[Bug] TestGemma4MoE::test_bf16 fails deterministically in pre-merge DGX_H100-4_GPUs-AutoDeploy-1: Not registered specs (bf16/auto)

Summary

accuracy/test_llm_api_autodeploy.py::TestGemma4MoE::test_bf16 fails deterministically with:

ValueError: Not registered specs: {'dtype': 'auto', 'quant_algo': None, 'kv_cache_quant_algo': None, 'spec_dec_algo': None, ...}

The accuracy reference files (mmlu/gsm8k/mmmu yaml) contain google/gemma-4-26B-A4B-it entries, but none matching the plain bf16/auto config this test requests.

Impact

• Stage DGX_H100-4_GPUs-AutoDeploy-1 is pre-merge, so any unrelated PR whose run selects this stage fails CI (observed on PR #15730: L0_MergeRequest_PR builds 47313 and 47345, base = current main in both cases).
• Because the stage is pre-merge-only, the post-merge waive automation never sees it — no auto-waive will come.

Ask

Either register the missing spec for the bf16/auto config, or waive the test until it is fixed. Also reported on #12710 (test origin).

cc @NVIDIA/trt-llm-autodeploy-devs

Start with accuracy/test_llm_api_autodeploy.py::TestGemma4MoE::test_bf16 and inspect the mmlu, gsm8k, and mmmu YAML reference entries for google/gemma-4-26B-A4B-it. Compare those registrations with the requested bf16/auto configuration and determine whether the missing spec should be registered or the test waived; done means the pre-merge stage no longer fails deterministically.