Fix GemmFloat8 C operand descriptor type

[Akshay Sonawane (apsonawane)]

Use the actual optional C tensor type for cuBLASLt layout descriptors, preserve output-typed fallback descriptors with an effective zero beta when C is absent, and cover mixed and scaled FP8 cases.

[Copilot]

Pull request overview

This PR fixes how the CUDA GemmFloat8 contrib op configures cuBLASLt’s C operand matrix layout descriptor by using the actual (optional) input C tensor type when present, and a safe output-typed fallback (with beta=0) when C is absent. This aligns descriptor types with the real inputs and adds coverage for mixed-type and scaled FP8 scenarios.

Changes:

• Update cuBLASLt Cdesc creation to use dtype_C when C exists, otherwise use the output type with an effective beta=0.
• Add a CUDA<12 runtime guard for the unsupported combination of C + nonzero beta.
• Extend unit tests to cover mixed C types, scaled FP8, and FP8 execution without C.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
onnxruntime/test/contrib_ops/gemm_float8_test.cc	Adds CUDA tests for mixed-type C, scaled FP8 cases, and the “no C” path.
onnxruntime/contrib_ops/cuda/math/gemm_float8.cu	Fixes cuBLASLt C layout descriptor typing and ensures beta is effectively zero when C is absent.
onnxruntime/contrib_ops/cuda/math/gemm_float8.cc	Adds the kernel type constraint for TC to match the schema’s allowed C types.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Tianlei Wu (tianleiwu)]

Two blocking FP8 issues remain. The non-FP8 mixed-C descriptor fix and the effective zero-beta handling are sound, but FP8 output currently selects unsupported C descriptor types, and the new FP8 regression tests never execute because their architecture threshold is impossible.

[Tianlei Wu (tianleiwu)]

Blocking: for FP8 D, cuBLASLt's supported type combinations require matrix C to be FP16 or BF16. This selects FP32 when the supplied C is float (as in the new scaled test), and selects FP8 when C is absent because c_cuda_type = d_cuda_type; neither (C=FP32, D=FP8) nor (C=FP8, D=FP8) is supported. Setting CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE below controls epilogue bias and does not make the matrix-C layout valid. Please use a supported FP16/BF16 C descriptor for FP8 output, reject or convert incompatible supplied C, and use a supported fallback descriptor with zero beta when C is absent.

[Tianlei Wu (tianleiwu)]

Blocking test gap: HasCudaEnvironment expects compute capability encoded as 100 * major + 10 * minor (for example, 890 for Ada), not the CUDA toolkit version. No GPU can satisfy >= 11080, so this test always returns before constructing the model. The helper used by the other new scaled FP8 test has the same pre-existing threshold, which means all FP8 coverage in this file is currently skipped. Please use the actual FP8 minimum (890) and preferably GTEST_SKIP() so an unsupported runner reports a skip explicitly.