feat(recipe): add qwen3-vl-8b NPU recipe with DAPO

[yidingdou]

Title: Adds Qwen3-VL-8B vision-language reasoning NPU recipe with DAPO on GEO3K

Why this recipe?

verl-recipe currently lacks vision-language (VL) reasoning recipes on NPU, despite the growing importance of multimodal reasoning.

This recipe is designed as VL reasoning with verifiable rewards:

Shows how run Qwen3-VL-8B dapo on NPU

Recipe Design

Model: Qwen3-VL-8B-Instruct (full-weight, NPU A2 * 2 nnodes)
Dataset: geo3k

Results

100 steps of DAPO on 2 A2(910B3) nnodes 64GB (~53 h):

算法: DAPO
Baseline (mean@1): 0.438
Test acc best (mean@1): 0.715

[gemini-code-assist]

Code Review

This pull request introduces a new bash script dapo/run_dapo_qwen3_vl_8b_fsdp2_npu.sh to run DAPO training for Qwen3-VL-8B on NPU using FSDP2. The review feedback highlights a critical bug where an undefined variable is referenced, which would cause the script to crash due to set -u. Additionally, it recommends using the + prefix for appending new Hydra configuration keys and utilizing the defined $total_training_steps variable instead of hardcoding its value.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.