Fix CUDA BeamSearch fp16 score mapping and add regression test

[Akshay Sonawane (apsonawane)]

This pull request introduces improvements to the Beam Search implementation and its test coverage, specifically for FP16 (half-precision float) support. The main changes include a bug fix in the CUDA implementation and the addition of a new test to verify the output type and shape for FP16 scores.

CUDA implementation fix:

• Changed the type mapping in LaunchBeamSearchScoreCopy from always using float to using the template parameter T, ensuring correct type handling for FP16 and other types. (onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.cu)

Test improvements:

• Added a new test, GptBeamSearchFp16_ScoresOutputTypeAndShape, which runs the GPT-2 Beam Search model with FP16 outputs and verifies that the output tensor has the correct type (FLOAT16) and expected shape. (onnxruntime/test/contrib_ops/beam_search_test.cc)

[Copilot]

Pull request overview

This PR fixes FP16 score output handling in the CUDA BeamSearch path by correcting the device-side output type mapping during score copying, and adds a regression test to validate FP16 scores output type/shape using a tiny GPT-2 beam search model.

Changes:

• Fix CUDA LaunchBeamSearchScoreCopy to map the output pointer type using the template parameter T (instead of always float) so FP16 scores are written with the correct element type.
• Add a new BeamSearch test that runs the FP16 GPT-2 beam search model and validates the scores output tensor element type is FLOAT16 and the shape matches expectations.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.cu	Corrects the CUDA output type mapping used when converting/copying final beam scores into the output tensor.
onnxruntime/test/contrib_ops/beam_search_test.cc	Adds a regression test for FP16 scores output type/shape (CUDA-only execution).

Suppressed comments (1)

onnxruntime/test/contrib_ops/beam_search_test.cc:290

• Inside the new CUDA-guarded block, the indentation is inconsistent (e.g., Ort::SessionOptions is not indented under the if, and #ifdef USE_CUDA has leading spaces). This reduces readability and deviates from the pattern used in the preceding GptBeamSearchFp16 test.

  if (enable_cuda) {
    Ort::SessionOptions session_options;
#ifdef USE_CUDA
    OrtCUDAProviderOptionsV2 cuda_options;
    cuda_options.use_tf32 = false;
    session_options.AppendExecutionProvider_CUDA_V2(cuda_options);
#endif