Issues

• float32 tf.matmul on rank-3 tensors returns wrong, nondeterministic results on CPU (oneDNN) for batch 2-16 and contraction >= 20000

  type:bug
  tensorflow/tensorflow#129498 · Filter by author italo-goncalves opened on Oct 23, 2026
  

• overlap_and_add returns non-empty tensor for 0-frame inputs when frame_length > frame_step

  type:bug
  tensorflow/tensorflow#129484 · Filter by author Abhirup0 opened on Oct 23, 2026
  

• tf.get_static_value() never folds StridedSlice, so tf.shape(x)[0] has no static value in graph mode

  2.21.0awaiting PR mergecomp:coretype:bug
  tensorflow/tensorflow#129438 · Filter by author Phoebus-Liu opened on Oct 22, 2026
  

• tf.raw_ops.For aborts on GPU for a valid float32 loop while the CPU execution succeeds

  2.21.0comp:gpucomp:opsstat:awaiting tensorflowertype:bug
  tensorflow/tensorflow#129396 · Filter by author lv20031001 opened on Oct 21, 2026
  1

  

• tf.raw_ops.SparseMatrixAdd with complex128 on GPU causes a segmentation fault

  2.21.0comp:opsstat:awaiting tensorflowertype:bug
  tensorflow/tensorflow#129392 · Filter by author lv20031001 opened on Oct 21, 2026
  1

  

• ImportEvent aborts the process for a multi-element event tensor instead of returning an error

  2.21.0awaiting PR mergecomp:opstype:bug
  tensorflow/tensorflow#129384 · Filter by author lv20031001 opened on Oct 21, 2026
  1

  

• ResourceAccumulatorApplyGradient aborts on accumulator/gradient dtype mismatch

  2.21.0awaiting PR mergecomp:opstype:bug
  tensorflow/tensorflow#129358 · Filter by author lv20031001 opened on Oct 20, 2026
  1

  

• RealDiv accepts integer dtypes in its OpDef, but eager/graph have no integer kernel and reject, while XLA (jit_compile=True) silently does truncating integer division — a dtype-support / semantic-consistency gap

  2.21.0comp:xlastat:awaiting tensorflowertype:bug
  tensorflow/tensorflow#129335 · Filter by author tinywisdom opened on Oct 20, 2026
  2

  

• [Numerical accuracy] Conv2DBackpropInput float32 deviation under severe cancellation on CPU

  comp:opsTF 2.16type:bug
  tensorflow/tensorflow#129166 · Filter by author L4AIsS opened on Oct 16, 2026
  1

  

• [XLA] UniformQuantize uses incorrect rounding at half-integer values

  2.21.0awaiting PR mergecomp:xlatype:bug
  tensorflow/tensorflow#129087 · Filter by author xyt66665000 opened on Oct 14, 2026
  1

  

• [XLA][CPU] atan2 returns incorrect quadrants for float64 subnormal inputs

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#129083 · Filter by author xyt66665000 opened on Oct 14, 2026
  1

  

• tf.linalg.set_diag with a band fails to trace when the input's dimensions aren't all known

  2.21.0awaiting PR mergecomp:opstype:bug
  tensorflow/tensorflow#129061 · Filter by author vishwakt opened on Oct 13, 2026
  1

  

• TF 2.20/2.21 load the runtime with RTLD_GLOBAL, exposing bundled BoringSSL/curl process-wide; 2.22.0rc0 is RTLD_LOCAL again. Please keep it (and consider 2.21.x)

  2.21.0subtype: ubuntu/linuxtype:bugtype:build/install
  tensorflow/tensorflow#129050 · Filter by author tyhust opened on Oct 13, 2026
  

• [XLA] betainc returns NaN at endpoints with infinite shape parameters

  2.21.0comp:xlastat:contribution welcometype:bug
  tensorflow/tensorflow#128991 · Filter by author xyt66665000 opened on Oct 12, 2026
  3

  

• [XLA] Complex log loses infinite real component for NaN + Inf*i

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128990 · Filter by author xyt66665000 opened on Oct 12, 2026
  2

  

• [XLA] Incorrect special-value handling in igamma and igammac

  2.21.0awaiting PR mergecomp:xlatype:bug
  tensorflow/tensorflow#128989 · Filter by author xyt66665000 opened on Oct 12, 2026
  2

  

• [XLA] Complex exp(-Inf + Inf*i) returns NaN on CPU XLA and GPU

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128988 · Filter by author xyt66665000 opened on Oct 12, 2026
  6

  

• [XLA] Zeta returns NaN instead of zero for large finite exponent

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128987 · Filter by author xyt66665000 opened on Oct 12, 2026
  1

  

• [XLA] reduce_min returns +0.0 instead of -0.0

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128986 · Filter by author xyt66665000 opened on Oct 12, 2026
  1

  

• [XLA] Complex sin/cos produce NaN for infinite imaginary inputs

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128985 · Filter by author xyt66665000 opened on Oct 12, 2026
  2

  

• [XLA] betainc returns NaN for valid large finite parameters on CPU and GPU

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128984 · Filter by author xyt66665000 opened on Oct 12, 2026
  1

  

• [XLA] Complex log1p(-1+0j) returns zero instead of -Inf on CPU and GPU

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128983 · Filter by author xyt66665000 opened on Oct 12, 2026
  2

  

• [XLA] Small-range float32 stateless_uniform collapses to a constant on CPU and GPU

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128963 · Filter by author xyt66665000 opened on Oct 12, 2026
  1

  

• [XLA] Low-precision tf.random.uniform produces the excluded upper bound

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128960 · Filter by author xyt66665000 opened on Oct 12, 2026
  1

  

• [XLA] Far-tail truncated normal sampling collapses to the lower bound

  2.21.0comp:xlatype:bug
  tensorflow/tensorflow#128959 · Filter by author xyt66665000 opened on Oct 12, 2026
  1