undefined symbol: cublasLtGetStatusString at Endpoint Deployment Error

Question

Hi

I am getting this error while deploying an end point. This was working fine for 3 months and I had to rebuild due to a minor change and started getting this error.

File "/azureml-envs/tensorflow-2.7/lib/python3.8/site-packages/azureml_inference_server_http/server/user_script.py", line 81, in load_script
main_module_spec.loader.exec_module(user_module)
File "", line 843, in exec_module
File "", line 219, in _call_with_frames_removed
File "/var/azureml-app/221123171914-1667710269/score.py", line 6, in
import ktrain
File "/azureml-envs/tensorflow-2.7/lib/python3.8/site-packages/torch/init.py", line 191, in
_load_global_deps()
File "/azureml-envs/tensorflow-2.7/lib/python3.8/site-packages/torch/init.py", line 153, in _load_global_deps
ctypes.CDLL(lib_path, mode=ctypes.RTLD_GLOBAL)
File "/azureml-envs/tensorflow-2.7/lib/python3.8/ctypes/init.py", line 373, in init
self._handle = _dlopen(self._name, mode)
OSError: /azureml-envs/tensorflow-2.7/lib/python3.8/site-packages/torch/lib/../../nvidia/cublas/lib/libcublas.so.11: undefined symbol: cublasLtGetStatusString, version libcublasLt.so.11

This is my environment:

FROM mcr.microsoft.com/azureml/openmpi4.1.0-cuda11.2-cudnn8-ubuntu20.04:20220714.v1

ENV AZUREML_CONDA_ENVIRONMENT_PATH /azureml-envs/tensorflow-2.7

Create conda environment

RUN conda create -p $AZUREML_CONDA_ENVIRONMENT_PATH \
python=3.8 pip=20.2.4

Prepend path to AzureML conda environment

ENV PATH $AZUREML_CONDA_ENVIRONMENT_PATH/bin:$PATH

Install pip dependencies

RUN HOROVOD_WITH_TENSORFLOW=1 pip install 'matplotlib~=3.5.0' \
'psutil~=5.8.0' \
'tqdm~=4.62.0' \
'scipy~=1.7.0' \
'numpy~=1.21.0' \
'ipykernel~=6.0' \
# upper bound azure-core to address typing-extensions conflict
'azure-core<1.23.0' \
'azureml-core==1.43.0' \
'azureml-defaults==1.43.0' \
'azureml-mlflow==1.43.0.post1' \
'azureml-telemetry==1.43.0' \
'azureml-inference-server-http==0.7.2' \
'pandas==1.4.1' \
'ktrain==0.30.0' \
'sentence-transformers==2.1.0' \
'tensorflow==2.7.0' \
'tokenizers==0.10.3' \
'protobuf~=3.19.1' \
'Flask==2.1.0' \
'transformers==4.10.3'

This is needed for mpi to locate libpython

ENV LD_LIBRARY_PATH $AZUREML_CONDA_ENVIRONMENT_PATH/lib:$LD_LIBRARY_PATH

Answer

Hi @romungi-MSFT

I managed to solve it

kept the container same as 20220729.v1
and added 'torch==1.12.0' \ 'torchvision==0.13.0' to the list. Didn't need the higher version as given in the link.
Thanks for that link

final env file

FROM mcr.microsoft.com/azureml/openmpi4.1.0-cuda11.2-cudnn8-ubuntu20.04:20220729.v1

ENV AZUREML_CONDA_ENVIRONMENT_PATH /azureml-envs/tensorflow-2.7

Create conda environment

RUN conda create -p $AZUREML_CONDA_ENVIRONMENT_PATH \
python=3.8 pip=20.2.4

Prepend path to AzureML conda environment

ENV PATH $AZUREML_CONDA_ENVIRONMENT_PATH/bin:$PATH

Install pip dependencies

RUN HOROVOD_WITH_TENSORFLOW=1 pip install 'matplotlib~=3.5.0' \
'psutil~=5.8.0' \
'tqdm~=4.62.0' \
'scipy~=1.7.0' \
'numpy~=1.21.0' \
'ipykernel~=6.0' \
# upper bound azure-core to address typing-extensions conflict
'azure-core<1.23.0' \
'azureml-core==1.43.0' \
'azureml-defaults==1.43.0' \
'azureml-mlflow==1.43.0.post1' \
'azureml-telemetry==1.43.0' \
'azureml-inference-server-http==0.7.2' \
'pandas==1.4.1' \
'ktrain==0.30.0' \
'sentence-transformers==2.1.0' \
'tensorflow==2.7.0' \
'tokenizers==0.10.3' \
'protobuf~=3.19.1' \
'Flask==2.1.0' \
'transformers==4.10.3' \
'torch==1.12.0' \
'torchvision==0.13.0'

This is needed for mpi to locate libpython

ENV LD_LIBRARY_PATH $AZUREML_CONDA_ENVIRONMENT_PATH/lib:$LD_LIBRARY_PATH

Share via

undefined symbol: cublasLtGetStatusString at Endpoint Deployment Error

Create conda environment

Prepend path to AzureML conda environment

Install pip dependencies

This is needed for mpi to locate libpython

1 answer

Create conda environment

Prepend path to AzureML conda environment

Install pip dependencies

This is needed for mpi to locate libpython

Your answer