Fail explicitly for unsupported dtypes in Rubin MoE atomic_add_func

[coderabbitai]

Summary

atomic_add_func in tensorrt_llm/_torch/cute_dsl_kernels/rubin/moe/utils.py emits atomic-add PTX only for cutlass.Float32 and cutlass.BFloat16.

For another dtype, including cutlass.Float16, the function currently emits no instruction. This can silently omit an accumulation.

Required change

Add an explicit unsupported-dtype branch after the Float32 and BFloat16 cases. Use the established error mechanism for this CuTe DSL path.

Rationale

The function must fail immediately when it receives a dtype that has no matching atomic-add implementation. It must not silently skip the accumulation.

Affected area

• tensorrt_llm/_torch/cute_dsl_kernels/rubin/moe/utils.py
• atomic_add_func

Acceptance criteria

• atomic_add_func retains its existing Float32 behavior.
• atomic_add_func retains its existing BFloat16 behavior.
• An unsupported dtype causes an explicit failure.
• The selected failure mechanism is compatible with CuTe DSL compilation.

Backlinks

• Pull request: [None][feat] Add Rubin SM107 CuTe DSL foundation and BF16 kernels #18311
• Review comment: [None][feat] Add Rubin SM107 CuTe DSL foundation and BF16 kernels #18311 (comment)
• Requested by: @farazkh80