Modify MIGraphX EP for Accuracy tests (#13455)

Allows MIGraphX EP to run the following additional tests. Also adds support to get MIGraphX to run eval_squad.py

Reference to the Rocm EP changes: https://github.com/microsoft/onnxruntime/pull/13306

Co-authored-by: Joseph Groenenboom <joseph.groenenboom@amd.com>
Co-authored-by: Ted Themistokleous <tthemist@amd.com>

onnxruntime/core/providers/migraphx/migraphx_execution_provider.cc

@@ -527,11 +527,6 @@ static bool IsUnsupportedOpMode(const onnxruntime::GraphViewer& graph_viewer, co
 }
 
 void SubgraphPostProcessing(const onnxruntime::GraphViewer& graph_viewer, std::vector<std::vector<NodeIndex>>& clusters, const logging::Logger& logger) {
-  // If the number of nodes in the graph is less than 5, do nothing
-  // this is to deal with onnx unit tests
-  if (graph_viewer.NumberOfNodes() <= 5) {
-    return;
-  }
 
   // Then check whether a subgraph should fallback to CPU
   // 1. Check whether a subgraph contains a RNN operator
@@ -556,11 +551,6 @@ void SubgraphPostProcessing(const onnxruntime::GraphViewer& graph_viewer, std::v
       }
     }
 
-    // if 6 operators or more
-    if (git.size() > 5) {
-      return false;
-    }
-
     // rnn operators, run on GPU
     if (std::any_of(git.begin(), git.end(), [&](auto nid) {
           const auto& node = graph_viewer.GetNode(nid);

onnxruntime/python/tools/transformers/benchmark_helper.py

@@ -164,9 +164,9 @@ def prepare_environment(cache_dir, output_dir, use_gpu, provider=None):
             ), "Please install onnxruntime-directml package to test GPU inference."
 
         else:
-            assert (
-                "CUDAExecutionProvider" in onnxruntime.get_available_providers()
-            ), "Please install onnxruntime-gpu package to test GPU inference."
+            assert set(onnxruntime.get_available_providers()).isdisjoint(
+                ["CUDAExecutionProvider", "ROCMExecutionProvider", "MIGraphXExecutionProvider"]
+            ), "Please install onnxruntime-gpu package, or install ROCm support, to test GPU inference."
 
     logger.info(f"PyTorch Version:{torch.__version__}")
     logger.info(f"Transformers Version:{transformers.__version__}")

onnxruntime/python/tools/transformers/optimizer.py

@@ -70,9 +70,13 @@ def optimize_by_onnxruntime(
         optimized_model_path (str): the path of optimized model
     """
     assert opt_level in [1, 2, 99]
+    from torch import version as torch_version
+
     import onnxruntime
 
-    if use_gpu and "CUDAExecutionProvider" not in onnxruntime.get_available_providers():
+    if use_gpu and not set(onnxruntime.get_available_providers()).isdisjoint(
+        ["CUDAExecutionProvider", "ROCMExecutionProvider", "MIGraphXExecutionProvider"]
+    ):
         logger.error("There is no gpu for onnxruntime to do optimization.")
         return onnx_model_path
 
@@ -99,13 +103,21 @@ def optimize_by_onnxruntime(
             onnx_model_path, sess_options, providers=["CPUExecutionProvider"], **kwargs
         )
     else:
-        session = onnxruntime.InferenceSession(
-            onnx_model_path, sess_options, providers=["CUDAExecutionProvider"], **kwargs
+        gpu_ep = []
+
+        if torch_version.cuda:
+            gpu_ep.append("CUDAExecutionProvider")
+        elif torch_version.hip:
+            gpu_ep.append("MIGraphXExecutionProvider")
+            gpu_ep.append("ROCMExecutionProvider")
+
+        session = onnxruntime.InferenceSession(onnx_model_path, sess_options, providers=gpu_ep, **kwargs)
+        assert set(onnxruntime.get_available_providers()).isdisjoint(
+            ["CUDAExecutionProvider", "ROCMExecutionProvider", "MIGraphXExecutionProvider"]
         )
-        assert "CUDAExecutionProvider" in session.get_providers()  # Make sure there is GPU
 
     assert os.path.exists(optimized_model_path) and os.path.isfile(optimized_model_path)
-    logger.debug("Save optimized model by onnxruntime to {}".format(optimized_model_path))
+    logger.debug("Save optimized model by onnxruntime to %s", optimized_model_path)
     return optimized_model_path
 
 

onnxruntime/test/python/transformers/parity_utilities.py

@@ -138,7 +138,15 @@ def create_ort_session(onnx_model_path, use_gpu=True):
     sess_options.graph_optimization_level = GraphOptimizationLevel.ORT_DISABLE_ALL
     sess_options.intra_op_num_threads = 2
     sess_options.log_severity_level = 2
-    execution_providers = ["CPUExecutionProvider"] if not use_gpu else ["CUDAExecutionProvider", "CPUExecutionProvider"]
+    execution_providers = ["CPUExecutionProvider"]
+
+    if use_gpu:
+        if torch.version.cuda:
+            execution_providers.append("CUDAExecutionProvider")
+        elif torch.version.hip:
+            execution_providers.append("MIGraphXExecutionProvider")
+            execution_providers.append("ROCMExecutionProvider")
+
     return InferenceSession(onnx_model_path, sess_options, providers=execution_providers)