[CUDA] Fix build errors (#27319)

Fix build errors that not caught by CI, but show up in post-merge
builds.

It was caused by two commits:
5a9877a
and
9adf238

One commit changed GQA interface, while the other commit uses old GQA
interface.

Author: tianleiwu

onnxruntime/contrib_ops/cuda/bert/group_query_attention_impl.cu

@@ -612,8 +612,8 @@ Status ExtremeDecoding(
       batch_size,
       parameters.seqlen_present_kv_cache,  // max_seqlen (capacity)
       data.past_seq_lens,
-      data.cos_cache,
-      data.sin_cache,
+      reinterpret_cast<const CudaT*>(data.cos_cache),
+      reinterpret_cast<const CudaT*>(data.sin_cache),
       parameters.do_rotary ? parameters.rotary_dim : 0,
       data.position_ids,
       parameters.rotary_interleaved,
@@ -1105,6 +1105,7 @@ Status QkvToContext(
 
 template struct GroupQueryAttentionData<half, half>;
 template struct GroupQueryAttentionData<__nv_bfloat16, __nv_bfloat16>;
+template struct GroupQueryAttentionData<BFloat16, BFloat16>;
 template struct GroupQueryAttentionData<half, int8_t>;
 
 template Status QkvToContext<half, half>(
@@ -1121,6 +1122,13 @@ template Status QkvToContext<__nv_bfloat16, __nv_bfloat16>(
     contrib::GroupQueryAttentionParameters& parameters,
     GroupQueryAttentionData<__nv_bfloat16, __nv_bfloat16>& data);
 
+template Status QkvToContext<BFloat16, BFloat16>(
+    const cudaDeviceProp& device_prop,
+    cublasHandle_t& cublas,
+    Stream* ort_stream,
+    contrib::GroupQueryAttentionParameters& parameters,
+    GroupQueryAttentionData<BFloat16, BFloat16>& data);
+
 template Status QkvToContext<half, int8_t>(
     const cudaDeviceProp& device_prop,
     cublasHandle_t& cublas,

onnxruntime/core/providers/cuda/llm/attention.cc

@@ -196,7 +196,7 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
     gqa_parameters.num_splits = 1;
 
     // Construct GroupQueryAttentionData
-    onnxruntime::contrib::cuda::GroupQueryAttentionData<CudaT> gqa_data;
+    onnxruntime::contrib::cuda::GroupQueryAttentionData<CudaT, CudaT> gqa_data;
 
     // Scratch buffers for flash/memory efficient attention
     IAllocatorUniquePtr<void> k_buffer;
@@ -355,6 +355,7 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
     // Centralized scratch buffer allocation using GQABufferRequirements
     auto buffer_req = onnxruntime::contrib::cuda::GQABufferRequirements::Compute<T>(
         gqa_parameters,
+        false,  // use_xqa
         gqa_data.use_flash_attention,
         gqa_data.use_flash_attention_fast_decode,
         gqa_data.use_memory_efficient_attention);
@@ -478,7 +479,7 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
     // Call GQA kernel (with flash or memory efficient attention)
     cublasHandle_t cublas = GetCublasHandle(context);
 
-    return onnxruntime::contrib::cuda::QkvToContext<CudaT>(
+    return onnxruntime::contrib::cuda::QkvToContext<CudaT, CudaT>(
         device_prop, cublas, context->GetComputeStream(), gqa_parameters, gqa_data);
   }