diff --git a/doc/scheduler-job-examples.tex b/doc/scheduler-job-examples.tex
index 7b8ebc6..2a50b3a 100644
--- a/doc/scheduler-job-examples.tex
+++ b/doc/scheduler-job-examples.tex
@@ -264,6 +264,25 @@ \subsection{Scheduling On The GPU Nodes}
 %And that there are no more GPUs available on that node (\texttt{hc:gpu=0}).
 %Note that no more than two GPUs can be requested for any one job. 
 
+% ------------------------------------------------------------------------------
+\subsubsection{P6 on Multi-GPU, Multi-Node}
+
+As described lines above, P6 cards are not compatible with Distribute and DataParallel functions
+(\texttt{Pytorch, Tensorflow}) when running on Multi-GPUs.
+One workaround is to run the job in Multi-node, single GPU per node; per example:
+
+\begin{verbatim}
+#SBATCH --nodes=2
+#SBATCH --gpus-per-node=1
+\end{verbatim}
+
+On P6 nodes: \texttt{speed-05, speed-17, speed-01}
+
+The example: 
+  \href{https://github.com/NAG-DevOps/speed-hpc/blob/master/src/pytorch-multinode-multigpu.sh}
+  {pytorch-multinode-multigpu.sh}
+illustrates a job for training on Multi-nodes, Multi-GPUs
+
 % ------------------------------------------------------------------------------
 \subsubsection{CUDA}
 
diff --git a/doc/scheduler-scripting.tex b/doc/scheduler-scripting.tex
index aa19c6c..d695d89 100644
--- a/doc/scheduler-scripting.tex
+++ b/doc/scheduler-scripting.tex
@@ -601,7 +601,7 @@ \subsubsection{Jupyter Notebooks}
 Create an \tool{ssh} tunnel between your computer and the node (\texttt{speed-XX}) where Jupyter is
 running (Using \texttt{speed-submit} as a ``jump server'') (Preferably: PuTTY, see \xf{fig:putty1} and \xf{fig:putty2})
 \begin{verbatim}
-ssh -L 8888:localhost:8888 speed-XX
+ssh -L 8888:speed-XX:8888 YOUR_USER@speed-submit.encs.concordia.ca
 \end{verbatim}
 Don't close the tunnel.
 
diff --git a/doc/speed-manual.pdf b/doc/speed-manual.pdf
index 474738f..eb4731c 100644
Binary files a/doc/speed-manual.pdf and b/doc/speed-manual.pdf differ
diff --git a/src/README.md b/src/README.md
index ce5d11b..0072959 100644
--- a/src/README.md
+++ b/src/README.md
@@ -21,6 +21,7 @@ These are examples either trivial or some are more elaborate. Some are described
   - `efficientdet.sh` -- `efficientdet` with Conda environment described below
   - `gurobi-with-python.sh` -- using Gurobi with Python and Python virtual environment
   - `pytorch-multicpu.txt` -- using Pytorch with Python virtual environment to run on CPUs; with instructions and code ready to paste.
+  - `pytorch-multinode-multigpu.sh` -- using Pytorch with Python virtual environment to run on Multinodes and MultiGpus
   - `lambdal-singularity.sh` -- an example use of the Singularity container to run LambdaLabs software stack on the GPU node. The container was built from the docker image as a [source](https://github.com/NAG-DevOps/lambda-stack-dockerfiles).
   - `openiss-reid-speed.sh` -- OpenISS computer vision exame for re-edentification, see [more](https://github.com/NAG-DevOps/speed-hpc/tree/master/src#openiss-reid-tfk) in its section
   - `openiss-yolo-cpu.sh`, `openiss-yolo-gpu.sh`, and `openiss-yolo-interactive.sh` -- OpenISS examples with YOLO, related to `reid`, see [more](https://github.com/NAG-DevOps/speed-hpc/tree/master/src#openiss-yolov3) in the corresponding section
diff --git a/src/pytorch-multinode-multigpu.sh b/src/pytorch-multinode-multigpu.sh
new file mode 100644
index 0000000..aa9989c
--- /dev/null
+++ b/src/pytorch-multinode-multigpu.sh
@@ -0,0 +1,29 @@
+#!/encs/bin/tcsh
+#SBATCH --job-name=pytorch_multinode_multigpu_train
+
+#SBATCH --nodes=2
+#SBATCH --gpus-per-node=1 #On P6 cards this value MUST be: 1
+#SBATCH --cpus-per-task=8
+#SBATCH --ntasks-per-node=1
+#SBATCH --mem=128G             ## Assign memory per node
+
+if ( $?SLURM_CPUS_PER_TASK ) then
+   setenv omp_threads $SLURM_CPUS_PER_TASK
+else
+   setenv omp_threads 1
+endif
+setenv OMP_NUM_THREADS $omp_threads
+
+setenv RDZV_HOST `hostname -s`
+setenv RDZV_PORT 29400
+setenv endpoint ${RDZV_HOST}:${RDZV_PORT}
+setenv CUDA_LAUNCH_BLOCKING 1
+setenv NCCL_BLOCKING_WAIT 1
+#setenv NCCL_DEBUG INFO
+setenv NCCL_P2P_DISABLE 1
+setenv NCCL_IB_DISABLE 1
+source /speed-scratch/$USER/tmp/Venv-Name/bin/activate.csh #path where you have created your python venv
+unsetenv CUDA_VISIBLE_DEVICES
+# nproc_per_node=1 On P6 cards
+srun torchrun --nnodes=$SLURM_JOB_NUM_NODES --nproc_per_node=1 --rdzv_id=$SLURM_JOB_ID --rdzv_backend=c10d --rdzv_endpoint=$endpoint main_multinode.py
+deactivate