Sandbox::https://github.com/tensorflow/tensorflow/issues

tf.keras.layers.GroupNormalization` large-constant beta offset loss

tensorflow/tensorflow#128754 · opened

on Oct 10, 2026

tf.keras.layers.BatchNormalization(training=True) `drops the` beta` offset for large constant batches

tensorflow/tensorflow#128753 · opened

on Oct 10, 2026

tf.keras.layers.LSTM `fails on zero-batch inputs while` SimpleRNN `and` GRU` return the expected empty output

tensorflow/tensorflow#128751 · opened

on Oct 10, 2026

1

tf.keras.layers.LSTM `fails on zero-batch inputs while` SimpleRNN `and` GRU` return the expected empty output

tensorflow/tensorflow#128750 · opened

on Oct 10, 2026

1

tf.keras.layers.Add`,` Average`, and` keras.ops.add `fail on rank-2 zero-batch tensors with` Biases must be 1D

tensorflow/tensorflow#128749 · opened

on Oct 10, 2026

1

GlobalMaxPooling1D/2D/3D `raise in eager mode for zero-batch inputs while` tf.function` returns the expected shape

tensorflow/tensorflow#128748 · opened

on Oct 10, 2026

1

oneDNN CPU convolution returns wrong output channel dimension for zero-batch inputs in eager mode

tensorflow/tensorflow#128746 · opened

on Oct 10, 2026

1

tf.keras.layers.Attention `and` AdditiveAttention` return nonzero outputs when all value positions are masked out

tensorflow/tensorflow#128745 · opened

on Oct 10, 2026

1

tf.keras.layers.LayerNormalization `returns different results in eager mode and` tf.function `for large float32 inputs with` center=True`

tensorflow/tensorflow#128743 · opened

on Oct 10, 2026

1

TensorFlow AveragePooling2D rejects int64 accepted by Torch

tensorflow/tensorflow#128733 · opened

on Oct 10, 2026

Saturated float32 softmax produces a nonzero gradient with oneDNN disabled for some tensor shapes

tensorflow/tensorflow#128728 · opened

on Oct 10, 2026

Eager Conv3DBackpropFilterV2 with oneDNN accepts 4 strides and returns a 4-D result

tensorflow/tensorflow#128647 · opened

on Oct 8, 2026

1

[RFC] Integrate KDNN for Linux ARM64 CPU Acceleration

tensorflow/tensorflow#128639 · opened

on Oct 8, 2026

_MklFusedInstanceNorm fusion with scalar scale/offset gives wrong channels in tf.function (oneDNN)

tensorflow/tensorflow#128638 · opened

on Oct 8, 2026

1

LeakyRelu remapper fuses maximum(x, alpha * x) for alpha > 1, wrong values in tf.function

tensorflow/tensorflow#128637 · opened

on Oct 8, 2026

1

_MklLayerNorm fusion uses the default epsilon for a hand-written layer norm, wrong results in tf.function

tensorflow/tensorflow#128636 · opened

on Oct 8, 2026

1

RemoveLogicalNot rewrites LogicalNot of float ordering comparisons, changing NaN results in tf.function

tensorflow/tensorflow#128635 · opened

on Oct 8, 2026

1

tf.while_loop preserves a 0-D NumPy loop variable in eager but tensorizes it under tf.function, changing body semantics

tensorflow/tensorflow#128634 · opened

on Oct 8, 2026

1

tf.image.non_max_suppression fails XLA compilation for legal float16 inputs with mixed T / T_threshold dtypes

tensorflow/tensorflow#128614 · opened

on Oct 8, 2026

1

tf.tensor_scatter_nd_add accepts rank-0 scalar `updates` under jit_compile=True while eager rejects them

tensorflow/tensorflow#128613 · opened

on Oct 8, 2026

1

N-D FFT kernel reads axes_shape[fft_rank - 1] when fft_rank is 0 (rank 0 input to tf.signal.fftnd and friends)

tensorflow/tensorflow#128588 · opened

on Oct 7, 2026

1

Optimize Concat GPU Performance: Eliminate Multi-Kernel Launch Overhead (N < 16), Add Contiguous Axis-0 Fast-Path, and Avoid Dynamic PCIe Marshalling

tensorflow/tensorflow#128543 · opened

on Oct 6, 2026

1

Incorporate flash attention natively to tensorflow

tensorflow/tensorflow#128509 · opened

on Oct 5, 2026

`tf.experimental.numpy.isclose` and `allclose` return inverted results for infinities

tensorflow/tensorflow#128480 · opened

on Oct 4, 2026

2

`tf.experimental.numpy.cbrt(-0.0)` returns `+0.0`

tensorflow/tensorflow#128479 · opened

on Oct 4, 2026

1

Navigation Menu

Repository navigation

Issues

Issues

tf.keras.layers.GroupNormalization` large-constant beta offset loss

tf.keras.layers.BatchNormalization(training=True) `drops the` beta` offset for large constant batches

tf.keras.layers.LSTM `fails on zero-batch inputs while` SimpleRNN `and` GRU` return the expected empty output

tf.keras.layers.LSTM `fails on zero-batch inputs while` SimpleRNN `and` GRU` return the expected empty output

tf.keras.layers.Add`,` Average`, and` keras.ops.add `fail on rank-2 zero-batch tensors with` Biases must be 1D

GlobalMaxPooling1D/2D/3D `raise in eager mode for zero-batch inputs while` tf.function` returns the expected shape

oneDNN CPU convolution returns wrong output channel dimension for zero-batch inputs in eager mode

tf.keras.layers.Attention `and` AdditiveAttention` return nonzero outputs when all value positions are masked out

tf.keras.layers.LayerNormalization `returns different results in eager mode and` tf.function `for large float32 inputs with` center=True`

TensorFlow AveragePooling2D rejects int64 accepted by Torch

Saturated float32 softmax produces a nonzero gradient with oneDNN disabled for some tensor shapes

Eager Conv3DBackpropFilterV2 with oneDNN accepts 4 strides and returns a 4-D result

[RFC] Integrate KDNN for Linux ARM64 CPU Acceleration

_MklFusedInstanceNorm fusion with scalar scale/offset gives wrong channels in tf.function (oneDNN)

LeakyRelu remapper fuses maximum(x, alpha * x) for alpha > 1, wrong values in tf.function

_MklLayerNorm fusion uses the default epsilon for a hand-written layer norm, wrong results in tf.function

RemoveLogicalNot rewrites LogicalNot of float ordering comparisons, changing NaN results in tf.function

tf.while_loop preserves a 0-D NumPy loop variable in eager but tensorizes it under tf.function, changing body semantics

tf.image.non_max_suppression fails XLA compilation for legal float16 inputs with mixed T / T_threshold dtypes

tf.tensor_scatter_nd_add accepts rank-0 scalar `updates` under jit_compile=True while eager rejects them

N-D FFT kernel reads axes_shape[fft_rank - 1] when fft_rank is 0 (rank 0 input to tf.signal.fftnd and friends)

Optimize Concat GPU Performance: Eliminate Multi-Kernel Launch Overhead (N < 16), Add Contiguous Axis-0 Fast-Path, and Avoid Dynamic PCIe Marshalling

Incorporate flash attention natively to tensorflow

`tf.experimental.numpy.isclose` and `allclose` return inverted results for infinities

`tf.experimental.numpy.cbrt(-0.0)` returns `+0.0`