Qualcomm AI Engine Direct - Remove input_list dependencies (#13411)

### Summary
- Current SimpleADB push logic require each script to prepare
input_list, and the name must follow the logic. Remove all these
dependencies.

### Test plan
All existing unit tests should already cover this change.

Co-authored-by: Cheng-Hsin Weng <[email protected]>

Author: chenweng-quic

backends/qualcomm/debugger/utils.py

@@ -267,11 +267,6 @@ def qnn_context_binary_generator(
         assert os.path.isfile(f"{self.tmp_dir}/{binary_name}.bin"), result.stderr
 
     def qnn_net_run(self, graph_name="forward.serialized"):
-        input_list = ""
-        for idx, _ in enumerate(self.sample_input):
-            input_name = f"input_{idx}_0.raw"
-            input_list += input_name + " "
-        input_list = input_list.strip() + "\n"
 
         self.config["backend_extension_config"]["backend_extensions"][
             "shared_library_path"
@@ -304,7 +299,6 @@ def qnn_net_run(self, graph_name="forward.serialized"):
         ]
         self.adb.push(
             inputs=self.sample_input,
-            input_list=input_list,
             files=files,
         )
         self.adb.execute(custom_runner_cmd=" ".join(cmds))

backends/qualcomm/tests/utils.py

@@ -193,13 +193,6 @@ def _save_model_and_expected_output(
         inputs: Tuple[torch.Tensor],
         dir_name: str,
     ) -> None:
-        # Save the input data list to be executed
-        input_list = ""
-        for idx, _ in enumerate(inputs):
-            input_name = f"input_0_{idx}.raw"
-            input_list += input_name + " "
-        input_list = input_list.strip() + "\n"
-
         ref_output = module(*inputs)
 
         # Save the expected output data to be verified
@@ -216,7 +209,7 @@ def _save_model_and_expected_output(
         with open(pte_fname, "wb") as file:
             file.write(buffer)
 
-        return input_list, ref_outputs, pte_fname
+        return ref_outputs, pte_fname
 
     def required_envs(self, conditions=None) -> bool:
         conditions = [] if conditions is None else conditions
@@ -247,7 +240,6 @@ def verify_output(  # noqa: C901
     ):
         with tempfile.TemporaryDirectory() as tmp_dir:
             (
-                input_list,
                 ref_outputs,
                 pte_fname,
             ) = self._save_model_and_expected_output(
@@ -319,9 +311,7 @@ def validate_intermediate_tensor():
                 )
 
             if self.enable_x86_64:
-                generate_inputs(
-                    tmp_dir, "input_list.txt", [processed_inputs], input_list
-                )
+                generate_inputs(tmp_dir, "input_list.txt", [processed_inputs])
                 make_output_dir(output_dir)
 
                 target = "x86_64-linux-clang"
@@ -434,7 +424,6 @@ def validate_intermediate_tensor():
                 )
                 adb.push(
                     inputs=[processed_inputs],
-                    input_list=input_list,
                     files=op_package_paths,
                 )
                 adb.extra_cmds += extra_cmds

examples/qualcomm/custom_op/custom_ops_1.py

@@ -102,15 +102,6 @@ def annotate_custom(gm: torch.fx.GraphModule) -> None:
         )
 
 
-def create_device_inputs(example_inputs):
-    input_list = ""
-    for idx, _ in enumerate(example_inputs):
-        input_name = f"input_0_{idx}.raw"
-        input_list += input_name + " "
-    input_list = input_list.strip() + "\n"
-    return input_list
-
-
 def _run(cmd, cwd=None):
     subprocess.run(cmd, stdout=sys.stdout, cwd=cwd, check=True)
 
@@ -204,7 +195,6 @@ def main(args):
     sample_input = (torch.ones(1, 32, 28, 28),)
     workspace = f"/data/local/tmp/executorch/{pte_filename}"
 
-    input_list = create_device_inputs(sample_input)
     soc_info = _soc_info_table[getattr(QcomChipset, args.model)]
 
     op_package_options, op_package_paths = prepare_op_package(
@@ -237,8 +227,7 @@ def main(args):
 
     if args.enable_x86_64:
         input_list_filename = "input_list.txt"
-        input_list = f"{args.artifact}/{input_list}"
-        generate_inputs(args.artifact, input_list_filename, sample_input, input_list)
+        generate_inputs(args.artifact, input_list_filename, sample_input)
         qnn_sdk = os.getenv("QNN_SDK_ROOT")
         assert qnn_sdk, "QNN_SDK_ROOT was not found in environment variable"
         target = "x86_64-linux-clang"
@@ -276,7 +265,7 @@ def main(args):
             host_id=args.host,
             soc_model=args.model,
         )
-        adb.push(inputs=sample_input, input_list=input_list, files=op_package_paths)
+        adb.push(inputs=sample_input, files=op_package_paths)
         adb.execute()
         adb.pull(output_path=args.artifact)
 

examples/qualcomm/oss_scripts/albert.py

@@ -51,7 +51,7 @@ def main(args):
             "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy."
         )
     else:
-        inputs, targets, input_list = get_masked_language_model_dataset(
+        inputs, targets = get_masked_language_model_dataset(
             args.dataset, tokenizer, data_size
         )
 
@@ -94,7 +94,7 @@ def main(args):
     make_output_dir(output_data_folder)
 
     # accuracy analysis
-    adb.push(inputs=inputs, input_list=input_list)
+    adb.push(inputs=inputs)
     adb.execute()
     adb.pull(output_path=args.artifact)
     # since the original nn.Module could not perform well on this task either

examples/qualcomm/oss_scripts/bert.py

@@ -50,7 +50,7 @@ def main(args):
             "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy."
         )
     else:
-        inputs, targets, input_list = get_masked_language_model_dataset(
+        inputs, targets = get_masked_language_model_dataset(
             args.dataset, tokenizer, data_size
         )
     module = AutoModelForMaskedLM.from_pretrained(
@@ -92,7 +92,7 @@ def main(args):
     make_output_dir(output_data_folder)
 
     # accuracy analysis
-    adb.push(inputs=inputs, input_list=input_list)
+    adb.push(inputs=inputs)
     adb.execute()
     adb.pull(output_path=args.artifact)
     goldens, predictions = [], []

examples/qualcomm/oss_scripts/conv_former.py

@@ -51,7 +51,7 @@ def main(args):
             "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy."
         )
     else:
-        inputs, targets, input_list = get_imagenet_dataset(
+        inputs, targets = get_imagenet_dataset(
             dataset_path=f"{args.dataset}",
             data_size=data_num,
             image_shape=(256, 256),
@@ -89,7 +89,7 @@ def main(args):
         soc_model=args.model,
         shared_buffer=args.shared_buffer,
     )
-    adb.push(inputs=inputs, input_list=input_list)
+    adb.push(inputs=inputs)
     adb.execute()
 
     # collect output data

examples/qualcomm/oss_scripts/cvt.py

@@ -106,7 +106,7 @@ def main(args):
             "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy."
         )
     else:
-        inputs, targets, input_list = get_imagenet_dataset(
+        inputs, targets = get_imagenet_dataset(
             dataset_path=f"{args.dataset}",
             data_size=data_num,
             image_shape=(256, 256),
@@ -146,7 +146,7 @@ def main(args):
         soc_model=args.model,
         shared_buffer=args.shared_buffer,
     )
-    adb.push(inputs=inputs, input_list=input_list)
+    adb.push(inputs=inputs)
     adb.execute()
 
     # collect output data

examples/qualcomm/oss_scripts/deit.py

@@ -55,7 +55,7 @@ def main(args):
             "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy."
         )
     else:
-        inputs, targets, input_list = get_imagenet_dataset(
+        inputs, targets = get_imagenet_dataset(
             dataset_path=f"{args.dataset}",
             data_size=data_num,
             image_shape=(height, width),
@@ -96,7 +96,7 @@ def main(args):
         host_id=args.host,
         soc_model=args.model,
     )
-    adb.push(inputs=inputs, input_list=input_list)
+    adb.push(inputs=inputs)
     adb.execute()
 
     # collect output data

examples/qualcomm/oss_scripts/dino_v2.py

@@ -49,7 +49,7 @@ def main(args):
         )
 
     img_size, data_num = 224, 100
-    inputs, targets, input_list = get_imagenet_dataset(
+    inputs, targets = get_imagenet_dataset(
         dataset_path=f"{args.dataset}",
         data_size=data_num,
         image_shape=(256, 256),
@@ -85,7 +85,7 @@ def main(args):
         host_id=args.host,
         soc_model=args.model,
     )
-    adb.push(inputs=inputs, input_list=input_list)
+    adb.push(inputs=inputs)
     adb.execute()
 
     # collect output data

examples/qualcomm/oss_scripts/distilbert.py

@@ -50,7 +50,7 @@ def main(args):
             "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy."
         )
     else:
-        inputs, targets, input_list = get_masked_language_model_dataset(
+        inputs, targets = get_masked_language_model_dataset(
             args.dataset, tokenizer, data_size
         )
     module = AutoModelForMaskedLM.from_pretrained(
@@ -92,7 +92,7 @@ def main(args):
     make_output_dir(output_data_folder)
 
     # accuracy analysis
-    adb.push(inputs=inputs, input_list=input_list)
+    adb.push(inputs=inputs)
     adb.execute()
     adb.pull(output_path=args.artifact)
     goldens, predictions = [], []