update api stability

Signed-off-by: peaceh <103117813+peaceh-nv@users.noreply.github.com>

Author: peaceh-nv

tensorrt_llm/_torch/custom_ops/cute_dsl_custom_ops.py

@@ -327,15 +327,12 @@ def get_dense_gemm_approximate_cta_nums(
         STATE_SIZE as DISTRIBUTED_TOPK_STATE_SIZE
     from ..cute_dsl_kernels.blackwell.top_k.single_pass_multi_cta_radix_topk import \
         SinglePassMultiCTARadixTopKKernel
-<<<<<<< HEAD
     from ..cute_dsl_kernels.blackwell.top_k.single_pass_multi_cta_radix_topk_cluster import \
         STATE_SIZE as CLUSTER_TOPK_STATE_SIZE
     from ..cute_dsl_kernels.blackwell.top_k.single_pass_multi_cta_radix_topk_cluster import (
         SinglePassMultiCTARadixTopKClusterKernel, _query_max_cluster_size)
     from ..cute_dsl_kernels.blackwell.dense_gemm_persistent import \
         PersistentDenseGemmKernel
-=======
->>>>>>> f6e66826d ([TRTLLM-11289][fix] Fix pre-commit formatting for CuTe DSL BF16 GEMM/BMM code)
     from ..cute_dsl_kernels.blackwell.utils import make_ptr
 
     class CuteDSLNVFP4BlackwellRunner(TunableRunner):

tests/unittest/api_stability/references/llm.yaml

@@ -251,6 +251,14 @@ methods:
         annotation: bool
         default: False
         status: prototype
+      use_cute_dsl_bf16_bmm:
+        annotation: bool
+        default: False
+        status: prototype
+      use_cute_dsl_bf16_gemm:
+        annotation: bool
+        default: False
+        status: prototype
     return_annotation: None
   generate:
     parameters: