fix a profiler bug

docker/Dockerfile.rocm

@@ -19,33 +19,24 @@ FROM docker.io/rocm/pytorch:rocm5.6_ubuntu20.04_py3.8_pytorch_2.0.1
 ENV UBSAN_OPTIONS=print_stacktrace=1
 ENV LC_ALL=C.UTF-8
 ENV LANG=C.UTF-8
-ADD ./docker/install/rocm_dev-requirements.txt dev-requirements.txt
 RUN groupadd -f render
 
 WORKDIR /
 
-ADD ./docker/install/ /Install
+RUN git clone -b merge_upstream https://github.com/ROCmSoftwarePlatform/AITemplate.git
+
+WORKDIR /AITemplate
 # necessary package
-RUN bash /Install/install_basic_dep.sh
+RUN bash ./docker/install/install_basic_dep.sh
 
 # for test
-RUN bash /Install/install_test_dep.sh
+RUN bash ./docker/install/install_test_dep.sh
 
 # for docs
-RUN bash /Install/install_doc_dep.sh
+RUN bash ./docker/install/install_doc_dep.sh
 
 # for detection
 RUN DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata
-RUN bash /Install/install_detection_deps.sh
-
-# Copy AITemplate to Docker
-RUN mkdir /AITemplate
-ADD ./COMMIT_INFO /AITemplate/COMMIT_INFO
-ADD ./python /AITemplate/python
-ADD ./3rdparty /AITemplate/3rdparty
-ADD ./examples /AITemplate/examples
-ADD ./tests /AITemplate/tests
-ADD ./docs /AITemplate/docs
-ADD ./static /AITemplate/static
-ADD ./licenses /AITemplate/licenses
-RUN cd /AITemplate/python && python setup.py install
+RUN bash ./docker/install/install_detection_deps.sh
+
+RUN cd ./python && python setup.py install

python/aitemplate/backend/profiler_runner.py

@@ -122,13 +122,6 @@ def process_task(task: Task) -> None:
 
     if len(stderr) > 0:
         # TODO: ugly fix, should remove when finish all profiler refactor
-        runtimes = PROF_RUNTIME_PATTERN.findall(stdout)
-        if len(runtimes) > 0:
-            single_file_profiler = True
-        if not single_file_profiler:
-            task._failed = True
-            return
-
         _LOGGER.debug(
             "Failed: [{name}][{algo}]\ncmd:\n{cmd}\nstderr:\n{stderr}".format(
                 name=task._name,
@@ -137,6 +130,13 @@ def process_task(task: Task) -> None:
                 stderr=stderr,
             ),
         )
+        runtimes = PROF_RUNTIME_PATTERN.findall(stdout)
+        if len(runtimes) > 0:
+            single_file_profiler = True
+        if not single_file_profiler:
+            task._failed = True
+            return
+
     task._ret, task._failed = extract_profile_result(
         stdout=stdout,
         return_ops=task._kwargs.get("return_ops", None),

python/aitemplate/backend/rocm/conv2d/common.py

@@ -284,7 +284,7 @@
   const int64_t stride = std::stoi(argv[8]);
   const int64_t pad = std::stoi(argv[9]);
   const int64_t dilation = std::stoi(argv[10]);
-  const int64_t group = std::stoi(argv[11]);
+  const int64_t group = std::stoi(argv[14]);
 """
 )