[diffusion][hot fix] fix accuracy bug caused by PR 14717

[yingluosanqian]

Motivation

PR14717 introduces an accuracy issue that causes some models to generate incorrect images, as shown below.

This PR fix it. cc @mickqian @BBuf

flux_image_t2i

flux_2_klein_image_t2

This issue was caused by a change to the default dtype of the RMSNorm weight. After fixing it, the image quality returns to normal.

flux_image_t2i

flux_2_klein_image_t2

Also fix a cutedsl sync issue. Though the model accuracy it's not caused by this issue.

before fix:

E       AssertionError: Tensor-likes are not close!
E       
E       Mismatched elements: 75 / 353894400 (0.0%)
E       Greatest absolute difference: 0.59765625 at index (0, 36219, 342) (up to 0.05 allowed)
E       Greatest relative difference: 13.1875 at index (0, 36219, 480) (up to 0.05 allowed)

python/sglang/jit_kernel/tests/test_fused_norm_scale_shift.py:175: AssertionError
======================================== short test summary info =========================================
FAILED python/sglang/jit_kernel/tests/test_fused_norm_scale_shift.py::TestFusedScaleResidualNormScaleShift::test_dtype_1[dtype1-layer] - AssertionError: Tensor-likes are not close!
===================================== 1 failed, 147 passed in 24.70s =

after fix:

python/sglang/jit_kernel/tests/test_fused_norm_scale_shift.py .................................... [ 24%]
.................................................................................................. [ 90%]
..............                                                                                     [100%]

========================================== 148 passed in 25.08s ==========================================

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[yingluosanqian]

fast_hunyuan_video

before fix:

after fix:

[BBuf]

/tag-and-rerun-ci