Map tensorflow/c/c_api_experimental.h header file #410
Conversation
This file contains 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
karllessard
added
CI build
|
Thanks @saudet , looks good, I’ve triggered a CI build before merging |
|
@saudet , I got a bunch of errors on Windows, any clue? |
|
Hum, it looks like the experimental APIs are not exported on Windows. I remember bumping into that previously. I suppose we can always hack the Bazel build to export them... TF Core's build seems to be using this Python script to exclude symbols, but I don't see where it's removing any of the "experimental" functions: Ah, yes, I remember now. The experimental APIs are not actually compiled at all by Bazel. We need to compile them separately, like this: java/tensorflow-core/tensorflow-core-api/pom.xml Lines 341 to 344 in ae74151 The problem with that though is that the experimental APIs themselves need to link with the internal API, which is not (cannot) be exported on Windows... 🤦 Why can't they just create multiple DLLs like PyTorch does it, seriously. It does work on Linux and Mac though, but then we're alienating Windows users... |
|
If we're doing this mainly for pluggable device support then I think the only extant device is the one Apple released for macOS, so it doesn't matter too much if it's not working on Windows. |
|
Well well well, look at that, the Python binaries on PyPI do export it on Windows: $ winedump -j export python/_pywrap_tensorflow_internal.pyd | grep TF_LoadPluggableDeviceLibrary
0089FAB0 42047 TF_LoadPluggableDeviceLibrarySo this sounds like a bug in the C/C++ build of TF Core. |
|
I exposed a bunch of things manually as part of the custom gradients PR, you might be able to add this to this list. |
Thanks! But it's not the symbols, but the functions themselves that are missing. For some reason Bazel doesn't compile the experimental APIs when building libraries for the C/C++ API, but it does compile them when building for the Python API. |
|
We'd need to see if it's supported by the binaries on PyPI for the relevant platforms as we now use those rather than building the TF native library ourselves. What platforms do you need? |
We need for 'linux-x86_64'. Our plugins work with the native TF (https://pypi.org/project/tensorflow/). |
|
Ok, we can take a look at it after 1.0 is released. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
So that we can call goodies like
TF_LoadPluggableDeviceLibrary(), see issues #409./cc @ashesfall