Safeq qwen test

[Matvezy]

What does this PR do?

Changed qwen to estimate if flash attn is available more safely, since it could break on t4

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Testing

• I have tested this change locally

Test details:
n/a

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context

---

Note

Improves robustness of attention backend selection for Qwen3-VL.

• Adds _get_qwen3vl_attn_implementation to return flash_attention_2 only if it’s available and importable on CUDA; otherwise falls back to eager
• Replaces inline attention selection with the new helper and passes attn_implementation to all from_pretrained model loading paths

^{Written by Cursor Bugbot for commit 2a024cc. This will update automatically on new commits. Configure here.}