Revert Implement DML copy for Lora Adapters (microsoft#22814)

Revert microsoft#22396
intel · Dec 11, 2024 · 633cdcc · 633cdcc
1 parent 9156269
commit 633cdcc
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 79 deletions.
diff --git a/onnxruntime/core/session/lora_adapters.cc b/onnxruntime/core/session/lora_adapters.cc
@@ -52,19 +52,6 @@ void LoraAdapter::MemoryMap(const std::filesystem::path& file_path) {
 
 static std::unique_ptr<IDataTransfer> GetDataTransfer(const OrtMemoryInfo& mem_info) {
   std::unique_ptr<IDataTransfer> data_transfer;
-  bool is_dml = false;
-  Status CopyTensor(const Tensor& src, Tensor& dst) const {
-    return data_transfer->CopyTensor(src, dst);
-  }
-  Status Sync() const {
-    if (is_dml) {
-      return ep->Sync();
-    } else {
-      return Status::OK();
-    }
-  }
-};
-}  // namespace
 
   if (strcmp(mem_info.name, onnxruntime::CPU) == 0) {
     return data_transfer;
@@ -77,17 +64,6 @@ static std::unique_ptr<IDataTransfer> GetDataTransfer(const OrtMemoryInfo& mem_i
       data_transfer = cuda_provider_info->CreateGPUDataTransfer();
     }
 #endif
-  } else if (strcmp(mem_info.name, onnxruntime::DML) == 0) {
-#ifdef USE_DML
-    auto ep_factory = onnxruntime::DMLProviderFactoryCreator::Create(ConfigOptions{}, 0, false, false, false);
-    dt.ep = ep_factory->CreateProvider();
-    dt.is_dml = true;
-    dt.data_transfer = dt.ep->GetDataTransfer();
-#else
-    status = ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "DML provider is not enabled in this build");
-#endif
-  } else {
-    status = ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Unsupported device allocator");
   }
 
   return data_transfer;

diff --git a/onnxruntime/test/lora/lora_test.cc b/onnxruntime/test/lora/lora_test.cc
@@ -200,15 +200,7 @@ TEST(LoraAdapterTest, Load) {
 }
 
 #ifdef USE_CUDA
-TEST(LoraAdapterTest, VerifyCudaDeviceCopy) {
-  if (DefaultCudaExecutionProvider() == nullptr) {
-    GTEST_SKIP() << "Skip This Test Due to this EP is null";
-  }
-#ifdef USE_DML
-  if (DefaultDmlExecutionProvider() != nullptr) {
-    GTEST_FAIL() << "It should not run with DML EP";
-  }
-#endif
+TEST(LoraAdapterTest, VerifyDeviceCopy) {
   auto cpu_ep = DefaultCpuExecutionProvider();
   auto cpu_allocator = cpu_ep->CreatePreferredAllocators()[0];
   auto cuda_ep = DefaultCudaExecutionProvider();
@@ -230,52 +222,7 @@ TEST(LoraAdapterTest, VerifyCudaDeviceCopy) {
     ASSERT_EQ(tensor_cpu.Shape().Size(), tensor_device.Shape().Size());
 
     Tensor copy(tensor_cpu.DataType(), tensor_cpu.Shape(), cpu_allocator);
-    ASSERT_TRUE(cuda_transfer->CanCopy(tensor_device.Location().device,
-                                       copy.Location().device));
-    ASSERT_STATUS_OK(cuda_transfer->CopyTensor(tensor_device, copy));
-
-    auto expected_span = tensor_cpu.DataAsSpan<float>();
-    auto copy_span = copy.DataAsSpan<float>();
-
-    ASSERT_EQ(expected_span, copy_span);
-  }
-}
-#endif
-
-#ifdef USE_DML
-TEST(LoraAdapterTest, VerifyDmlDeviceCopy) {
-  // NO_DML_TEST is set, DML test is skipped
-  if (DefaultDmlExecutionProvider() == nullptr) {
-    GTEST_SKIP() << "Skip This Test Due to this EP is null";
-  }
-
-#ifdef USE_CUDA
-  if (DefaultCudaExecutionProvider() != nullptr) {
-    GTEST_FAIL() << "It should not run with CUDA EP";
-  }
-#endif
-
-  auto cpu_ep = DefaultCpuExecutionProvider();
-  auto cpu_allocator = cpu_ep->CreatePreferredAllocators()[0];
-
-  auto dml_allocator = DefaultDmlExecutionProvider()->CreatePreferredAllocators()[0];
-  auto dml_transfer = DefaultDmlExecutionProvider()->GetDataTransfer();
-
-  auto test_params = GenerateTestParameters<float>()();
-  lora::LoraAdapter adapter(std::move(dml_allocator));
-  adapter.Load(std::move(test_params));
-
-  auto [begin, end] = adapter.GetParamIterators();
-  for (; begin != end; ++begin) {
-    const auto& [_, param] = *begin;
-    const auto& tensor_device = param.GetDeviceOrMapped().Get<Tensor>();
-    ASSERT_EQ(0, strcmp(tensor_device.Location().name, onnxruntime::DML));
-
-    const auto& tensor_cpu = param.GetMapped().Get<Tensor>();
-    ASSERT_EQ(tensor_cpu.Shape().Size(), tensor_device.Shape().Size());
-
-    Tensor copy(tensor_cpu.DataType(), tensor_cpu.Shape(), cpu_allocator);
-    ASSERT_TRUE(dml_transfer->CanCopy(tensor_device.Location().device,
+    ASSERT_TRUE(gpu_transfer->CanCopy(tensor_device.Location().device,
                                       copy.Location().device));
     ASSERT_STATUS_OK(gpu_transfer->CopyTensor(tensor_device, copy));