaca299cd1bc1ac8019bddb2f4db6 | [ROCm]: Updates for rocm_dnn header dependency | Rahul Batra | rahbatra@amd.com | 1,664,294,606,000 |
9ff97264b103695012907913ebce1d0a3d3b0930 | Integrate LLVM at llvm/llvm-project@2baf4a06ef06
Updates LLVM usage to match
[2baf4a06ef06](https://github.com/llvm/llvm-project/commit/2baf4a06ef06)
PiperOrigin-RevId: 566960191 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,220,772,000 |
0d09ec0c1e2cce294b338e8f2e0e52ce9ccdc5e3 | Move missed bazel config to LLVM17 toolchain
PiperOrigin-RevId: 566966919 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,222,433,000 |
a7e3dec0b6681ca070b29a4cb9d8a3d709b1f9ba | Merge pull request #61917 from mo271:scorecard_link
PiperOrigin-RevId: 566969250 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,223,431,000 |
a8866e20fafb66161e403493f188a5639a023028 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/752d6d83d403986227dffe42beb5014843cf2ddb.
PiperOrigin-RevId: 566971528 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,223,380,000 |
69923fd52754c1c501e45f8cd87a7d93a74bcc4c | Fix layout logic in IrArray::Index::Linearize.
The logic previously assumed the layout was major-to-minor. In practice, I believe this assumption is correct in all places where Linearize is called, because LayoutNormalization converts most instructions to be major-to-minor. But in general, IrArray supports arbitrary l... | Reed Wanderman-Milne | reedwm@google.com | 1,695,223,472,000 |
5f7dd0c0e3df973e1aff926fec11f08df7fcf859 | Internal, fix build error.
PiperOrigin-RevId: 566975236 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,224,283,000 |
95060833cffbc61eecea769d07bb05914f50d502 | Update Docker image tags and digests in Kokoro and Bazel
This is changing the Kokoro job config back to its original state (using the latest-python3.X docker image tags) and is updating the Bazel remote toolchain config to the latest Docker containers. (The ones created here https://github.com/tensorflow/tensorflow/ac... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,224,301,000 |
1f9a25d2f55aeab6c57ece9173f27fb1af3ffda2 | Restore TSL/XLA headers in tensorflow/include
The headers were previously in tensorflow/include/tensorflow/tsl and tensorflow/include/tensorflow/compiler/xla, but they can now be found in tensorflow/include/{tsl,xla}. C projects built against the TF wheel should now build unless they directly reference one of these he... | Jake Harmon | jakeharmon@google.com | 1,695,226,526,000 |
1e035a5f6055c0531f35b1ec8508a54fcf2737ca | Proposal to create a cluster Bridge API for TF:
V1 - Used when the bridge is called via the Function API, which roughly aligns to TF2.
V0 - Used when the Bridge is called via the Session API. We're limited to not supporting infeed / outfeed ops.
After this change, existing bridge.cc logic gets pulled out into here.
... | Mason Chang | masonchang@google.com | 1,695,227,125,000 |
ab54126289a5ee3976bc4de6d93032cf6b750ffa | [NFC] Log NCCL calls for AllToAll
PiperOrigin-RevId: 566988443 | Rahul Joshi | jurahul@google.com | 1,695,227,314,000 |
105a1a38382f2574222072d86bf4dbc8ae2fb7ae | Add tf_mlrt.await_all_control op if there is any unused future.
PiperOrigin-RevId: 566992749 | Kuangyuan Chen | chky@google.com | 1,695,228,219,000 |
d03c477d727b93b71ac1710885c6c918d7754361 | Add licenses and notices for third party libraries
PiperOrigin-RevId: 566995194 | Raviteja Gorijala | gorijala@google.com | 1,695,228,713,000 |
879a5e37eb350f556debec3cd5c9b6dcb8809318 | Rename tsl/cuda/BUILD to tsl/cuda/BUILD.bazel.
Unify _stub and _lib targets in tsl/cuda (under an unsuffixed name), and update users to point directly to merged target. This also allows us to remove some .bzl macros.
PiperOrigin-RevId: 566996155 | Peter Hawkins | phawkins@google.com | 1,695,228,915,000 |
8ddb2e6879fedfa6928bdbcb97c2e665464338ef | Refer to CUDA stubs directly from TSL, rather than using an alias defined in xla/stream_executor.
Remove the aliases in xla/stream_executor.
PiperOrigin-RevId: 567025507 | Peter Hawkins | phawkins@google.com | 1,695,234,047,000 |
5c30dcb3d2d62751a11fbace1c8aa79eb86c0617 | added host_linkopt | Raunak | mayank.kumar.raunak@intel.com | 1,695,235,244,000 |
9770efd3b28709716705ffd11af55fa1c7a60637 | PR #5749: [ROCm] gpu command buffer for ROCm
Imported from GitHub PR https://github.com/openxla/xla/pull/5749
ROCm enable command buffer https://github.com/openxla/xla/commit/ec74b3b29d549ba97f499751651684800ab55aa6
@ezhulenev @akuegel Thanks in advance
Copybara import of the project:
--
aca5aa0e853b201c94420f41c22... | Chao | cchen104@amd.com | 1,695,235,038,000 |
f17f1a7b0c0abc9b1e0b11a461f71685b51aae2d | add id shuffling to the sparse core preprocess ops.
PiperOrigin-RevId: 567040448 | Ziyin Huang | ziyinh@google.com | 1,695,236,811,000 |
6dc6334a711da5b8135fe6da1445a0b259ab69ac | Rollback of commit a9075649062815e7859544728a0c3e0af9311a0c
Pass along the `shape` param from `tf.compat.v1.get_variable` to its underlying variable creator. At this point, we have already checked that either `shape` is compatible w/ `initial_value`, or that it is not specified (`None`). Therefore it should alway...
... | Mason Chang | masonchang@google.com | 1,695,237,030,000 |
04c9a3483f02e6d04bb90032df5812065e5d4143 | [XLA:GPU] Triton GEMM: fix fusion traversal logic.
PiperOrigin-RevId: 567046240 | Ilia Sergachev | sergachev@google.com | 1,695,238,077,000 |
0e3480236cec19ea558cd93dd017013e5cfee1b3 | include THIRD_PARTY_NOTICES.txt in the wheel.
PiperOrigin-RevId: 567049176 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,238,766,000 |
5b534b09c7ec440176d3ac85f8c0976217c0830b | Update metrics name from v0 to v1
PiperOrigin-RevId: 567061192 | Mason Chang | masonchang@google.com | 1,695,241,305,000 |
9258e1aeb9fdbdd8566847f31a7c64243ba6d599 | Update tpu build flags in .bazelrc
PiperOrigin-RevId: 567062395 | Kanglan Tang | kanglan@google.com | 1,695,241,539,000 |
9d138e582de5f9de67afc7608ac9258a7d6670f1 | Modify ARM64 docker images to build under TF repo
PiperOrigin-RevId: 567065712 | Michael Hudgins | michaelhudgins@google.com | 1,695,242,213,000 |
2ebb1b5b001d730a7077d8d60ff459443a54fb7c | Remove fallback to the old xla builder bridge and rely on the combined bridge's call.
PiperOrigin-RevId: 567068282 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,242,715,000 |
03c58a3fa7229078dd4fab969dc18dc2c8a7265a | [stream_executor] Remove unused gpu_launch_dim.h
PiperOrigin-RevId: 567068762 | Eugene Zhulenev | ezhulenev@google.com | 1,695,242,801,000 |
ea0637ac525b4b1e1794373014f1159a7e13ebcb | Mesh creation API should also support taking a list of DeviceSpec as input
PiperOrigin-RevId: 567074638 | Bing Hu | binghu@google.com | 1,695,243,944,000 |
8fcaa1bb3130fdcc2f859820c4de49f029e355b2 | Roll back the accidental change for the xla pad size in GetMinibatchOp.
PiperOrigin-RevId: 567077638 | Ziyin Huang | ziyinh@google.com | 1,695,244,542,000 |
e2624e25d0c9465f631389f4c3375d7a3df5ea47 | Support storing bfloat16 in tensor_content format.
PiperOrigin-RevId: 567079988 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,245,080,000 |
caf5b4d890095df814d8b39d2ac9717aa32ecf1a | Remove obsolete bincount code since forward compatibility window has expired
PiperOrigin-RevId: 567085770 | Edward Schwartz | schwartzedward@google.com | 1,695,246,276,000 |
de064617a4e8c07a2f4646447e84dfa45a66c80d | [XLA] Fix DCHECK in ShapeUtil::ByteSizeOf that failed when shape didn't have a layout
PiperOrigin-RevId: 567088606 | Berkin Ilbeyi | berkin@google.com | 1,695,246,870,000 |
7974a0b3afd2b53e92491a1bdea3b54294e0f6e7 | [xla:gpu] NFC: Print node type instead of node pointer in error messages
PiperOrigin-RevId: 567102686 | Anlun Xu | anlunx@google.com | 1,695,249,993,000 |
c2519c4aa372eecc6a26dfd4dedee866f3f0f2e3 | Pass along the shape param from tf.compat.v1.get_variable to its underlying variable creator. At this point, we have already checked that either shape is compatible w/ initial_value, or that it is not specified (None). Therefore it should always be safe to pass it. Having it can help subsequent nested variable creators... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,251,015,000 |
af50677da4f39028b85fe42df41cf1ebab68bea5 | Strip `external/local_tsl` prefix during zip of tsl protos | William Muir | wamuir@gmail.com | 1,695,251,397,000 |
5e3184466b9f28423f5538eceec6b61274e0ade8 | Exercise new version of TF/XLA bridge in legalize_tf_quant_test
PiperOrigin-RevId: 567108435 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,251,341,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.