Issues

• [TPU] Silent BF16 gather corruption in T16-to-T8 fusion with debug callback (libtpu 0.0.48)

  openxla/xla#51205 · Filter by author fdz-1999 opened on Oct 20, 2026
  2

• [XLA:GPU] triton-gemm-rewriter moves a layout-changing copy below a bitcast; the f64 dot then reads a transposed operand

  openxla/xla#51186 · Filter by author toshiki-ms opened on Oct 20, 2026

• AllGatherOptimizer misses moving subtraction before compatible all-gathers

  openxla/xla#51155 · Filter by author nikokurumi opened on Oct 19, 2026

• [CPU] x / broadcast(r) is rewritten to x * broadcast(1/r) without fast math: ~29% of float32 quotients are 1 ulp off

  openxla/xla#51148 · Filter by author AshishKumar4 opened on Oct 19, 2026

• live_devices + more than one GPU per process causes a data race on GpuExecutableRunOptions::incarnations_

  openxla/xla#51142 · Filter by author ymlei opened on Oct 19, 2026

• [XLA:CPU] Wrong results from a small loop fusion: LLVM LoopStrengthReduce adds an incorrect nuw flag

  openxla/xla#51048 · Filter by author etiferrier opened on Oct 18, 2026
  1

• [XLA:GPU] Triton GEMM fusion computes some int8 dots as 2x or 4x the product; with CUDA 13, cuBLASLt then refuses them under deterministic ops

  openxla/xla#51016 · Filter by author AshishKumar4 opened on Oct 17, 2026

• CPU: jit returns inf/NaN/~1e38 garbage for a bfloat16 graph with reductions fused by YNNPACK (__ynn_fusion); eager is correct

  openxla/xla#51014 · Filter by author carlosgmartin opened on Oct 17, 2026

• [GPU] Autotuning makes a small broadcast-multiply-reduce fusion take ~1 minute to compile

  openxla/xla#51012 · Filter by author carlosgmartin opened on Oct 17, 2026

• [XLA:CPU] jaxlib 0.11.2: a chain of dynamic-update-slice adds in a loop is 1.3–2.8× slower than in 0.11.1

  openxla/xla#50977 · Filter by author etiferrier opened on Oct 16, 2026
  1

• [XLA:GPU] HostOffloader CHECK-fails when a host-offloaded buffer is the data operand of send (host_offloader.cc:351)

  openxla/xla#50971 · Filter by author YuanchengJiang opened on Oct 16, 2026

• [XLA:GPU] HloComputationDeduplicator makes a non-inlined call target also a while body, then CollectivePipeliner's dataflow analysis CHECK-fails (hlo_dataflow_analysis.cc:1479)

  openxla/xla#50970 · Filter by author YuanchengJiang opened on Oct 16, 2026

• [XLA:GPU] Triton one-shot all-reduce fails to compile for more than 32 ranks: "size mismatch when packing elements for LLVM struct"

  bugerr:BuildGPUNVIDIA-GPU
  openxla/xla#50877 · Filter by author ksew1 opened on Oct 15, 2026

• [ROCm] GPU sort returns corrupted results on gfx1151 (Strix Halo) for 1k-16k elements

  AMD-GPUbugGPU
  openxla/xla#50861 · Filter by author FIM43-Redeye opened on Oct 15, 2026
  

• [XLA:GPU] Triton normalization fusion fails on a reduce whose reducer returns a parameter (emitter_helpers.cc:1295 !to_emit.empty())

  bugGPUNVIDIA-GPU
  openxla/xla#50830 · Filter by author YuanchengJiang opened on Oct 15, 2026

• [XLA:GPU] Constants inside a _xla_stream_annotation call are never sanitized: ConstantNameToGlobalName DCHECK fails (buffer_assignment_util.cc:80)

  bugGPU
  openxla/xla#50829 · Filter by author YuanchengJiang opened on Oct 15, 2026

• [XLA:GPU] HloComputation::ReplaceInstruction / ReplaceWithNewInstruction hit DCHECK(changed) when the replaced instruction has control dependencies (DegenerateDimensionRewriter, ReshapeMover)

  bugGPUNVIDIA-GPU
  openxla/xla#50826 · Filter by author YuanchengJiang opened on Oct 15, 2026

• [XLA:GPU] HLO text parser CHECK-fails on a shape-less get-tuple-element whose operand is not a tuple or whose index is out of range

  bugerr: RuntimeGPU
  openxla/xla#50825 · Filter by author YuanchengJiang opened on Oct 15, 2026

• [XLA:GPU] [SPMD] Sharding propagation CHECK-fails (IsTiledLeaf()) when an elementwise op's operands carry {unreduced} sharding

  bugGPUshardingstat:awaiting openxla-eng
  openxla/xla#50824 · Filter by author YuanchengJiang opened on Oct 15, 2026
  1

• [XLA] AllGatherBroadcastReorder misses reordering when only the gather dimension is broadcasted

  bugGPUstat:awaiting openxla-eng
  openxla/xla#50671 · Filter by author nikokurumi opened on Oct 13, 2026
  1

• [GPU] CUDA illegal address during fusion-emitter autotuning while compiling an f64 module (jaxlib 0.8.1, A100)

  bugGPUNVIDIA-GPU
  openxla/xla#50641 · Filter by author zekun-shi opened on Oct 13, 2026

• [XLA:CPU] Integer power ignores all exponent bits above bit 5 (wrong results for odd bases with exponent >= 64)

  bugCPUerr: Runtime
  openxla/xla#50550 · Filter by author YuanchengJiang opened on Oct 12, 2026

• [XLA] HloCSE CHECK-fails combining a duplicate constant that has control successors

  bugCPU
  openxla/xla#50549 · Filter by author YuanchengJiang opened on Oct 12, 2026

• [XLA:CPU] Deserialized executables always attach an HloProto copy, roughly doubling resident memory (ignores --xla_embed_ir_in_executable)

  bugCPUerr:performance
  openxla/xla#50471 · Filter by author firstprinciples opened on Oct 11, 2026
  1

• [XLA:GPU] Conv fusion compiles any scalar reduce after a convolution as cuDNN AMAX

  bugGPUNVIDIA-GPU
  openxla/xla#50470 · Filter by author CarloLucibello opened on Oct 11, 2026