experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[tf.data] Apply single-device multi component function optimization in RunSync().

Browse Source 
Previously, `RunSync()` was missing the optimization from e5c6881c77 that enables a single-component multi-device function from being executed as a local function. When 624f1a0f82 switched all tf.data functions to be considered as multi-device functions, the optimization became load-bearing, because occasionally it will dispatch a function to a ProcessFunctionLibraryRuntime that does not actually support multi-device functions, because it has no rendezvous factory.

The fix is simple: update the call to `GetHandleOnDevice()` in `PrepareRunSync()` to pass the optional `include_multi_device = true` argument.

PiperOrigin-RevId: 309481098
Change-Id: Id12d01cffc50399bab5711c7de4756c542873ec2
This commit is contained in:
Derek Murray 2020-05-01 15:04:34 -07:00 committed by TensorFlower Gardener
parent 412b12ea8d
commit fe1ff3c01a
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
tensorflow/core/common_runtime/function.cc
View File

@ -1235,7 +1235,8 @@ Status FunctionLibraryRuntimeImpl::PrepareRunSync(
run_opts->create_rendezvous = false;
}
LocalHandle local_handle = parent_->GetHandleOnDevice(device_name_, handle);
LocalHandle local_handle = parent_->GetHandleOnDevice(
device_name_, handle, /*include_multi_device=*/true);
if (local_handle == kInvalidLocalHandle) {
*out_item = nullptr;
return Status::OK();