v2.19.0-rc0
Breaking Changes📦 tensorflowView on GitHub →
⚠ 2 breaking✨ 1 features🐛 1 fixes⚡ 1 deprecations🔧 5 symbols
Summary
TensorFlow 2.19.0 adds bfloat16 support for the tfl.Cast op, deprecates tf.lite.Interpreter in favor of ai_edge_litert.interpreter, and changes two interpreter constants to const references, requiring code updates.
⚠️ Breaking Changes
- The public constants tflite::Interpreter::kTensorsReservedCapacity and tflite::Interpreter::kTensorsCapacityHeadroom are now const references instead of constexpr, breaking code that expects compile‑time constants.
- tf.lite.Interpreter now emits a deprecation warning and will be removed in TensorFlow 2.20; code must import ai_edge_litert.interpreter instead.
Migration Steps
- Replace imports of tf.lite.Interpreter with ai_edge_litert.interpreter as described in the migration guide.
- If code relied on the constexpr values of tflite::Interpreter::kTensorsReservedCapacity or kTensorsCapacityHeadroom, update it to use the new const reference values (e.g., access them at runtime rather than as compile‑time constants).
✨ New Features
- tfl.Cast operation now supports bfloat16 in the runtime kernel.
🐛 Bug Fixes
- Publishing of libtensorflow packages has been stopped; the library can still be extracted from the PyPI package.
🔧 Affected Symbols
tflite::Interpreter::kTensorsReservedCapacitytflite::Interpreter::kTensorsCapacityHeadroomtf.lite.Interpreterai_edge_litert.interpretertfl.Cast⚡ Deprecations
- tf.lite.Interpreter is deprecated and will be removed in TF 2.20; use ai_edge_litert.interpreter.