1.13.0

TensorFlow Serving using TensorFlow 1.13.1

Major Features and Improvements

• Support for TensorRT 5.0 (GPU docker image built against CUDA 10 and TensorRT 5.0)
• Support for listening gRPC over UNIX socket (commit: a25b0da)
• New GPU version of TensorFlow Serving API PIP package. This depends on the tensorflow-gpu instead of tensorflow PIP package, but is otherwise identical. (commit: 525c1af)
• TF Serving end-to-end colab! Training with Keras, serving with TF Serving and REST API (commit: 1ff8aad)

Breaking Changes

• No breaking changes.

Bug Fixes and Other Changes

• Make error message for input size mismatch in Predict call even more actionable. (commit: 7237fb5)
• Document how to use the version policy to pin a specific version, or serve multiple versions, of a model. (commit: 2724bfe)
• Document config reloading and model version labels. (commit: f4890af)
• Fix the compile error on ARM-32 in net_http/server. (commit: 5446fd9)
• Adds ModelSpec to SessionRunResponse. (commit: 58a2263)
• Add MKL support (commit: 8f79253)
• Fix default path of Prometheus metrics endpoint (commit: 9d05b0c)
• Add monitoring metrics for saved model (export_dir) warm up latency. (commit: de0935b)
• Add more details/clarification to model version labels documentation. (commit: f9e6ac4)
• Split --tensorflow_session_parallelism flag into two new flags: --tensorflow_intra_op_parallelism and --tensorflow_inter_op_parallelism (commit: 71092e4)
• Update CPU Docker images to Ubuntu 18.04 (commit: 8023fba)
• Upgrade to Bazel 0.20.0 (commit: fc0b75f)
• Update Python 2 scripts to be compatible with both Python 2 and 3 (commit: 846d443)

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Daniel Shi, Karthik Vadla, lapolonio, robert, Shintaro Murakami, Siju, Tom Forbes, Ville TöRhöNen

1.13.0-rc1

TensorFlow Serving using TensorFlow 1.13.0-rc1

Major Features and Improvements

• Support for TensorRT 5.0 (GPU docker image built against CUDA 10 and TensorRT 5.0)
• Support for listening gRPC over UNIX socket (commit: a25b0da)
• New GPU version of TensorFlow Serving API PIP package. This depends on the tensorflow-gpu instead of tensorflow PIP package, but is otherwise identical. (commit: 525c1af)
• TF Serving end-to-end colab! Training with Keras, serving with TF Serving and REST API (commit: 1ff8aad)

Breaking Changes

• No breaking changes.

Bug Fixes and Other Changes

• Make error message for input size mismatch in Predict call even more actionable. (commit: 7237fb5)
• Document how to use the version policy to pin a specific version, or serve multiple versions, of a model. (commit: 2724bfe)
• Document config reloading and model version labels. (commit: f4890af)
• Fix the compile error on ARM-32 in net_http/server. (commit: 5446fd9)
• Adds ModelSpec to SessionRunResponse. (commit: 58a2263)
• Add MKL support (commit: 8f79253)
• Fix default path of Prometheus metrics endpoint (commit: 9d05b0c)
• Add monitoring metrics for saved model (export_dir) warm up latency. (commit: de0935b)
• Add more details/clarification to model version labels documentation. (commit: f9e6ac4)
• Split --tensorflow_session_parallelism flag into two new flags: --tensorflow_intra_op_parallelism and --tensorflow_inter_op_parallelism (commit: 71092e4)
• Update CPU Docker images to Ubuntu 18.04 (commit: 8023fba)
• Upgrade to Bazel 0.20.0 (commit: fc0b75f)
• Update Python 2 scripts to be compatible with both Python 2 and 3 (commit: 846d443)

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Daniel Shi, Karthik Vadla, lapolonio, robert, Shintaro Murakami, Siju, Tom Forbes, Ville TöRhöNen

1.12.0

TensorFlow Serving using TensorFlow 1.12.0

Major Features and Improvements

• Add new REST API to get model status from ModelServer (commit: 00e459f)
• Add new REST API to get model metadata from ModelServer (fixes #1115) (commit: 9768702)
• Support accepting gzipped REST API requests (fixes #1091) (commit: b94f6c8)

Breaking Changes

None

Bug Fixes and Other Changes

• Update MKL build (commit: e11bd51)
• Remove version pinning on pip packages (commit: 462072c)
• Update basic serving tutorials (commit: 33a4b05)
• Replacing legacy_init_op argument in SavedModelBuilder with main_op. (commit: 2fda31f)
• Add git hash for version metadata of model server and add tags for dev and nightly builds. (commit: 5c7740f)
• Add error messages for specific cases when json for REST requests (commit: a17c892)
• Python examples now run in a hermetic environment with all required dependencies (commit: 793fd90)

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Charles Verge, demfier, Kamidi Preetham, Lihang Li, naurril, vfdev, Yu Zheng

1.12.0-rc0

TensorFlow Serving using TensorFlow 1.12.0-rc2

Major Features and Improvements

• Add new REST API to get model status from ModelServer (commit: 00e459f)
• Add new REST API to get model metadata from ModelServer (fixes #1115) (commit: 9768702)
• Support accepting gzipped REST API requests (fixes #1091) (commit: b94f6c8)

Breaking Changes

Bug Fixes and Other Changes

• Update MKL build (commit: e11bd51)
• Remove version pinning on pip packages (commit: 462072c)
• Update basic serving tutorials (commit: 33a4b05)
• Replacing legacy_init_op argument in SavedModelBuilder with main_op. (commit: 2fda31f)
• Add git hash for version metadata of model server and add tags for dev and nightly builds. (commit: 5c7740f)
• Add error messages for specific cases when json for REST requests (commit: a17c892)
• Python examples now run in a hermetic environment with all required dependencies (commit: 793fd90)

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Charles Verge, demfier, Kamidi Preetham, Lihang Li, naurril, vfdev, Yu Zheng

1.11.1

TensorFlow Serving using TensorFlow 1.11.0

Bug Fixes and Other Changes

• Fix version of model server binary (Fixes #1134)
• Range check floating point numbers correctly (Fixes #1136).
• Fix docker run script for same user and group name (Fixes #1137).
• Fix GPU build (Fixes #1150)

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

vfdev

1.11.0

TensorFlow Serving using TensorFlow 1.11.0

Major Features and Improvements

• Prometheus exporter for TF metrics (see 021efbd for details).

Breaking Changes

• No breaking changes

Bug Fixes and Other Changes

• Built against TensorFlow 1.11.0
• Accept integers for float/doubles in JSON REST API requests
• TF Serving API is now pre-built into Docker development images
• GPU Docker images are now built against cuDNN 7.2
• Add --max_num_load_retries flag to ModelServer (fixes #1099)
• Add user-configured model version labels to the stand-alone ModelServer binary.
• Directly import tensor.proto.h (the transitive import will be removed from tensor.h soon)
• Building optimized TensorFlow Serving binaries is now easier (see docs for details)
• Adds columnar format support for input/output tensors in Predict REST API (fixes #1047)
• Development Dockerfiles now produce a more optimized ModelServer
• Fixed TensorFlow Serving API PyPi package overwriting TensorFlow package.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Feisan, joshua.horowitz, Prashanth Reddy Basani, tianyapiaozi, Vamsi Sripathi, Yu Zheng

1.11.0-rc1

TensorFlow Serving using TensorFlow 1.11.0-rc1

1.11.0-rc0

Release 1.11.0-rc0

Major Features and Improvements

• Prometheus exporter for TF metrics (see 021efbd for details).
• Added new REST API to get status of model(s) from ModelServer.

Breaking Changes

• No breaking changes

Bug Fixes and Other Changes

• Built against TensorFlow 1.11.0-rc0.
• Directly import tensor.proto.h (the transitive import will be removed from tensor.h soon)
• Building optimized TensorFlow Serving binaries is now easier (see docs for details)
• Adds columnar format support for input/output tensors in Predict REST API (fixes #1047)
• Development Dockerfiles now produce a more optimized ModelServer
• Fixed TensorFlow Serving API PyPi package overwriting TensorFlow package.

TensorFlow Serving 1.10.1

Release 1.10.1

Bug Fixes and Other Changes

• Fixed TensorFlow Serving API PyPi package overwriting TensorFlow package.
• Update Docker configs to fix tensorflow/tensorflow#21518

TensorFlow Serving 1.10.0

Release 1.10.0

Major Features and Improvements

• No major features or improvements.

Breaking Changes

• TensorFlow Serving API now uses gRPC's GA release. The beta gRPC API has been deprecated, and will be removed in a future version of TensorFlow Serving. Please update your gRPC client code (sample)
• Docker images for GPU are built against NCCL 2.2, in following with TensorFlow 1.10.

Bug Fixes and Other Changes

• Built against TensorFlow 1.10.0
• Added GPU serving Docker image.
• Repo cloning and shell prompt in example readme.
• Updated Docker instructions.
• Updated min Bazel version (0.15.0).
• Convert TF_CHECK_OKs to TF_ASSERT_OK in some unit tests.
• Remove error suppression (.IgnoreError()) from BasicManager.
• Add new bazel_in_docker.sh tool for doing hermetic bazel builds.
• Fix erroneous formatting of numbers in REST API output that are larger than 6 digits.
• Add support for Python 3 while also compatible with Python 2.7 in mnist_saved_model.py.
• Fix an incorrect link to Dockerfile.devel-gpu.
• Add util for get model status.
• Adding support for secure channel to ModelServer.
• Add version output to model server binary.
• Change ServerRequestLogger::Update to only create new and delete old loggers if needed.
• Have the Model Server interpret specific hard-coded model version labels "stable" and "canary" as the smallest and largest version#, respectively.
• Add half_plus_two CPU and GPU models to test data.