Major Features and Improvements
- No major features or improvements.
Breaking Changes
- No breaking changes.
Bug Fixes and Other Changes
- Init
python_wheel_version_suffix_repositoryin tensorflow-serving after
tensorflow/tensorflow@805775f.
(commit: 21c6421) - Use tf_workspace* alias when initializing TensorFlow's external dependencies. (commit: ebb194c)
- Replace deprecated TF status and mutex types with absl replacement types. (commit: bec6e1c)
- Update release notes for 2.18.1 (commit: 564d335)
- Fix a bug in ServerRequestLogger::Update() (commit: d83a861)
- Fix a absl log MACRO redefinition bug. (commit: bfd832d)
- Provide Separate Validate()/Apply() calls for Logger Update (commit: 8c8a286)
- Add 2 new resource kinds constants for RAM (commit: dff9dd8)
- Clarifying in the API that some request options are only supported for disagg serving. (commit: a1d3566)
- Align Protobuf version with Tensorflow (commit: a692125)
- PredictRequest API update to enable cancellation control messages (commit: 8c846b7)
- Update release notes for 2.19.0-rc0 (commit: 305a2f7)
- Update release notes for 2.19.0 (commit: cb03205)
- Introduce kSteadyHeapRamBytes and kSharedHeapRamBytes to resource kinds. (commit: e4b44e8)
- Allow model server to reload a config if it doesn't change from the previous loading attempt. (commit: a54796b)
- Do not return
SignatureDef.defaultsfromGetModelMetadata(commit: 5083127) - Add disable_host_compilation option to allow disabling host compilation after load. (commit: f105002)
- Use a copy of ModelWarmupOptions for model warmup. (commit: fabf1fd)
- Remove sm_50 from cuda_clang build configs to match TF (commit: 40c49d7)
- Log the old and new configs to ERROR when we fail to accept the transition. (commit: b9b5f52)
- Porting the grappler batch prioritization rewriter to TFRT. (commit: 429ddb2)
- Porting the grappler batch prioritization rewriter to TFRT. (commit: ca86c54)
- Bump bazel version from 6.5.0 to 7.4.1 to align with TF.
(commit: 7491ad5) - Increase Bazel version requirement to 7.4.1 in root.workspace (commit: b16e36f)
- Flip back incompatible_enable_cc_toolchain_resolution to bypass cpp toolchain resolution error. (commit: 96718f3)
- Porting the grappler batch prioritization rewriter to TFRT. (commit: dc7f1bb)
- Extract the untracked part of
RunSavedModelWarmupto a separate function (commit: a419266) - Enable use of absl nullability annotations now that the absl version used provides them. (commit: 482cdcd)
- Update release notes for 2.19.1 (commit: 89dfd99)
- Add a command line flag
remote_predict_op_use_tensor_contentto reduce request size (commit: 1d1a029) - changing Orbax Model version file uniquely identifiable (commit: a01a31a)
- Plumb priority into RunHandlerPool (commit: 4184bff)
- Fix a check failure in ServerCore. (commit: 223ac59)
- Add
max_cache_lengthtoPredictRequest.RequestOptions. (commit: 6db34f8) - Add
max_cache_lengthtoPredictRequest.RequestOptions. (commit: 6d2fd08) - Add max_cache_length to PredictRequest.RequestOptions. (commit: 1c88766)
- Use std::numeric_limits for max int32 value. (commit: a33d9da)
- Removal of tsl-specific integral types. (commit: d1f407f)
- Add a new sampling mode to infra logging. (commit: b05d65f)
- TF Serving Dockerfiles upgraded for a hermetic build. MKL stays non-hermetic at this time because of strict linker checks in llvm_openmp. (commit: ecb468f)
- Fixed Serving server startup in docker environment (commit: 275ec5c)
- Optimized disk space usage in Dockerfile.devel-gpu (commit: 824ad7a)
- Add option to allow resplitting for priority aware scheduler. (commit: 9baa5a2)
- add mutable version which can save a copy (commit: a5f6f0b)
- Update version for 2.20.0 release. (#4135) (commit: bc7e9d2)