use torch wheel that has patched fixes (#396)

serving/docker/deepspeed.Dockerfile

@@ -12,7 +12,7 @@
 ARG version=11.6.1-cudnn8-devel-ubuntu20.04
 FROM nvidia/cuda:$version
 ARG djl_version=0.20.0~SNAPSHOT
-ARG torch_version=1.12.1
+ARG torch_wheel="https://aws-pytorch-unified-cicd-binaries.s3.us-west-2.amazonaws.com/r1.12.1_ec2/20221208-234008/d3dae914337cde7e182d28544aed5efce29255c4/torch-1.12.1%2Bcu116-cp38-cp38-linux_x86_64.whl"
 ARG accelerate_version=0.13.2
 ARG deepspeed_wheel="https://publish.djl.ai/deepspeed/deepspeed-0.7.5%2Bbf16-py2.py3-none-any.whl"
 ARG transformers_version=4.23.1
@@ -44,7 +44,7 @@ RUN apt-get update && \
     scripts/install_python.sh && \
     scripts/install_s5cmd.sh x64 && \
     DEBIAN_FRONTEND=noninteractive apt-get install -yq libaio-dev libopenmpi-dev && \
-    pip3 install torch==${torch_version} --extra-index-url https://download.pytorch.org/whl/cu116 && \
+    pip3 install ${torch_wheel} && \
     pip3 install ${deepspeed_wheel} &&  \
     pip3 install transformers==${transformers_version} && \
     pip3 install triton==2.0.0.dev20221202 mpi4py sentencepiece accelerate==${accelerate_version} bitsandbytes && \