Bug#907192: [done] Re: RFS: tensorflow/1.10.0+dfsg-A1 [ITP]
control: close -1
I'll sponsor myself shortly as long as nothing goes wrong in the last build.
This will be an ~300MB initial upload.
On Fri, Sep 07, 2018 at 01:23:03PM +0000, Mo Zhou wrote:
> Hi mentors,
>
> Does anyone have idea about the following very last blocker for libtensorflow?
> We are really ready to upload TF once the C API unit tests passed without fatal error.
Finally I sorted out the very last bug with a hack instead of really
looking into it. I don't understand why upstream store the CPU frequency
in MHz as an int64 instead of a float. And that might be the root of the
very last problem. Anyway let's upload it first since the patched shared
object appears to be working correctly.
Some final comments:
0. The build system of our TF package is written by me from scratch,
but spiritually it tries its best to produce similar result compared
to bazel.
1. I know most people use TF's python interface, but please let me
postpone the python interface in the future uploads. We currently
have C and C++ interface::
libtensorflow-framework1.10
a bare framework shared object which ships no OP or Kernel.
libtensorflow1.10
C interface
libtensorflow-cc1.10
C++ interface
1. There are 421 commits in the git repo including many one-line-change commits.
2. The latest upload count is 42 (A1u42 means stage 1 upload 42).
3. Whlie creating this package I became a big fan of ninja-build, which
works overwhelmingly better than Make when doing incremental build,
FTBFS debugging, and checking screent output with -j72.
marking this RFS as done.
Reply to: