Qualcomm AI Engine Direct - CI for Non-LLM GA model

.ci/scripts/test_model.sh

@@ -188,6 +188,24 @@ test_model_with_qnn() {
     EXPORT_SCRIPT=edsr
     # Additional deps for edsr
     pip install piq
+  elif [[ "${MODEL_NAME}" == "cvt" ]]; then
+    EXPORT_SCRIPT=cvt
+  elif [[ "${MODEL_NAME}" == "deit" ]]; then
+    EXPORT_SCRIPT=deit
+  elif [[ "${MODEL_NAME}" == "dit" ]]; then
+    EXPORT_SCRIPT=dit
+  elif [[ "${MODEL_NAME}" == "efficientnet" ]]; then
+    EXPORT_SCRIPT=efficientnet
+  elif [[ "${MODEL_NAME}" == "focalnet" ]]; then
+    EXPORT_SCRIPT=focalnet
+  elif [[ "${MODEL_NAME}" == "mobilevit_v1" ]]; then
+    EXPORT_SCRIPT=mobilevit_v1
+  elif [[ "${MODEL_NAME}" == "mobilevit_v2" ]]; then
+    EXPORT_SCRIPT=mobilevit_v2
+  elif [[ "${MODEL_NAME}" == "pvt" ]]; then
+    EXPORT_SCRIPT=pvt
+  elif [[ "${MODEL_NAME}" == "swin" ]]; then
+    EXPORT_SCRIPT=swin_transformer
   elif [[ "${MODEL_NAME}" == "albert" ]]; then
     EXPORT_SCRIPT=albert
   elif [[ "${MODEL_NAME}" == "bert" ]]; then
@@ -196,6 +214,8 @@ test_model_with_qnn() {
     EXPORT_SCRIPT=distilbert
   elif [[ "${MODEL_NAME}" == "eurobert" ]]; then
     EXPORT_SCRIPT=eurobert
+  elif [[ "${MODEL_NAME}" == "roberta" ]]; then
+    EXPORT_SCRIPT=roberta
   else
     echo "Unsupported model $MODEL_NAME"
     exit 1
@@ -210,10 +230,13 @@ test_model_with_qnn() {
     "dl3"|"mv3"|"mv2"|"ic4"|"ic3"|"vit"|"mb"|"w2l")
         SCRIPT_FOLDER=scripts
         ;;
-    "albert"|"bert"|"distilbert")
+    "cvt"|"deit"|"dit"|"focalnet"|"mobilevit_v2"|"pvt"|"swin")
+        SCRIPT_FOLDER=oss_scripts
+        ;;
+    "albert"|"bert"|"distilbert"|"roberta"|"efficientnet"|"mobilevit_v1")
         pip install evaluate
         SCRIPT_FOLDER=oss_scripts
-        # Bert models running in 16bit will encounter op validation fail on some operations,
+        # 16bit models will encounter op validation fail on some operations,
         # which requires CHIPSET >= SM8550.
         QNN_CHIPSET=SM8550
         ;;

.github/workflows/trunk.yml

@@ -470,7 +470,7 @@ jobs:
       docker-image: executorch-ubuntu-22.04-qnn-sdk
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 900
+      timeout: 90
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
@@ -480,23 +480,49 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_model.sh ${{ matrix.model }} "cmake" "qnn"
 
-  test-qnn-optimum-model:
-    name: test-qnn-optimum-model
+  test-qnn-optimum-model1:
+    name: test-qnn-optimum-model1
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
       contents: read
     strategy:
       matrix:
         dtype: [fp32]
-        model: [albert, bert, distilbert] # eurobert requires transfomer >= 4.48.0, skip for now
+        model: [cvt, deit, dit, efficientnet, focalnet, mobilevit_v1, mobilevit_v2] # eurobert requires transfomer >= 4.48.0, skip for now
       fail-fast: false
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-qnn-sdk
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 900
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool cmake
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_model.sh ${{ matrix.model }} "cmake" "qnn"
+
+  test-qnn-optimum-model2:
+    name: test-qnn-optimum-model2
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      matrix:
+        dtype: [fp32]
+        model: [pvt, swin, albert, bert, distilbert, roberta] # eurobert requires transfomer >= 4.48.0, skip for now
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-qnn-sdk
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")

backends/qualcomm/builders/op_slice_copy.py

@@ -56,7 +56,7 @@ def define_node(
         if start < 0:
             start = start % input_tensor.shape[dim]
 
-        if len(node.args) > 3:
+        if len(node.args) > 3 and node.args[3] is not None:
             end = min(cast(int, node.args[3]), input_tensor.shape[dim])
             if end < 0:
                 end = end % input_tensor.shape[dim]

backends/qualcomm/tests/test_qnn_delegate.py

@@ -4585,14 +4585,13 @@ def test_gMLP(self):
                 self.assertGreaterEqual(msg["top_1"], 60)
                 self.assertGreaterEqual(msg["top_5"], 85)
 
-    @unittest.skip("Only outputs good accuracy in QNN 2.29")
-    def test_mobilevit_v2(self):
+    def test_mobilevit_v1(self):
         if not self.required_envs([self.image_dataset]):
             self.skipTest("missing required envs")
 
         cmds = [
             "python",
-            f"{self.executorch_root}/examples/qualcomm/oss_scripts/mobilevit_v2.py",
+            f"{self.executorch_root}/examples/qualcomm/oss_scripts/mobilevit_v1.py"
             "--dataset",
             self.image_dataset,
             "--artifact",
@@ -4610,8 +4609,6 @@ def test_mobilevit_v2(self):
         ]
         if self.host:
             cmds.extend(["--host", self.host])
-        if self.shared_buffer:
-            cmds.extend(["--shared_buffer"])
 
         p = subprocess.Popen(cmds, stdout=subprocess.DEVNULL)
         with Listener((self.ip, self.port)) as listener:
@@ -4621,17 +4618,22 @@ def test_mobilevit_v2(self):
             if "Error" in msg:
                 self.fail(msg["Error"])
             else:
-                self.assertGreaterEqual(msg["top_1"], 50)
+                self.assertGreaterEqual(msg["top_1"], 70)
                 self.assertGreaterEqual(msg["top_5"], 85)
 
-    def test_pvt(self):
+    @unittest.skip("Only outputs good accuracy in QNN 2.29")
+    def test_mobilevit_v2(self):
         if not self.required_envs([self.image_dataset]):
             self.skipTest("missing required envs")
 
         cmds = [
             "python",
-            f"{self.executorch_root}/examples/qualcomm/oss_scripts/pvt.py",
+            f"{self.executorch_root}/examples/qualcomm/oss_scripts/mobilevit_v2.py",
+            "--dataset",
             self.image_dataset,
+            "--artifact",
+            self.artifact_dir,
+            "--build_folder",
             self.build_folder,
             "--device",
             self.device,
@@ -4644,6 +4646,8 @@ def test_pvt(self):
         ]
         if self.host:
             cmds.extend(["--host", self.host])
+        if self.shared_buffer:
+            cmds.extend(["--shared_buffer"])
 
         p = subprocess.Popen(cmds, stdout=subprocess.DEVNULL)
         with Listener((self.ip, self.port)) as listener:
@@ -4653,21 +4657,17 @@ def test_pvt(self):
             if "Error" in msg:
                 self.fail(msg["Error"])
             else:
-                self.assertGreaterEqual(msg["top_1"], 65)
+                self.assertGreaterEqual(msg["top_1"], 50)
                 self.assertGreaterEqual(msg["top_5"], 85)
 
-    def test_mobilevit1(self):
+    def test_pvt(self):
         if not self.required_envs([self.image_dataset]):
             self.skipTest("missing required envs")
 
         cmds = [
             "python",
-            f"{self.executorch_root}/examples/qualcomm/oss_scripts/mobilevit1.py"
-            "--dataset",
+            f"{self.executorch_root}/examples/qualcomm/oss_scripts/pvt.py",
             self.image_dataset,
-            "--artifact",
-            self.artifact_dir,
-            "--build_folder",
             self.build_folder,
             "--device",
             self.device,
@@ -4689,7 +4689,7 @@ def test_mobilevit1(self):
             if "Error" in msg:
                 self.fail(msg["Error"])
             else:
-                self.assertGreaterEqual(msg["top_1"], 70)
+                self.assertGreaterEqual(msg["top_1"], 65)
                 self.assertGreaterEqual(msg["top_5"], 85)
 
     def test_regnet(self):

examples/qualcomm/oss_scripts/deit.py

@@ -6,10 +6,12 @@
 
 import getpass
 import json
+import logging
 import os
 from multiprocessing.connection import Client
 
 import numpy as np
+import torch
 from executorch.backends.qualcomm._passes.qnn_pass_manager import (
     get_capture_program_passes,
 )
@@ -46,16 +48,23 @@ def main(args):
     data_num = 100
     height = config.image_size
     width = config.image_size
-    inputs, targets, input_list = get_imagenet_dataset(
-        dataset_path=f"{args.dataset}",
-        data_size=data_num,
-        image_shape=(height, width),
-        crop_size=(height, width),
-    )
+
+    if args.ci:
+        inputs = [(torch.rand(1, 3, height, width),)]
+        logging.warning(
+            "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(
+            dataset_path=f"{args.dataset}",
+            data_size=data_num,
+            image_shape=(height, width),
+            crop_size=(height, width),
+        )
 
     # Get the Deit model.
     model = get_instance()
-    pte_filename = "deit_qnn"
+    pte_filename = "deit_qnn_q8"
 
     # lower to QNN
     passes_job = get_capture_program_passes()
@@ -120,8 +129,8 @@ def main(args):
     parser.add_argument(
         "-a",
         "--artifact",
-        help="path for storing generated artifacts and output by this example. Default ./deit_qnn",
-        default="./deit_qnn",
+        help="path for storing generated artifacts and output by this example. Default ./deit",
+        default="./deit",
         type=str,
     )
 
@@ -134,7 +143,7 @@ def main(args):
             "for https://www.kaggle.com/datasets/ifigotin/imagenetmini-1000)"
         ),
         type=str,
-        required=True,
+        required=False,
     )
 
     args = parser.parse_args()

examples/qualcomm/oss_scripts/mobilevit1.py → examples/qualcomm/oss_scripts/mobilevit_v1.py

@@ -84,7 +84,7 @@ def main(args):
         .to("cpu")
     )
 
-    pte_filename = "mobilevit1_qnn_q16"
+    pte_filename = "mobilevit_v1_qnn_q16"
     build_executorch_binary(
         module.eval(),
         inputs[0],
@@ -157,8 +157,8 @@ def main(args):
         "-a",
         "--artifact",
         help="path for storing generated artifacts by this example. "
-        "Default ./mobilevit1",
-        default="./mobilevit1",
+        "Default ./mobilevit_v1",
+        default="./mobilevit_v1",
         type=str,
     )
 

examples/qualcomm/oss_scripts/roberta.py

@@ -6,6 +6,7 @@
 
 import getpass
 import json
+import logging
 import os
 from multiprocessing.connection import Client
 
@@ -38,16 +39,29 @@ def main(args):
     skip_node_id_set, skip_node_op_set = parse_skip_delegation_node(args)
 
     os.makedirs(args.artifact, exist_ok=True)
-    data_size = 100
 
     tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")
-    inputs, targets, input_list = get_masked_language_model_dataset(
-        args.dataset, tokenizer, data_size
-    )
+    data_size = 100
+    if args.ci:
+        random_ids = torch.randint(low=0, high=100, size=(1, 100), dtype=torch.int32)
+        attention_mask = torch.ones((1, 100), dtype=torch.float32)
+        inputs = [
+            (
+                random_ids,
+                attention_mask,
+            )
+        ]
+        logging.warning(
+            "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(
+            args.dataset, tokenizer, data_size
+        )
 
     # Get the Roberta model.
     model = get_instance(args)
-    pte_filename = "roberta_qnn"
+    pte_filename = "roberta_qnn_q16"
 
     # lower to QNN
     passes_job = get_capture_program_passes()
@@ -137,7 +151,7 @@ def main(args):
         "-a",
         "--artifact",
         help="path for storing generated artifacts and output by this example. Default ./Roberta_qnn",
-        default="./Roberta_qnn",
+        default="./roberta",
         type=str,
     )
     parser.add_argument(
@@ -149,7 +163,7 @@ def main(args):
             "for https://www.kaggle.com/datasets/mikeortman/wikipedia-sentences"
         ),
         type=str,
-        required=True,
+        required=False,
     )
 
     args = parser.parse_args()

examples/qualcomm/oss_scripts/swin_transformer.py

@@ -89,7 +89,7 @@ def main(args):
 
     data_num = 100
     if args.ci:
-        inputs = [torch.rand(1, 3, 224, 224)]
+        inputs = [(torch.rand(1, 3, 224, 224),)]
         logging.warning(
             "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy."
         )
@@ -181,8 +181,9 @@ def main(args):
     parser.add_argument(
         "-a",
         "--artifact",
-        help="path for storing generated artifacts by this example. " "Default ./swin",
-        default="./swin",
+        help="path for storing generated artifacts by this example. "
+        "Default ./swin_transformer",
+        default="./swin_transformer",
         type=str,
     )