[ci] chore: change two-node Ascend RayJob E2E workflow from A2 to A3

Author: wangdonglei

.github/workflows/e2e_ascend_multinode.yml

@@ -40,17 +40,17 @@ permissions:
 
 env:
   SUBMITTER_URL: http://rayjob-submitter.ci-system.svc.cluster.local
-  IMAGE: quay.io/ascend/verl:verl-9.0.0-910b-ubuntu22.04-py3.11-latest
+  IMAGE: swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/ascend-ci/verl/verl:verl-9.0.0-a3-ubuntu22.04-py3.11-latest
   WORKER_REPLICAS: "2"
 
 jobs:
   ascend-rayjob-e2e:
     if: github.repository_owner == 'verl-project'
     name: E2E Ascend RayJob multinode PPO
-    runs-on: linux-aarch64-a2b3-8
+    runs-on: linux-aarch64-a3-8
     timeout-minutes: 240
     container:
-      image: quay.io/albert54/actions-runner:ubuntu-24.04
+      image: swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/ascend-ci/verl/verl:verl-9.0.0-a3-ubuntu22.04-py3.11-latest
       options: >-
         --dns 10.96.0.10
     steps:

tests/special_npu/run_qwen3_06b_ppo_multinode.sh

@@ -50,7 +50,7 @@ python3 -m verl.trainer.main_ppo \
     trainer.logger='["console"]' \
     trainer.project_name='verl_ppo_example_gsm8k_qwen3' \
     trainer.experiment_name='qwen3_06b_fsdp_multinode_2x4' \
-    trainer.n_gpus_per_node=4 \
+    trainer.n_gpus_per_node=8 \
     trainer.nnodes=2 \
     trainer.save_freq=-1 \
     trainer.test_freq=-1 \