TensorFlow Serving using TensorFlow 1.14.0-rc0.
Major Features and Improvements
- Use MKL-DNN contraction kernels by default. (commit: a075ebe)
- Add option to refuse to unload the last servable version. (commit: c8496b1)
- Add ability to disable periodic filesystem polling (#1295). (commit: 7245055)
Breaking Changes
- No breaking changes.
Bug Fixes and Other Changes
- Add
enforce_session_run_timeout
insideServer::Options
. (commit: de03064) - Add -o option, to pass params to
docker
command. (commit: dd59021) - Stop using reader locks (tf_shared_lock) on the read path of FastReadDynamicPtr. (commit: f04e583)
- Add saved model tags to logging metadata. These tags are used by (commit: 6320701)
- Adds an option in
SessionBundleConfig
to repeat warmup replay n times per request. (commit: 15cd202) - Improve tpu server warm up (commit: 63d31a3)
- Official PIP package releases are now tied to a specific version of TensorFlow (commit: 9514c37)
- Bump the minimal Bazel version to 0.24.1 (commit: 96a716c)
- Add new device type for TPU. (commit: c74861d)
- Fix incorrect formatting of decimal numbers in JSON output (#1332) (commit: d7c3b3d)
- Fixed the gzip uncompression support in the HTTP server for large request bodies. (commit: fb7835c)
- Add stack memory resource kind. (commit: e56e72b)
- Adds ModelServer test for loading SavedModel exported from Keras Sequential API (commit: 9578f3d)
- Ignore SIGPIPE for libevent,prevent the SIGPIPE signal from being raised (#1257) (commit: 8d88a5b)
- Fix #1367 (commit: 58af901)
- Update Serving_REST_simple.ipynb (commit: 3870ba5)
- Updates README with link to architecture overview (commit: d233a82)
- Update example section to use Docker (commit: a5fc8bb)
Thanks to our Contributors
This release contains contributions from many people at Google, as well as:
G. Hussain Chinoy, Karthik Vadla, mkim301, yjhjstz