Major Features and Improvements
- No major features or improvements.
Breaking Changes
- No breaking changes.
Bug Fixes and Other Changes
- Increase memory limit for ASan builds of py_binary targets. (commit: 34ba5df)
- Fix build issue. (commit: b1756b6)
- Increase cpp_link.mem for ASAN builds to fix OOM failures. (commit: de4ebf0)
- Add RPC deadline and cancellation signal to Run Options for task-level cancellation in batching scheduler. (commit: 14a8723)
- Use portpicker to pick unused ports in model server tests. (commit: 192efc5)
- Update release notes for TensorFlow Serving 2.20.0. (commit: 9903247)
- Add option to enable async IFRT in TFRT SavedModel. (commit: d802d85)
- Updates the TensorFlow dependency to fixbuild failures (commit: 62f7e45)
- Add option for enable_batching_task_lazy_cancellation (commit: ffee590)
- Add HERMETIC_NCCL_VERSION="2.29.7" to the cuda_clang bazelrc config. (commit: c005b7d)
- Use
commoninstead ofbuildfor hermetic CUDA repo_env flags (commit: ebab0b2) - Make RequestLogger::UniformSampler thread-safe to prevent race conditions during concurrent logging. (commit: 849fd89)
- Replace std::mt19937 with absl::BitGen in RequestLogger's UniformSampler. (commit: 3918641)
- Zero-init dummy values in gzip_zlib (commit: d2d02d7)
- Align TensorFlow Serving dependencies and toolchains to TF 2.21 release. (commit: d442e7d)
- Fix TensorFlow 2.21 build in root.workspace, resolve dso_loader header dependency under CUDA, and update devel Dockerfile logging and flags. (commit: e9057d7)
- Pass PredictRequest by pointer in PredictStreamedContext::ProcessRequest callers. (commit: b4a35ef)
- Add SetResponsesCompleteCallback to PredictStreamedContext interface to allow servables to notify handlers when response generation is complete. (commit: 5782006)
- Migrate ABSL_MUST_USE_RESULT to C++17 [[nodiscard]] (commit: afbc8b3)
- Clean up const PredictRequest& ProcessRequest API. (commit: 254a63c)
- Update TF Text to v2.21.1. (commit: aaf360b)
- Pin TensorFlow back to the 2.21.0 release commit. (commit: 880a735)
- Support shortcutting RPC during model warmup in TFServingRemotePredict op. (commit: caffa2d)
- Verify both valid GitHub SHA256 hashes for llvm-raw archive in TF 2.21.0 and protect org_tensorflow pin from auto-releaser overwrite. (commit: b677136)
- Upgrade log level for "no versions found" from WARNING to ERROR. (commit: 9fdd3b8)
- Regenerate OSS prediction_service_pb2.py and model_service_pb2.py with Protobuf 6.31.1 for TensorFlow 2.21.0 compatibility. (commit: 4f4f484)
- Update version for 2.21.0 release. (#4152) (commit: 75c8de0)
- This release is based on TF version 2.21.0.