diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
new file mode 100644
index 0000000000..ec50a345f8
--- /dev/null
+++ b/.github/workflows/build.yml
@@ -0,0 +1,38 @@
+name: Build Docker Images
+
+on:
+  workflow_dispatch:
+    inputs:
+      onmt_version:
+        description: "OpenNMT version"
+        required: true
+        type: string
+  # to facilitate initial tests in PR
+  push:
+    branches:
+      - "docker"
+
+run-name: ${{ github.workflow }} -- ${{ inputs.onmt_version || 'test' }}
+
+env:
+  ONMT_VERSION: ${{ inputs.onmt_version || 'test' }}
+
+jobs:
+  build:
+    runs-on: ubuntu-22.04
+    strategy:
+      matrix:
+        cuda_version: [11.8.0, 12.1.0]
+    permissions: write-all
+    steps:
+    - name: Checkout opennmt repo
+      uses: actions/checkout@v4
+    - name: Login to ghcr
+      uses: docker/login-action@v3
+      with:
+        registry: ghcr.io
+        username: ${{ github.actor }}
+        password: ${{ secrets.GITHUB_TOKEN }}
+    - name: Build
+      run: |
+        docker/build.sh ${{ env.ONMT_VERSION }} ${{ matrix.cuda_version}}
diff --git a/README.md b/README.md
index ec41122105..818e063d89 100644
--- a/README.md
+++ b/README.md
@@ -53,6 +53,30 @@ If you used previous versions of OpenNMT-py, you can check the [Changelog](https
 
 ## Setup
 
+### Using docker
+
+To facilitate setup and reproducibility, some docker images are made available via the Github Container Registry:
+https://github.com/OpenNMT/OpenNMT-py/pkgs/container/opennmt-py
+
+You can adapt the workflow and build your own image(s) depending on specific needs by using `build.sh` and `Dockerfile` in the `docker` directory of the repo.
+
+```
+docker pull ghcr.io/opennmt/opennmt-py:3.4.3-ubuntu22.04-cuda12.1
+```
+
+Example oneliner to run a container and open a bash shell within it
+```
+docker run --rm -it --runtime=nvidia ghcr.io/opennmt/opennmt-py:test-ubuntu22.04-cuda12.1
+```
+Note: you need to have the [Nvidia Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) (formerly nvidia-docker) installed to properly take advantage of the CUDA/GPU features.
+
+Depending on your needs you can add various flags:
+- `-p 5000:5000` to forward some exposed port from your container to your host;
+- `-v /some/local/directory:/some/container/directory` to mount some local directory to some container directory;
+- `--entrypoint some_command` to directly run some specific command as the container entry point (instead of the default bash shell);
+
+### Installing locally
+
 OpenNMT-py requires:
 
 - Python >= 3.8
@@ -78,7 +102,7 @@ Note: if you encounter a `MemoryError` during installation, try to use `pip` wit
 pip install -r requirements.opt.txt
 ```
 
-## Manual installation of some dependencies
+### Manual installation of some dependencies
 
 Apex is highly recommended to have fast performance (especially the legacy fusedadam optimizer and FusedRMSNorm)
 
diff --git a/docker/Dockerfile b/docker/Dockerfile
new file mode 100644
index 0000000000..e469409c58
--- /dev/null
+++ b/docker/Dockerfile
@@ -0,0 +1,52 @@
+ARG CUDA_VERSION=11.8.0
+FROM nvidia/cuda:$CUDA_VERSION-devel-ubuntu22.04
+
+RUN apt-get update && apt-get install -y locales gcc g++ python3-dev
+RUN apt-get update && apt-get install -y \
+    git \
+    python3-pip \
+    python3-dev \
+    libprotobuf-dev \
+    libprotobuf-c-dev
+
+RUN pip3 install --upgrade pip
+RUN pip3 install packaging
+
+# Install torch
+RUN CU=$(echo "${CUDA_VERSION%.*}" | sed 's/\.//g'); pip3 install torch --index-url "https://download.pytorch.org/whl/cu$CU"
+
+# Install apex
+RUN mkdir /setup
+WORKDIR /setup
+RUN git clone https://github.com/nvidia/apex
+WORKDIR /setup/apex
+RUN pip3 install ninja
+ENV TORCH_CUDA_ARCH_LIST="6.0;6.1;6.2;7.0;7.5;8.6"
+RUN pip3 install -v --no-build-isolation \
+    --config-settings --global-option="--cpp_ext" \
+    --config-settings --global-option="--cuda_ext" \
+    --config-settings --global-option="--deprecated_fused_adam" \
+    --global-option="--xentropy" \
+    --global-option="--fast_multihead_attn" \
+    ./
+
+# Install flash-attention
+RUN pip install flash-attn --no-build-isolation
+
+# Install llm-awq
+RUN git clone https://github.com/mit-han-lab/llm-awq && \
+    cd llm-awq && \
+    pip install -e . && \
+    cd ..
+
+# Install AutoAWQ
+RUN pip install autoawq
+
+COPY . /opennmt-py
+WORKDIR /opennmt-py
+RUN pip install -r requirements.opt.txt
+RUN pip install -e .
+
+WORKDIR /
+
+ENTRYPOINT /bin/bash
\ No newline at end of file
diff --git a/docker/build.sh b/docker/build.sh
new file mode 100755
index 0000000000..29d1e2bb7a
--- /dev/null
+++ b/docker/build.sh
@@ -0,0 +1,28 @@
+#!/bin/bash
+#
+# Build and push version X of OpenNMT-py with CUDA Y:
+# ./build.sh X Y
+
+set -e
+
+# allow user to run this script from anywhere
+# from https://stackoverflow.com/a/246128
+# one-liner which will give you the full directory name
+# of the script no matter where it is being called from
+unset CDPATH
+DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
+
+ROOT_DIR=$DIR/..
+cd $ROOT_DIR
+
+ONMT_VERSION="$1"
+CUDA_VERSION="$2"
+[ -z "$CUDA_VERSION" ] && CUDA_VERSION="11.8.0"
+
+IMAGE="ghcr.io/opennmt/opennmt-py"
+TAG="$ONMT_VERSION-ubuntu22.04-cuda${CUDA_VERSION%.*}"
+
+echo "Building $IMAGE:$TAG with CUDA_VERSION=$CUDA_VERSION"
+
+docker build -t $IMAGE:$TAG --progress=plain -f docker/Dockerfile --build-arg CUDA_VERSION=$CUDA_VERSION .
+docker push $IMAGE:$TAG
\ No newline at end of file
diff --git a/onmt/opts.py b/onmt/opts.py
index 861e96172e..b6b34c1c7c 100644
--- a/onmt/opts.py
+++ b/onmt/opts.py
@@ -493,6 +493,13 @@ def distributed_opts(parser):
         type=int,
         help="Port of master for torch.distributed training.",
     )
+    group.add(
+        "--timeout",
+        "-timeout",
+        default=60,
+        type=int,
+        help="Timeout for one GOU to wait for the others.",
+    )
 
 
 def model_opts(parser):
diff --git a/onmt/trainer.py b/onmt/trainer.py
index a570039707..d580b19fd1 100644
--- a/onmt/trainer.py
+++ b/onmt/trainer.py
@@ -328,9 +328,10 @@ def train(
             )
 
             if valid_iter is not None and step % valid_steps == 0:
-                valid_stats = self.validate(
-                    valid_iter, moving_average=self.moving_average
-                )
+                if self.parallel_mode == "tensor_parallel" or self.gpu_rank <= 0:
+                    valid_stats = self.validate(
+                        valid_iter, moving_average=self.moving_average
+                    )
 
             if step % valid_steps == 0 and self.gpu_rank <= 0:
                 self._report_step(
diff --git a/onmt/utils/distributed.py b/onmt/utils/distributed.py
index c7e6051a7c..e6779c397f 100644
--- a/onmt/utils/distributed.py
+++ b/onmt/utils/distributed.py
@@ -29,7 +29,7 @@ def multi_init(opt, device_id):
         init_method=dist_init_method,
         world_size=dist_world_size,
         rank=opt.gpu_ranks[device_id],
-        timeout=timedelta(seconds=60),
+        timeout=timedelta(seconds=opt.timeout),
     )
     gpu_rank = torch.distributed.get_rank()
     if not is_master(opt, device_id):
diff --git a/requirements.opt.txt b/requirements.opt.txt
index f2d13037ca..cdd42a0cbd 100644
--- a/requirements.opt.txt
+++ b/requirements.opt.txt
@@ -6,3 +6,4 @@ scipy
 bitsandbytes>=0.41.2
 safetensors
 spacy
+gradio