June 2026 performance summary focused on strengthening NPU compatibility, expanding VoxCPM2 support across devices, and elevating deployment documentation for Qwen3-Omni. Key work included NPU-ready Qwen3-TTS Code2Wav initialization/config improvements, a bug fix for conv2d runtime on NPU, upgrading ModelRunner to v0.22.0 with compressed token scheduling, and introducing VoxCPM2 audio processing paths with cross-device encoder compatibility. Documentation updates for Qwen3-Omni-30B-A3B-Thinking deployments completed to improve onboarding and operational reliability across environments.

April 2026 (2026-04) focused on improving deployment reliability for Qwen3-Omni-30B via targeted documentation updates in the vllm-ascend repository. The work reduced risk of HcclAllreduce failures by clarifying required environment variables and aligned guidance with the vLLM main baseline, delivering clearer, more actionable instructions for users and maintainers.

March 2026 (2026-03) highlights quantization optimization and stability improvements for vLLM on Ascend NPU. Key deliverables include: Quantization Optimization for Qwen3-Omni on Ascend NPU with Auto-Quantization Tuning enhancements; multiple quantization and attention stability fixes across Qwen-Omni and ViT in Qwen2.5VL; and a multimodal embedding merge fix. These efforts improved quantization accuracy, stability, and performance, enabling more reliable deployment on Ascend hardware and reducing run-time errors.