PT2.16

Issues

• [RFC] Migrate torch-xpu-ops DeviceProperties.h to XPUContext.h

  module: rfctest: ut
  Task

  intel/torch-xpu-ops#5245 · Filter by author Stonepia opened on Sep 8, 2026 · PT2.16
  2

  

• XPU Autocast does not support fp32 dtypes

  dtype: float32module: coretest: ut
  Feature

  intel/torch-xpu-ops#2680 · Filter by author kaixuanliu opened on Jan 5, 2026 · PT2.16
  8

  

• torch xpu event has ~0.1ms latency, which is too large

  dependency component: oneAPImodule: coreperformancetest: ut
  Bug

  intel/torch-xpu-ops#3227 · Filter by author jianyizh opened on Mar 31, 2026 · PT2.16
  6

  

• Host may stuck when submit too many kernels when event recording

  dependency component: driverfunctionalitymodule: coretest: ut
  Bug

  intel/torch-xpu-ops#2467 · Filter by author jianyizh opened on Dec 5, 2025 · PT2.16
  2

  

• Enable FP8/MXFP8 Ops with requests and CUDA alignment

  agent:needs-humandtype: float8dtype: mxfp8module: infratest: ut
  Task

  intel/torch-xpu-ops#2207 · Filter by author CuiYifeng opened on Oct 23, 2025 · PT2.16
  9

  

• The operator 'aten::_grouped_mm' is not currently implemented for the XPU device.

  functionalitymodule: aotest: ut
  Feature

  intel/torch-xpu-ops#1574 · Filter by author githubsgi opened on Apr 11, 2025 · PT2.16
  5

  

• Implement torch._scaled_grouped_mm for xpu backend

  module: aotest: ut
  Feature

  intel/torch-xpu-ops#3060 · Filter by author kgajdamo opened on Mar 16, 2026 · PT2.16
  2

  

• [BMG] PyTorch 2.12 performance cases where inductor is slower than eager (speedup < 1)

  dtype: bfloat16hw: BMGinferencemodule: inductorperformancetest: e2e
  Bug

  intel/torch-xpu-ops#3761 · Filter by author mengfei25 opened on May 27, 2026 · PT2.16
  5

  

• [TorchAO] Float8 training support on XPU

  module: aotest: ut
  Feature

  intel/torch-xpu-ops#2325 · Filter by author liangan1 opened on Nov 11, 2025 · PT2.16
  4

  

• [TorchAO] MX training native PyTorch on XPU

  module: aotest: ut
  Feature

  intel/torch-xpu-ops#2326 · Filter by author liangan1 opened on Nov 11, 2025 · PT2.16
  4

  

• [TorchAO] MOE training enabling on XPU

  dependency component: oneDNNmodule: aotest: uttraining
  Feature

  intel/torch-xpu-ops#2323 · Filter by author liangan1 opened on Nov 11, 2025 · PT2.16
  4

  

• Torchbench model functorch_dp_cifar10 got fail_accuracy

  Accuracyhw: BMGhw: PVCneed_splitos: Windowstest: e2e
  Bug

  intel/torch-xpu-ops#1877 · Filter by author libohao1201 opened on Jul 23, 2025 · PT2.16
  3

  

• [release/2.12] performance regressions compared with 2.11

  dependency component: communitydtype: amp_fp16hw: PVCinferencemodule: inductorperformanceregressiontest: e2e
  Bug

  intel/torch-xpu-ops#3432 · Filter by author mengfei25 opened on Apr 23, 2026 · PT2.16
  14

  

• Timm_efficientdet NotImplementedError: The original model code forces the use of CUDA.

  Accuracydependency component: third_partydtype: amp_bf16dtype: amp_fp16dtype: bfloat16dtype: float16dtype: float32duplicatefunctionalitymodule: infratest: e2etrainingtriaged
  Bug

  intel/torch-xpu-ops#492 · Filter by author mengfei25 opened on Jun 27, 2024 · PT2.16
  5

  

• [cpp extension] Provide a clear error message when using inconsistent oneapi versions.

  dependency component: oneAPIfunctionalitymodule: buildtest: ut
  Bug

  intel/torch-xpu-ops#1649 · Filter by author ZhaoqiongZ opened on May 9, 2025 · PT2.16
  9

  

• [TorchAO][BMG] Llama-3.2-1B-Instruct Dynamic INT8 got 10% performance drop with oneDNN 3.11.1

  dependency component: oneDNNdtype: int8hw: BMGmodule: aoperformanceregressiontest: ut
  Bug

  intel/torch-xpu-ops#3076 · Filter by author LifengWang opened on Mar 18, 2026 · PT2.16
  2

  

• [OOB][Perf] aten::clone unnecessary memory copies on XPU (18 models affected)

  dtype: float16inferencemodule: opsperformancetest: oob
  Bug

  intel/torch-xpu-ops#3783 · Filter by author jianyizh opened on May 29, 2026 · PT2.16
  1

  

• [OOB] use sdpa fusion in eager models

  inferencemodule: sdpatest: oob
  Feature

  intel/torch-xpu-ops#3778 · Filter by author jianyizh opened on May 29, 2026 · PT2.16
  1

  

• [OOB] models missing sdpa fusion in torch compile inference

  inferencemodule: inductorperformancetest: oob
  Bug

  intel/torch-xpu-ops#3779 · Filter by author jianyizh opened on May 29, 2026 · PT2.16
  1

  

• [OOB][Perf] aten::addmm (GEMM) performance gap on XPU vs CUDA (62 models affected)

  dependency component: oneDNNdtype: float16inferencemodule: gemmperformancetest: oob
  Bug

  intel/torch-xpu-ops#3782 · Filter by author jianyizh opened on May 29, 2026 · PT2.16
  1

  

• [OOB][Perf] aten::cat performance gap on XPU vs CUDA (54 models affected)

  dtype: float16inferencemodule: opsperformancetest: oob
  Bug

  intel/torch-xpu-ops#3780 · Filter by author jianyizh opened on May 29, 2026 · PT2.16
  1

  

• [oneAPI][backward compatibility] libur_loader.so.0: version `LIBUR_LOADER_0.11' not found

  dependency component: oneAPIfunctionalitymodule: buildtest: ut
  Feature

  intel/torch-xpu-ops#2349 · Filter by author dvrogozh opened on Nov 12, 2025 · PT2.16
  1

  

• [release/2.12][PVC] BERT_pytorch amp_bf16 inference got fail_accuracy

  Accuracydtype: amp_bf16hw: PVCinferencemodule: inductortest: e2e
  Bug

  intel/torch-xpu-ops#3463 · Filter by author mengfei25 opened on Apr 24, 2026 · PT2.16
  2

  

• Register spill in XeFMHAFwdKernel on PVC (SIMD16, 256 GRF)

  enhancementhw: PVCmodule: sdpaperformancetest: ut
  Feature

  intel/torch-xpu-ops#3740 · Filter by author CuiYifeng opened on May 22, 2026 · PT2.16
  1

  

• Expose allowTF32OneDNN through the public fp32_precision API so XPU TF32 is controllable from device-agnostic code

  dtype: float32module: coretest: utut_upstream
  Feature

  intel/torch-xpu-ops#4919 · Filter by author daisyden opened on Aug 12, 2026 · PT2.16
  3