more launch scripts

codeaudit · Jul 9, 2022 · b80ed48 · b80ed48
1 parent 529bf6a
commit b80ed48
Show file tree

Hide file tree

Showing 4 changed files with 118 additions and 0 deletions.
diff --git a/scripts/slurm/resume_512_improvedaesthetic/launcher.sh b/scripts/slurm/resume_512_improvedaesthetic/launcher.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+export NODE_RANK=${SLURM_NODEID}
+echo "##########################################"
+echo MASTER_ADDR=${MASTER_ADDR}
+echo MASTER_PORT=${MASTER_PORT}
+echo NODE_RANK=${NODE_RANK}
+echo WORLD_SIZE=${WORLD_SIZE}
+echo "##########################################"
+# debug environment worked great so we stick with it
+# no magic there, just a miniconda python=3.9, pytorch=1.12, cudatoolkit=11.3
+# env with pip dependencies from stable diffusion's requirements.txt
+eval "$(/fsx/stable-diffusion/debug/miniconda3/bin/conda shell.bash hook)"
+conda activate stable
+cd /fsx/stable-diffusion/stable-diffusion
+
+CONFIG=configs/stable-diffusion/txt2img-1p4B-multinode-clip-encoder-high-res-512-improvedaesthetic.yaml
+EXTRA="model.params.ckpt_path=/fsx/stable-diffusion/stable-diffusion/logs/2022-07-07T16-15-18_txt2img-1p4B-multinode-clip-encoder-high-res-512/checkpoints/last.ckpt"
+DEBUG="-d True lightning.callbacks.image_logger.params.batch_frequency=5"
+
+python main.py --base $CONFIG --gpus 0,1,2,3,4,5,6,7 -t --num_nodes ${WORLD_SIZE} --scale_lr False $EXTRA #$DEBUG
diff --git a/scripts/slurm/resume_512_improvedaesthetic/sbatch.sh b/scripts/slurm/resume_512_improvedaesthetic/sbatch.sh
@@ -0,0 +1,39 @@
+#!/bin/bash
+#SBATCH --partition=compute-od-gpu
+#SBATCH --job-name=stable-diffusion-512cont-improvedaesthetic
+#SBATCH --nodes=20
+#SBATCH --gpus-per-node=8
+#SBATCH --cpus-per-gpu=4
+#SBATCH --ntasks-per-node=1
+#SBATCH --output=%x_%j.%n.out
+
+# nccl / efa stuff
+module load intelmpi
+source /opt/intel/mpi/latest/env/vars.sh
+export LD_LIBRARY_PATH=/opt/aws-ofi-nccl/lib:/opt/amazon/efa/lib64:/usr/local/cuda-11.0/efa/lib:/usr/local/cuda-11.0/lib:/usr/local/cuda-11.0/lib64:/usr/local/cuda-11.0:/opt/nccl/build/lib:/opt/aws-ofi-nccl-install/lib:/opt/aws-ofi-nccl/lib:$LD_LIBRARY_PATH
+export NCCL_PROTO=simple
+export PATH=/opt/amazon/efa/bin:$PATH
+export LD_PRELOAD="/opt/nccl/build/lib/libnccl.so"
+export FI_EFA_FORK_SAFE=1
+export FI_LOG_LEVEL=1
+export FI_EFA_USE_DEVICE_RDMA=1 # use for p4dn
+export NCCL_DEBUG=info
+export PYTHONFAULTHANDLER=1
+export CUDA_LAUNCH_BLOCKING=0
+export OMPI_MCA_mtl_base_verbose=1
+export FI_EFA_ENABLE_SHM_TRANSFER=0
+export FI_PROVIDER=efa
+export FI_EFA_TX_MIN_CREDITS=64
+export NCCL_TREE_THRESHOLD=0
+
+# pytorch multinode vars
+# node rank should be set in launcher script
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=11338
+export WORLD_SIZE=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | wc -l)
+
+echo MASTER_ADDR=${MASTER_ADDR}
+echo MASTER_PORT=${MASTER_PORT}
+echo WORLD_SIZE=${WORLD_SIZE}
+
+srun --output=%x_%j.%n.out bash /fsx/stable-diffusion/stable-diffusion/scripts/slurm/resume_512_improvedaesthetic/launcher.sh
diff --git a/scripts/slurm/resume_768_hr/launcher.sh b/scripts/slurm/resume_768_hr/launcher.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+export NODE_RANK=${SLURM_NODEID}
+echo "##########################################"
+echo MASTER_ADDR=${MASTER_ADDR}
+echo MASTER_PORT=${MASTER_PORT}
+echo NODE_RANK=${NODE_RANK}
+echo WORLD_SIZE=${WORLD_SIZE}
+echo "##########################################"
+# debug environment worked great so we stick with it
+# no magic there, just a miniconda python=3.9, pytorch=1.12, cudatoolkit=11.3
+# env with pip dependencies from stable diffusion's requirements.txt
+eval "$(/fsx/stable-diffusion/debug/miniconda3/bin/conda shell.bash hook)"
+conda activate stable
+cd /fsx/stable-diffusion/stable-diffusion
+
+CONFIG=configs/stable-diffusion/txt2img-multinode-clip-encoder-f16-768-laion-hr.yaml
+EXTRA="model.params.ckpt_path=/fsx/stable-diffusion/stable-diffusion/checkpoints/f16-33k+12k-hr_pruned.ckpt"
+DEBUG="-d True lightning.callbacks.image_logger.params.batch_frequency=5"
+
+python main.py --base $CONFIG --gpus 0,1,2,3,4,5,6,7 -t --num_nodes ${WORLD_SIZE} --scale_lr False $EXTRA #$DEBUG
diff --git a/scripts/slurm/resume_768_hr/sbatch.sh b/scripts/slurm/resume_768_hr/sbatch.sh
@@ -0,0 +1,39 @@
+#!/bin/bash
+#SBATCH --partition=compute-od-gpu
+#SBATCH --job-name=stable-diffusion-768cont-resumehr
+#SBATCH --nodes=20
+#SBATCH --gpus-per-node=8
+#SBATCH --cpus-per-gpu=4
+#SBATCH --ntasks-per-node=1
+#SBATCH --output=%x_%j.%n.out
+
+# nccl / efa stuff
+module load intelmpi
+source /opt/intel/mpi/latest/env/vars.sh
+export LD_LIBRARY_PATH=/opt/aws-ofi-nccl/lib:/opt/amazon/efa/lib64:/usr/local/cuda-11.0/efa/lib:/usr/local/cuda-11.0/lib:/usr/local/cuda-11.0/lib64:/usr/local/cuda-11.0:/opt/nccl/build/lib:/opt/aws-ofi-nccl-install/lib:/opt/aws-ofi-nccl/lib:$LD_LIBRARY_PATH
+export NCCL_PROTO=simple
+export PATH=/opt/amazon/efa/bin:$PATH
+export LD_PRELOAD="/opt/nccl/build/lib/libnccl.so"
+export FI_EFA_FORK_SAFE=1
+export FI_LOG_LEVEL=1
+export FI_EFA_USE_DEVICE_RDMA=1 # use for p4dn
+export NCCL_DEBUG=info
+export PYTHONFAULTHANDLER=1
+export CUDA_LAUNCH_BLOCKING=0
+export OMPI_MCA_mtl_base_verbose=1
+export FI_EFA_ENABLE_SHM_TRANSFER=0
+export FI_PROVIDER=efa
+export FI_EFA_TX_MIN_CREDITS=64
+export NCCL_TREE_THRESHOLD=0
+
+# pytorch multinode vars
+# node rank should be set in launcher script
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=11338
+export WORLD_SIZE=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | wc -l)
+
+echo MASTER_ADDR=${MASTER_ADDR}
+echo MASTER_PORT=${MASTER_PORT}
+echo WORLD_SIZE=${WORLD_SIZE}
+
+srun --output=%x_%j.%n.out bash /fsx/stable-diffusion/stable-diffusion/scripts/slurm/resume_768_hr/launcher.sh  # srun vs mpirun?