ultralytics 8.4.111

v8.4.111 - Add Huawei Ascend NPU training support (#25500)

on Python PyPI

🌟 Summary

🚀 Ultralytics 8.4.111 expands hardware support with validated Huawei Ascend NPU training, broader accelerator compatibility, and important tracking, MPS, deployment, and documentation improvements.

📊 Key Changes

• Huawei Ascend NPU training support 🧠

  • Added single- and multi-NPU training and validation through torch_npu.
  • Supports standard training features including AMP, checkpointing, resume, AutoBatch, profiling, memory management, and validation.
  • Multi-NPU training uses Huawei’s HCCL distributed backend.
  • Example device selections include device=npu:0 and device=npu:0,1.
  • Ascend documentation now covers the workflow from training through .om export and deployment.

• Unified accelerator handling ⚙️

  • Device behavior is now derived from the selected PyTorch device rather than being hard-coded for CUDA.
  • Improves compatibility with:
    • Huawei Ascend npu
    • Intel xpu
    • AMD ROCm through PyTorch’s standard CUDA-style device interface
  • Distributed training now selects the appropriate backend: NCCL for NVIDIA, HCCL for Ascend, and XCCL for Intel.

• AMD ROCm integration guide 🔴

  • Added documentation for training, validation, and inference on supported AMD GPUs using PyTorch ROCm.
  • Clarifies that ROCm uses device=0 or device=cuda:0, not device=rocm:0.
  • Clearly distinguishes supported ROCm workflows from currently unsupported or separate technologies such as MIGraphX, DirectML, and Ryzen AI NPU.

• Improved accelerator-aware data loading and profiling 📈

  • Dataloaders, pinned memory, synchronization, memory checks, profiling, automatic batch sizing, and mixed precision now account for more device types.
  • NPU and XPU execution avoids unsupported operations such as torchvision NMS kernels where necessary.

• Tracking and numerical stability improvements 🎯

  • Object counting now detects objects that pass through polygon regions between frames, even when no centroid is sampled inside the region.
  • Deep OC-SORT preserves homography precision.
  • Kalman filter states consistently use float64.
  • Re-identification features are stored safely as float32.
  • GMC matching now retains valid zero-variance and boundary-distance matches.
  • NumPy assignment now raises a clear error for infeasible cost matrices instead of potentially hanging.

• Apple MPS reliability fixes 🍎

  • Avoids problematic in-place operations on strided MPS tensors across affected macOS and PyTorch versions.
  • Prevents autocast crashes on MPS with PyTorch versions older than 2.5.

• Documentation and platform updates 📚

  • Added LabelMe dataset import instructions for converting offline annotations to YOLO format and uploading them to the Ultralytics Platform.
  • Updated Rust inference documentation for ultralytics-inference 0.0.32, including Intel CPU, GPU, and NPU device options.
  • Refreshed Raspberry Pi 5 YOLO26 benchmarks with LiteRT and clarified that LiteRT export must be performed off-device.
  • Added a monocular depth estimation tutorial to the YOLO26 documentation.
  • Normalized GPU names in usage telemetry to improve reporting consistency.

🎯 Purpose & Impact

• 🌍 Broader hardware choice: Users can train Ultralytics models on Huawei Ascend NPUs, AMD GPUs, Intel accelerators, and NVIDIA GPUs through more consistent device handling.
• 🚀 Easier enterprise and edge deployment: Ascend training now connects directly to existing export and deployment workflows, while AMD and Raspberry Pi guidance makes hardware-specific setup clearer.
• 🧩 Less backend-specific code: The shared PyTorch device abstraction reduces the need for separate trainers or parallel implementations for each accelerator.
• 📊 More dependable tracking: Object counters and trackers should behave more accurately in fast-motion, low-variance, and mixed-precision scenarios.
• 🛡️ Improved reliability: MPS fixes prevent crashes and inconsistent detections, while the assignment fallback now fails safely on impossible inputs.
• 📖 Better onboarding: New LabelMe, AMD, Ascend, Rust, Raspberry Pi, and depth resources help users move from dataset preparation to training and deployment more easily.
• 🔢 The package version is updated to 8.4.111.

What's Changed

• Document LabelMe Platform dataset import by @glenn-jocher in #25491
• Normalize GPU names in usage events by @glenn-jocher in #25503
• Bump ultralytics-inference version to 0.0.32 in documentation by @onuralpszr in #25502
• Refactor inference documentation to improve header placement and formatting by @onuralpszr in #25505
• Add https://youtu.be/i-V1kRCJD0M to docs by @RizwanMunawar in #25506
• Preserve the GMC homography precision in Deep OC-SORT by @raimbekovm in #25504
• Fix autocast crash on MPS with torch<2.5.0 by @Y-T-G in #25494
• Fix tracker docstring examples to run as written by @raimbekovm in #25509
• Keep gmc matches when the spatial distance variance is zero by @raimbekovm in #25507
• Declare the Kalman filter state dtype as float64 by @raimbekovm in #25499
• Fix PyTorch MPS strided tensor bugs in in-place operations by @Rahulbiradar9 in #25496
• Count objects that cross an ObjectCounter region between frames by @JESUSROYETH in #25492
• Store ReID appearance features as float32 by @raimbekovm in #25484
• Update Raspberry Pi 5 benchmarks with LiteRT by @lakshanthad in #25489
• Raise on an infeasible cost matrix in the NumPy linear_sum_assignment fallback by @ErenAta16 in #25508
• Add Huawei Ascend NPU training support by @glenn-jocher in #25500

New Contributors

• @Rahulbiradar9 made their first contribution in #25496
• @ErenAta16 made their first contribution in #25508

Full Changelog: v8.4.110...v8.4.111