adjust atol value for mini_qwen2_vl FP32 in test_mini_model_multimodal

[faaany]

Summary

test_mini_model_multimodal has 2 test models: mini_mllama and mini_qwen2_vl. On XPU, mini_mllama cases pass but mini_qwen2_vl cases fail. In the 2 failed cases: one fails at 1e-2 decimal point, which we need to fix; but the other fails at 1e-6 decimal point, which we suspect that the atol is set too strict for non-cuda hardware accelerators as can be seen below:

FAILED test_mini_models_multimodal.py::test_mini_model_multimodal[mini_qwen2_vl-32-0.0001-dtype0-1e-08-1e-05-0.005-1e-05-0.005-1e-05] - AssertionError: Number of mismatched elements: 3
Mismatch at index (0, 21): tensor1[(0, 21)] = 0.03207247704267502, tensor2[(0, 21)] = 0.0320717953145504
Mismatch at index (0, 24): tensor1[(0, 24)] = 0.03104054555296898, tensor2[(0, 24)] = 0.03104141354560852
Mismatch at index (0, 27): tensor1[(0, 27)] = 0.04820457845926285, tensor2[(0, 27)] = 0.04820366948843002

This PR tries to fix this. What do you guys think?

• Hardware Type:
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

[faaany]

cc @yao-matrix and @sywangyi

[tyler-romero]

This LGTM but I'm going to pass to someone at LI to provide a final review: @shivam15s

[faaany]

close this PR, as it is included in #552