diff --git a/docker/Dockerfile.gpu b/docker/Dockerfile.gpu
index 4e8a8085..8dd7a5d0 100644
--- a/docker/Dockerfile.gpu
+++ b/docker/Dockerfile.gpu
@@ -1,11 +1,13 @@
-FROM nvidia/cuda:12.2.0-runtime-ubi8 as base
+FROM nvidia/cuda:11.7.1-runtime-ubi8 as base
 
 ARG cnlpt_version
 
 RUN yum -y install python39 python39-pip
-RUN pip3.9 install cython torch
+RUN pip3.9 install cython
 RUN pip3.9 install cnlp-transformers==$cnlpt_version
 
+# pytorch can't find the cudnn library with our setup, so just point at it directly
+ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/python3.9/site-packages/nvidia/cudnn/lib/"
 
 WORKDIR /opt/cnlp/
 # this copy is to support the preload of train models in the downstream images
diff --git a/docker/MAINTAINER.md b/docker/MAINTAINER.md
index 48bf2a2a..722703c2 100644
--- a/docker/MAINTAINER.md
+++ b/docker/MAINTAINER.md
@@ -14,9 +14,11 @@ Pass `--help` to see all your options.
 
 ### Local Testing
 Use the `./build.py` script to build the image you care about,
-and then run something like the following, depending on your model:
-```shell
+and then run something like one of the following, depending on your model and processor:
+
+```
 docker run --rm -p 8000:8000 smartonfhir/cnlp-transformers:termexists-latest-cpu
+docker run --rm -p 8000:8000 --gpus all smartonfhir/cnlp-transformers:termexists-latest-gpu
 ```
 
 With that specific example of the `termexists` model, you could smoke test it like so: