February 2026 achieved measurable performance and reliability improvements across two repositories: (1) vllm-ascend delivered an adaptive block size selection for the linear_persistent kernel, optimizing throughput and latency for batch-invariant linear operations in LLM inference without API changes; (2) Verl stabilized MoE and async vLLM weight loading by fixing an AttributeError and ensuring post-load processing executes only after all weights are loaded, significantly reducing risk of partial updates during rollout; (3) Qwen3Next training support on NPU was added, including a training script that leverages FSDP and the VLLM backend to broaden hardware coverage and accelerate development. These changes were validated with targeted tests and aligned with CI/review goals, improving production performance, stability, and platform reach.

November 2025 monthly summary: Focused on delivering model support and stabilizing runtime patches across Verl and vLLM-Ascend. Key features include adding Qwen3NextForCausalLM support in Verl and ensuring robust runtime execution on Ascend devices via dynamic patch resolution. These changes reduce loading errors and improve cross-environment compatibility, enabling reliable experimentation and production workloads with Qwen3Next models.