fixes

Author: Vladimir Paramuzov

src/plugins/intel_gpu/src/kernel_selector/cl_kernels/sdpa_micro.cl

@@ -265,7 +265,7 @@ KERNEL(micro_sdpa)(OPTIONAL_SHAPE_INFO_ARG
 
 #if WITH_ATTN_MASK
         mask_tile_type mask_tile;
-        tile_load_t(&mask_tile, msk, k, q, q, sg_j0_kq + wg_j0, k0 + sg_i0_kq);
+        tile_load_t(&mask_tile, msk, q, k, q, sg_j0_kq + wg_j0, k0 + sg_i0_kq);
 #endif
 
 #if REMAINDER_K

src/plugins/intel_gpu/src/kernel_selector/kernels/sdpa/sdpa_kernel_micro.cpp

@@ -187,7 +187,7 @@ void SDPAKernelMicro::init_microkernels(const sdpa_params& params, micro::Packag
     auto& out = params.outputs[0];
     const auto head_size = params.conf.head_size;
     const auto d_max = get_d_max(head_size);
-    const Tensor::Dim n_keys = K.X().v; //get_seq_length(K, params.input1_order);
+    const Tensor::Dim n_keys = get_seq_length(K, params.input1_order);
     const Tensor::Dim n_queries = get_seq_length(Q, params.input0_order);
     const Tensor::Dim n_values = V.X();
     const auto batch = out.Batch().v * out.Feature().v;
@@ -585,7 +585,7 @@ void SDPAKernelMicro::GetUpdateDispatchDataFunc(KernelData& kd) const {
         s_q.t = ScalarDescriptor::Types::INT32;
         s_q.v.s32 = static_cast<uint32_t>(n_queries.v);
 
-        const bool is_prefill = true;//n_queries.v > 1;
+        const bool is_prefill = n_queries.v > 1;
 
         OPENVINO_ASSERT(kernel_data.kernels.size() == 2, "[GPU] Invalid kernels size for update dispatch data func");