Bring this branch up to current main (DSA path moved out of nvidia/)

[drakosha]

Brings this branch up to current main (da329cc30) so it merges cleanly again.

Your two commits stopped applying because the fused DSA path moved in August: vllm/models/deepseek_v32/nvidia/attention.py is now deepseek_v32/attention.py, and nvidia/kernels.py is now common/kernels.py. Nothing else about the fix changed. Your regression test came along and now lives next to the existing cases in tests/kernels/test_fused_deepseek_v32_norm_rope.py.

It is a merge commit, so it applies to your branch without conflicts. After merging, vllm-project#50005 loses needs-rebase and shows three files, 110 lines against main. The diff GitHub displays here is large only because the branch is catching up with a month of main.

Verified on 4xH200 with GLM-5.2-NVFP4, TP4 + DCP4, fp8_ds_mla:

• DCP=4 output is byte-identical to DCP=1 under greedy decoding. Unpatched main emits repeated-token garbage (locklocklock...) on the same prompt.
• pytest tests/kernels/test_fused_deepseek_v32_norm_rope.py: 67 passed with the fix. Without it, your new case is the only failure (1 failed, 66 passed), so it does catch the bug it describes.

Authorship on the commit is yours; I signed off as the submitter. AI assistance was used for the rebase and the test runs.

[foraxe]

Thanks @drakosha - reviewed and merged. I confirmed that upstream vllm-project#50005 is now mergeable and shows only the three intended files. I also reran the focused suite on GB200 (67 passed) and all applicable changed-file pre-commit hooks passed. The upstream PR and issue descriptions now use the current two-defect scope and credit your H200 validation.