[recipe, diffusion] fix: let Ray set Ascend visible devices in Qwen-Image NPU script

[Sky-Trigger]

What does this PR do?

This PR removes RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1 from the Qwen-Image OCR LoRA NPU FlowGRPO example script:

examples/flowgrpo_trainer/qwen_image/run_qwen_image_ocr_lora_npu.sh

With this environment variable set, Ray skips rewriting ASCEND_RT_VISIBLE_DEVICES for Ascend/NPU workers. In this FlowGRPO Qwen-Image NPU recipe, this can lead to device placement mismatch across workers.

Observed error:

AssertionError: Expects tensor to be on the compute device npu:15, was on npu:0

Removing this override allows Ray to manage the per-worker Ascend device visibility normally, which fixes the device mismatch in the NPU FlowGRPO training script.

Why is this needed?

RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1 is useful in some vLLM-Ascend distributed serving setups where device mapping is handled outside Ray. However, for this verl-omni FlowGRPO training recipe, keeping it enabled prevents Ray from setting the expected per-worker device visibility and may cause tensors to be created on the wrong NPU.

This change keeps the example script aligned with Ray-managed NPU placement and avoids the runtime device assertion failure.

Test

Tested by running the Qwen-Image OCR LoRA NPU FlowGRPO example.

Before this change, the script failed with:

AssertionError: Expects tensor to be on the compute device npu:15, was on npu:0

After removing RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1, the script can proceed without this device mismatch error.

[gemini-code-assist]

Code Review

This pull request removes the environment variable export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1 from the run_qwen_image_ocr_lora_npu.sh script. There are no review comments to evaluate, and I have no additional feedback to provide on this change.

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.