AFAIK when adding new ops, you need to install from sources:
https://www.tensorflow.org/how_tos/adding_an_op/

From your description, I think you maybe installed the pip package?
When I look at our repo, both master and r0.12 branch has the file:
https://github.com/tensorflow/tensorflow/blob/r0.12/tensorflow/core/platform/stream_executor.h
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/platform/stream_executor.h

So I am not sure what you installed, or what the reason might be.
Please try installing from sources and if the problem persists, include more debugging information, detailing each step you took to reproduce the problem.

It's possible our stream executor headers are not packaged in pip. I can take a look tomorrow

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/pip_package/setup.py#L159 it looks like we're missing tensorflow/stream_executor there

I'm not sure if we're missing anything else, unfortunately, but I can add that change.

This is not complete yet. When I try to include stream_executor.h:

#include "tensorflow/core/platform/stream_executor.h"

I get:
/u/zeyer/.local/lib/python2.7/site-packages/tensorflow/include/tensorflow/stream_executor/dso_loader.h:32:30: fatal error: cuda/cuda_config.h: No such file or directory

I get the same error with stream_executor.h, when compiling a custom op with nvcc and 1.3.0 binaries:

In file included from /usr/local/lib/python3.5/dist-packages/tensorflow/include/tensorflow/core/platform/default/stream_executor.h:26:0,
                 from /usr/local/lib/python3.5/dist-packages/tensorflow/include/tensorflow/core/platform/stream_executor.h:24,
                 from /usr/local/lib/python3.5/dist-packages/tensorflow/include/tensorflow/core/util/cuda_kernel_helper.h:26,
                 from roi_align_gpu.cu.cc:8:
/usr/local/lib/python3.5/dist-packages/tensorflow/include/tensorflow/stream_executor/dso_loader.h:32:30: fatal error: cuda/cuda_config.h: No such file or directory

Same problem here with 1.3.0 compiled from source when I try to compile a custom op. Does anyone have a workaround?

@pronobis it seems like you're suggesting a workaround yourself in #12860:

Copying cuda_config.h to /site-packages/tensorflow/include/tensorflow/stream_executor/cuda solves the problem.

Are you asking for something else?

I suppose #12860 is there now to track this problem. Generally, I'm looking for a better permanent solution, but again, that's what #12860 is for.

I wonder what the state is here? Wouldn't it be possible to just add cuda_config.h to the pip package? Wouldn't that solve the problem?

It has been 14 days with no activity and this issue has an assignee.Please update the label and/or status accordingly.

@av8ramit could you check if we have all the dependencies of stream_executor.h as a part of the pip package, except for cuda headers?

@gunan, the dependencies are in the pip package

Nagging Assigneee: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

All the dependencies are in the pip package.

Folding all the duplicates into #15002