March 2026 monthly summary for modelscope/ms-swift. This period focused on improving data visibility for NPU performance and refining training workflows on the Mindspeed backend. Key features delivered include documentation improvements to clarify NPU performance data collection in Megatron-SWIFT and training script enhancements for Qwen3 Omni on Mindspeed. No major bugs fixed this month. Overall impact includes clearer performance data guidance, smoother training workflow, and better alignment between code paths and data collection logic, enabling faster decision-making and more reliable performance analyses.

Feb 2026 — NPU Profiling Performance Enhancements in verl: Optimized GPU memory utilization and introduced a checkpoint engine configuration to boost profiling performance and accuracy. Addressed flaky profiling scripts with a targeted fix, stabilizing benchmarks and enabling faster iteration. This work reduces memory pressure, accelerates performance tuning, and improves reliability of profiling data across deployments.

December 2025 monthly summary for modelscope/ms-swift: Delivered key enhancements to the Mindspeed-backed Qwen3-Omni training workflow and improved NPU reliability. The new fine-tuning script enables streamlined, parameterized training for Qwen3-Omni on Mindspeed backend, facilitating faster experimentation and production-ready deployments. A compatibility fix bypasses the mcore version check when Torch NPU is available, eliminating version-mismatch errors on NPU hardware and reducing setup friction. Together, these efforts improve end-to-end model customization throughput, stability of NPU-backed workflows, and cross-hardware reliability.

November 2025 monthly work summary for modelscope/ms-swift focusing on delivering NPU-accelerated deployment optimizations and operator fusion for modeling_qwen2.

During August 2025 for volcengine/verl, two profiling-focused features were delivered with strengthened test coverage and discrete-mode profiling enabled across NPU and NsightSystems. The work enhanced profiling reliability, validated cross-device performance on Ascend devices, and established a solid foundation for future performance optimizations.

July 2025: Delivered NPU Profiling for VERL framework FSDP backend, enhancing performance visibility on NPU devices and enabling data-driven optimizations. Consolidated profiler configuration into a single actor_rollout_ref.profiler, added optional role selection in discrete mode, and refreshed documentation and tooling to streamline profiling workflows. These changes improved data collection efficiency and reduced profiling parsing time, enabling faster turnaround on performance analyses.