fixes

.github/workflows/lint.yml

@@ -52,6 +52,7 @@ jobs:
           level: warning
           filter_mode: added
           lib: true
+          pyright_version: 1.1.291
       - name: pylint
         uses: dciborow/[email protected]
         with:

onnxruntime/python/onnxruntime_inference_collection.py

@@ -105,7 +105,6 @@ class Session:
     """
 
     def __init__(self):
-
         # self._sess is managed by the derived class and relies on bindings from C.InferenceSession
         self._sess = None
         self._enable_fallback = True

onnxruntime/python/tools/quantization/calibrate.py

@@ -633,7 +633,6 @@ def collect_value(self, name_to_arr):
                 )
 
     def merge_histogram(self, old_histogram, data_arr, new_min, new_max, new_threshold):
-
         (old_hist, old_hist_edges, old_min, old_max, old_threshold) = old_histogram
 
         if new_threshold <= old_threshold:
@@ -849,7 +848,6 @@ def create_calibrator(
     use_external_data_format=False,
     extra_options={},
 ):
-
     calibrator = None
     if calibrate_method == CalibrationMethod.MinMax:
         # default settings for min-max algorithm

onnxruntime/python/tools/quantization/onnx_quantizer.py

@@ -49,7 +49,6 @@ def __init__(
         op_types_to_quantize,
         extra_options=None,
     ):
-
         if not model_has_infer_metadata(model):
             model = save_and_reload_model(model)
         self.value_infos = {vi.name: vi for vi in model.graph.value_info}
@@ -808,7 +807,11 @@ def __quantize_inputs(
             initializer = find_by_name(node_input, self.model.initializer())
             if initializer is not None:
                 if self.per_channel and op_level_per_channel:
-                    (q_weight_name, zp_name, scale_name,) = self.quantize_weight_per_channel(
+                    (
+                        q_weight_name,
+                        zp_name,
+                        scale_name,
+                    ) = self.quantize_weight_per_channel(
                         initializer.name,
                         self.weight_qType if initializer_use_weight_qType else self.activation_qType,
                         axis,

onnxruntime/python/tools/tensorrt/perf/benchmark.py

@@ -233,7 +233,6 @@ def get_latency_result(runtimes, batch_size):
 
 
 def get_ort_session_inputs_and_outputs(name, session, ort_input):
-
     sess_inputs = {}
     sess_outputs = None
 
@@ -428,7 +427,6 @@ def inference_ort(
 
 
 def inference_ort_and_get_prediction(name, session, ort_inputs):
-
     ort_outputs = []
     for ort_input in ort_inputs:
         sess_inputs, sess_outputs = get_ort_session_inputs_and_outputs(name, session, ort_input)
@@ -548,7 +546,6 @@ def generate_onnx_model_random_input(test_times, ref_input):
     inputs = []
 
     for i in range(test_times):
-
         input_data = []
         for tensor in ref_input:
             shape = tensor.shape
@@ -743,7 +740,6 @@ def update_metrics_map_ori(model_to_metrics, name, ep_to_operator):
 #
 ###################################################################################################
 def update_fail_model_map(model_to_fail_ep, model_name, ep, e_type, e):
-
     if model_name in model_to_fail_ep and ep in model_to_fail_ep[model_name]:
         return
 
@@ -766,7 +762,6 @@ def update_fail_model_map(model_to_fail_ep, model_name, ep, e_type, e):
 
 
 def update_fail_model_map_ori(model_to_fail_ep, fail_results, model_name, ep, e_type, e):
-
     if model_name in model_to_fail_ep and ep in model_to_fail_ep[model_name]:
         return
 
@@ -785,7 +780,6 @@ def update_fail_model_map_ori(model_to_fail_ep, fail_results, model_name, ep, e_
 
 
 def skip_ep(model_name, ep, model_to_fail_ep):
-
     if model_name not in model_to_fail_ep:
         return False
 
@@ -969,7 +963,6 @@ def find_test_data_directory(path):
 
 
 def parse_models_info_from_directory(path, models):
-
     test_data_dir = find_test_data_directory(path)
 
     if test_data_dir:
@@ -996,15 +989,13 @@ def parse_models_info_from_directory(path, models):
 
 
 def parse_models_info_from_file(root_dir, path, models):
-
     # default working directory
     root_working_directory = root_dir + "perf/"
 
     with open(path) as f:
         data = json.load(f)
 
         for row in data:
-
             if "root_working_directory" in row:
                 root_working_directory = row["root_working_directory"]
                 continue
@@ -1185,7 +1176,6 @@ def output_details(results, csv_filename):
 
 
 def output_fail(model_to_fail_ep, csv_filename):
-
     with open(csv_filename, mode="w", newline="") as csv_file:
         column_names = ["model", "ep", "error type", "error message"]
 
@@ -1220,7 +1210,6 @@ def add_status_dict(status_dict, model_name, ep, status):
 
 
 def build_status(status_dict, results, is_fail):
-
     if is_fail:
         for model, model_info in results.items():
             for ep, ep_info in model_info.items():
@@ -1240,7 +1229,6 @@ def build_status(status_dict, results, is_fail):
 
 
 def output_status(results, csv_filename):
-
     need_write_header = True
     if os.path.exists(csv_filename):
         need_write_header = False
@@ -1533,7 +1521,6 @@ def output_metrics(model_to_metrics, csv_filename):
 
         results = []
         for model, ep_info in model_to_metrics.items():
-
             result = {}
             result_fp16 = {}
             result["model_name"] = model
@@ -1663,7 +1650,6 @@ def test_models_eps(args, models):
         ep_results = {"latency": {}, "metrics": {}, "session": {}}
 
         for exec_provider in ep_list:
-
             # Skip model + EP combinations that have already failed in a previous run.
             if skip_ep(name, exec_provider, model_to_fail_ep):
                 continue
@@ -1752,7 +1738,6 @@ def run_model_on_ep(
 
     # use float16.py for cuda fp16 only
     if cuda_fp16 == exec_provider:
-
         # handle model
         if "model_path_fp16" in model_info:
             model_path = os.path.normpath(os.path.join(model_work_dir, model_info["model_path_fp16"]))
@@ -1942,7 +1927,6 @@ def benchmark_model_on_ep(
             return
 
     if result:
-
         ep_results["latency"][exec_provider] = {}
         ep_results["latency"][exec_provider]["average_latency_ms"] = result["average_latency_ms"]
         ep_results["latency"][exec_provider]["latency_90_percentile"] = result["latency_90_percentile"]

onnxruntime/python/tools/tensorrt/perf/perf_utils.py

@@ -119,7 +119,6 @@ def pretty_print(pp, json_object):
 
 
 def parse_single_file(f):
-
     try:
         data = json.load(f)
     except Exception as e:

onnxruntime/python/tools/transformers/benchmark.py

@@ -178,7 +178,12 @@ def run_onnxruntime(
                         fusion_options,
                     )
             if "tf" in model_source:
-                (onnx_model_file, is_valid_onnx_model, vocab_size, max_sequence_length,) = export_onnx_model_from_tf(
+                (
+                    onnx_model_file,
+                    is_valid_onnx_model,
+                    vocab_size,
+                    max_sequence_length,
+                ) = export_onnx_model_from_tf(
                     model_name,
                     MODELS[model_name][1],
                     MODELS[model_name][2],

onnxruntime/python/tools/transformers/bert_perf_test.py

@@ -517,7 +517,7 @@ def main():
     with open(summary_file, "w+", newline="") as tsv_file:
         tsv_writer = csv.writer(tsv_file, delimiter="\t", lineterminator="\n")
         headers = None
-        for (key, perf_result) in sorted_results:
+        for key, perf_result in sorted_results:
             params = key.split(",")
             if headers is None:
                 headers = [

onnxruntime/python/tools/transformers/compare_bert_results.py

@@ -90,7 +90,6 @@ def run_test(
     segment_ids_name,
     input_mask_name,
 ):
-
     # Try deduce input names from optimized model.
     input_ids, segment_ids, input_mask = get_bert_inputs(
         optimized_model, input_ids_name, segment_ids_name, input_mask_name

onnxruntime/python/tools/transformers/convert_generation.py

@@ -1165,7 +1165,6 @@ def generate_gpt2_init_decoder(
 
     # Try without the Casts before and after the MatMuls
     if logits_matmul_to_residual_add_path is None:
-
         # Normalization Node is : LayerNormalization
         logits_matmul_to_residual_add_path = gpt2_init_decoder_model.match_parent_path(
             logits_matmul_node,

onnxruntime/python/tools/transformers/fusion_embedlayer.py

@@ -112,7 +112,12 @@ def check_attention_subgraph(
                 logger.debug("No Attention like subgraph in children of LayerNormalization")
                 return False
         else:
-            if children_types != ["Add", "MatMul", "MatMul", "MatMul",] and children_types != [
+            if children_types != [
+                "Add",
+                "MatMul",
+                "MatMul",
+                "MatMul",
+            ] and children_types != [
                 "MatMul",
                 "MatMul",
                 "MatMul",

onnxruntime/python/tools/transformers/models/bart/export.py

@@ -32,7 +32,6 @@ def print_args(args):
 
 
 def user_command():
-
     parent_parser = argparse.ArgumentParser(add_help=False)
     parent_parser.add_argument("--max_length", type=int, default=20, help="default to 20")
     parent_parser.add_argument("--min_length", type=int, default=0, help="default to 0")
@@ -66,7 +65,6 @@ def user_command():
 
 
 if __name__ == "__main__":
-
     args = user_command()
     if args.opset_version < 14:
         raise ValueError(f"The minimum supported opset version is 14! The given one was {args.opset_version}.")

onnxruntime/python/tools/transformers/models/bart/utils/export_summarization_edinit.py

@@ -89,7 +89,6 @@ def _create_encoder_export(args, config: BartConfig):
     def _prepare_encoder_decoder_kwargs_for_generation(
         self, input_ids: torch.Tensor, model_kwargs, model_input_name: Optional[str] = None
     ) -> Dict[str, Any]:
-
         # retrieve encoder hidden states
         # 1. get encoder
         encoder = self.get_encoder()
@@ -189,7 +188,6 @@ def export_encoder(args):
     config, tokenizer = export_helper.initialize_config(args)
 
     with torch.no_grad():
-
         model, input_data = export_helper.initialize_model(config, tokenizer, args)
         start_time = time.time()
         model._prepare_encoder_decoder_kwargs_for_generation = _create_encoder_export(args, config).__get__(

onnxruntime/python/tools/transformers/models/bart/utils/export_summarization_enc_dec_past.py

@@ -249,7 +249,6 @@ def export_decoder(args):
     config = decoder_config_update(config)
 
     with torch.no_grad():
-
         model, input_data = export_helper.initialize_model(config, tokenizer, args)
         start_time = time.time()
 

onnxruntime/python/tools/transformers/models/bart/utils/onnx_inference.py

@@ -32,7 +32,6 @@ def run_inference(args):
     config, tokenizer = export_helper.initialize_config(args)
 
     with torch.no_grad():
-
         model, input_data = export_helper.initialize_model(config, tokenizer, args)
         batch_num = 3
         input_data = input_data.repeat(batch_num, 1)

onnxruntime/python/tools/transformers/models/gpt2/gpt2_tester.py

@@ -130,7 +130,6 @@ def __init__(
         top_k=20,
         top_k_required_order=False,
     ):
-
         self.batch_size = input_ids.shape[0]
         self.input_length = input_ids.shape[1]
         self.n_layer = num_layer
@@ -462,7 +461,10 @@ def test_generation(
                     )
                     Gpt2Helper.auto_increase_buffer_size(output_buffers, output_shapes)
 
-                    (onnx_io_output, avg_latency_ms,) = Gpt2Helper.onnxruntime_inference_with_binded_io(
+                    (
+                        onnx_io_output,
+                        avg_latency_ms,
+                    ) = Gpt2Helper.onnxruntime_inference_with_binded_io(
                         session,
                         onnx_io_runner.get_inputs(),
                         output_buffers,

onnxruntime/python/tools/transformers/models/longformer/convert_to_onnx.py

@@ -149,7 +149,6 @@ def parse_arguments():
 
 # Create a dummy input for ONNX export.
 def get_dummy_inputs(config, export_padding, device):
-
     # When sequence length is multiple of windows size, there is no padding logic in ONNX graph
     sequence_length = config.attention_window[0] + 1 if export_padding else config.attention_window[0]
 

onnxruntime/python/tools/transformers/models/longformer/generate_test_data.py

@@ -264,7 +264,6 @@ def create_longformer_test_data(
     global_mask_name,
     num_global_tokens,
 ):
-
     input_ids, input_mask, global_mask = get_longformer_inputs(model, input_ids_name, input_mask_name, global_mask_name)
     all_inputs = generate_test_data(
         batch_size,

onnxruntime/python/tools/transformers/models/t5/t5_decoder.py

@@ -93,7 +93,6 @@ def __init__(self, decoder, lm_head, config):
         self.config = config
 
     def forward(self, decoder_input_ids, encoder_attention_mask, encoder_hidden_states, *past):
-
         past_key_values = PastKeyValuesHelper.group_by_layer(past, self.config.num_layers)
 
         decoder_outputs = self.decoder(

onnxruntime/python/tools/transformers/models/t5/t5_encoder_decoder_init.py

@@ -255,7 +255,7 @@ def verify_onnx(
 
         test_cases = [(4, 11), (1, 2), (3, 1), (8, 5)]
         test_cases_max_diff = []
-        for (batch_size, encode_sequence_length) in test_cases[:max_cases]:
+        for batch_size, encode_sequence_length in test_cases[:max_cases]:
             inputs = T5EncoderDecoderInitInputs.create_dummy(
                 model.config,
                 batch_size,

onnxruntime/python/tools/transformers/onnx_exporter.py

@@ -461,7 +461,6 @@ def export_onnx_model_from_pt(
     model_fusion_statistics,
     fusion_options,
 ):
-
     config, model = load_pt_model(model_name, model_class, cache_dir, config_modifier)
     # config, model = load_pt_model_from_tf(model_name)
     model.cpu()

onnxruntime/python/tools/transformers/onnx_model_bart.py

@@ -159,7 +159,6 @@ def fuse(self, normalize_node, input_name_to_nodes, output_name_to_node):
             return
 
         if matmul_v.input[0] == root_input and matmul_q.input[0] == root_input and matmul_v.input[0] == root_input:
-
             mask_nodes = []
             mask_index = None
             attention_last_node = reshape_qkv_2

onnxruntime/python/tools/transformers/onnx_model_tnlr.py

@@ -40,7 +40,6 @@ def create_attention_node(
         output: str,
         add_qk_str: str,
     ) -> Union[NodeProto, None]:
-
         assert num_heads > 0
         if hidden_size > 0 and (hidden_size % num_heads) != 0:
             logger.debug(f"input hidden size {hidden_size} is not a multiple of num of heads {num_heads}")

onnxruntime/python/tools/transformers/shape_optimizer.py

@@ -107,7 +107,6 @@ def add_extra_graph_output(self, extra_outputs):
         names_to_evaluate = []
         output_names = [output.name for output in self.model.graph.output]
         for name in extra_outputs:
-
             if self.get_initializer(name) is not None:  # already a constant
                 continue
             names_to_evaluate.append(name)