diff --git a/.github/workflows/integration.yaml b/.github/workflows/integration.yaml
index 6563997..e70a926 100644
--- a/.github/workflows/integration.yaml
+++ b/.github/workflows/integration.yaml
@@ -1,8 +1,9 @@
 name: Integration test
 
 on:
-  pull_request_review:
-    types: [submitted, edited]
+  push:
+    branches:
+      - master
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
@@ -10,8 +11,6 @@ concurrency:
 
 jobs:
   integration-test:
-    # Integration test only works while the Pull Request is approved or the source repository is trusted repository.
-    if: github.event.review.state == 'APPROVED' || github.event.pull_request.head.repo.full_name == 'aliyun/pai-python-sdk'
     runs-on: ubuntu-latest
     env:
       GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
diff --git a/noxfile.py b/noxfile.py
index 244846e..68d4a0d 100644
--- a/noxfile.py
+++ b/noxfile.py
@@ -50,6 +50,12 @@ def integration(session: Session):
         if os.environ.get(key, value) is not None
     }
 
+    # set worker to 2 * cpu_count (physical cores) if not specified
+    if "-n" not in session.posargs and "--numprocesses" not in session.posargs:
+        pos_args = session.posargs + ["-n", str(os.cpu_count() * 2)]
+    else:
+        pos_args = session.posargs
+
     session.run(
         "pytest",
         "--cov-config=.coveragerc",
@@ -57,7 +63,7 @@ def integration(session: Session):
         "--cov-report=html",
         "--cov=pai",
         os.path.join("tests", "integration"),
-        *session.posargs,
+        *pos_args,
         env=env,
     )
     session.run(
diff --git a/pai/estimator.py b/pai/estimator.py
index 5441692..6510ee2 100644
--- a/pai/estimator.py
+++ b/pai/estimator.py
@@ -29,6 +29,9 @@
 from datetime import datetime
 from typing import Any, Dict, List, Optional, Union
 
+from Tea.exceptions import TeaException
+
+from .api.base import PaginatedResult
 from .api.entity_base import EntityBaseMixin
 from .common import ProviderAlibabaPAI, git_utils
 from .common.consts import INSTANCE_TYPE_LOCAL_GPU, FileSystemInputScheme, JobType
@@ -1521,7 +1524,9 @@ def _normalize_name(name: str) -> str:
         for name, value in self.estimator.hyperparameters.items():
             env[_TrainingEnv.ENV_PAI_HPS_PREFIX + _normalize_name(name)] = str(value)
             user_args.extend(["--" + name, shlex.quote(str(value))])
-        env[_TrainingEnv.ENV_PAI_USER_ARGS] = shlex.join(user_args)
+        env[_TrainingEnv.ENV_PAI_USER_ARGS] = " ".join(
+            [shlex.quote(v) for v in user_args]
+        )
         env[_TrainingEnv.ENV_PAI_HPS] = json.dumps(
             {name: str(value) for name, value in self.estimator.hyperparameters.items()}
         )
@@ -1878,15 +1883,27 @@ def __init__(
         self._future = None
         self._stop = False
 
-    def _list_logs(self):
-        page_number, page_offset = 1, 0
-        # print training job logs.
-        while not self._stop:
+    def _list_logs_api(self, page_number=1):
+        try:
             res = self.session.training_job_api.list_logs(
                 self.training_job_id,
                 page_number=page_number,
                 page_size=self.page_size,
             )
+            return res
+        except TeaException as e:
+            # hack: Backend service may raise an exception when the training job
+            # instance is not found.
+            if e.code == "TRAINING_JOB_INSTANCE_NOT_FOUND":
+                return PaginatedResult(items=[], total_count=0)
+            else:
+                raise e
+
+    def _list_logs(self):
+        page_number, page_offset = 1, 0
+        # print training job logs.
+        while not self._stop:
+            res = self._list_logs_api(page_number=page_number)
             # 1. move to next page
             if len(res.items) == self.page_size:
                 # print new logs starting from page_offset
@@ -1904,11 +1921,7 @@ def _list_logs(self):
         # When _stop is True, wait and print remaining logs.
         time.sleep(10)
         while True:
-            res = self.session.training_job_api.list_logs(
-                self.training_job_id,
-                page_number=page_number,
-                page_size=self.page_size,
-            )
+            res = self._list_logs_api(page_number=page_number)
             # There maybe more logs in the next page
             if len(res.items) == self.page_size:
                 self._print_logs(logs=res.items[page_offset:])
diff --git a/pai/huggingface/model.py b/pai/huggingface/model.py
index 7b46b7a..1771470 100644
--- a/pai/huggingface/model.py
+++ b/pai/huggingface/model.py
@@ -259,7 +259,7 @@ def _get_supported_tf_versions_for_inference(self) -> List[str]:
         return res
 
     def _get_latest_tf_version_for_inference(self) -> str:
-        """Return the latest Transformers version for inference."""
+        """Return the latest transformers version for inference."""
         res = self._get_supported_tf_versions_for_inference()
         return max(
             res,
diff --git a/pai/model.py b/pai/model.py
index b631231..f480561 100644
--- a/pai/model.py
+++ b/pai/model.py
@@ -987,12 +987,18 @@ def _deploy_local(
 
         # build command to install requirements
         if requirements_list:
-            install_requirements = shlex.join(
-                ["python", "-m", "pip", "install"] + requirements_list
+            install_requirements = " ".join(
+                [
+                    shlex.quote(s)
+                    for s in ["python", "-m", "pip", "install"] + requirements_list
+                ]
             )
         elif requirements_path:
-            install_requirements = shlex.join(
-                ["python", "-m", "pip", "install", "-r", requirements_path]
+            install_requirements = " ".join(
+                [
+                    shlex.quote(s)
+                    for s in ["python", "-m", "pip", "install", "-r", requirements_path]
+                ]
             )
         else:
             install_requirements = ""