[Bug Fix] Fix vllm fused_add_rms_norm 6-arg crash on non-AITER path

[xiaohong42]

Motivation

When running SGLang on ROCm with SGLANG_USE_AITER=0, the vllm fallback path for fused_add_rms_norm crashes with:

TypeError: fused_add_rms_norm() takes 4 positional arguments but 6 were given

This is because the current code calls fused_add_rms_norm with 6 arguments (matching aiter's rmsnorm2d_fwd_with_add signature), but vllm's fused_add_rms_norm uses a 4-arg in-place API: fused_add_rms_norm(input, residual, weight, eps).

Modifications

• RMSNorm.forward_hip: Changed fused_add_rms_norm from 6-arg call to vllm's 4-arg in-place API. Removed unnecessary torch.empty_like allocations for out and residual_out.
• GemmaRMSNorm.forward_hip: Same fix as above.
• Added residual = residual.contiguous() before the in-place call to ensure the tensor is contiguous.

Accuracy Tests

N/A - This is a bug fix that makes the vllm fallback path callable. The aiter path (SGLANG_USE_AITER=1) is unaffected.

Speed Tests and Profiling

N/A - No performance-sensitive logic changed. The fix removes two torch.empty_like allocations, which should be neutral or slightly beneficial.

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.