Nsight (#343)

* Added NCCL Nsight files Signed-off-by: Ankur Srivastava <awsankur@amazon.com> * Added NCCL Nsight files Signed-off-by: Ankur Srivastava <awsankur@amazon.com> * Updateed nccl and readme and fsdp Signed-off-by: Ankur Srivastava <awsankur@amazon.com> * Updated NCCL Signed-off-by: Ankur Srivastava <awsankur@amazon.com> * Added almost everything Signed-off-by: Ankur Srivastava <awsankur@amazon.com> * Removed token Signed-off-by: Ankur Srivastava <awsankur@amazon.com> --------- Signed-off-by: Ankur Srivastava <awsankur@amazon.com>
aws-samples · Jun 3, 2024 · 44e448e · 44e448e
1 parent 765e627
commit 44e448e
Show file tree

Hide file tree

Showing 22 changed files with 1,269 additions and 19 deletions.
diff --git a/4.validation_and_observability/5.nsight/2.generate_recipes.sh b/4.validation_and_observability/5.nsight/2.generate_recipes.sh
@@ -0,0 +1,20 @@
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_api_sum --input ${NSIGHT_REPORT_NAME}.nsys-rep
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_api_sync --input ${NSIGHT_REPORT_NAME}.nsys-rep
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_gpu_kern_pace --input ${NSIGHT_REPORT_NAME}.nsys-rep --name ncclDevKernel_ReduceScatter_Sum_f32_RING_LL
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_gpu_kern_pace --input ${NSIGHT_REPORT_NAME}.nsys-rep --name ncclDevKernel_AllGather_RING_LL
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_gpu_kern_sum --input ${NSIGHT_REPORT_NAME}.nsys-rep
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_gpu_mem_size_sum --input ${NSIGHT_REPORT_NAME}.nsys-rep
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_gpu_mem_time_sum --input ${NSIGHT_REPORT_NAME}.nsys-rep
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe cuda_gpu_time_util_map --input ${NSIGHT_REPORT_NAME}.nsys-rep
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe nccl_sum --input ${NSIGHT_REPORT_NAME}.nsys-rep
+
+/fsx/nsight-efa/target-linux-x64/nsys recipe nccl_gpu_time_util_map --input ${NSIGHT_REPORT_NAME}.nsys-rep
diff --git a/4.validation_and_observability/5.nsight/EKS/Dockerfile.llama2-efa b/4.validation_and_observability/5.nsight/EKS/Dockerfile.llama2-efa
@@ -0,0 +1,198 @@
+apiVersion: "kubeflow.org/v1"
+kind: PyTorchJob
+metadata:
+  name: fsdp
+spec:
+  elasticPolicy:
+    rdzvBackend: etcd
+    rdzvHost: etcd
+    rdzvPort: 2379
+    minReplicas: 1
+    maxReplicas: 96
+    maxRestarts: 100
+    #metrics:
+    #  - type: Resource
+    #    resource:
+    #      name: cpu
+    #      target:
+    #        type: Utilization
+    #        averageUtilization: 80
+  pytorchReplicaSpecs:
+    Worker:
+      replicas: 2
+      restartPolicy: OnFailure
+      template:
+        metadata:
+          labels:
+            app: fsdp
+            nvidia-devtools-sidecar-injector: enabled
+        spec:
+          volumes:
+            - name: shmem
+              #emptyDir:
+              #  medium: Memory
+              hostPath:
+                path: /dev/shm
+          #nodeSelector:
+          #  node.kubernetes.io/instance-type: "p5.48xlarge"
+          containers:
+            - name: pytorch
+              image: 159553542841.dkr.ecr.us-west-2.amazonaws.com/fsdp:llama2-efa-main-02-13
+              imagePullPolicy: Always
+              resources:
+                requests:
+                  nvidia.com/gpu:
+                  vpc.amazonaws.com/efa: 4
+                limits:
+                  nvidia.com/gpu:
+                  vpc.amazonaws.com/efa: 4
+              env:
+              # for P5 FI_* should be commented out
+              #- name: LOGLEVEL
+              #  value: "DEBUG"
+              - name: FI_PROVIDER
+                value: efa
+              - name: FI_EFA_USE_DEVICE_RDMA
+                value: "1"
+              - name: FI_EFA_FORK_SAFE
+                value: "1"
+              - name: FI_LOG_LEVEL
+                value: "1"
+              - name: FI_EFA_ENABLE_SHM_TRANSFER
+                value: "1"
+             #- name: NCCL_DEBUG
+             #   value: "INFO"
+              - name: NCCL_ASYNC_ERROR_HANDLING
+                value: "1"
+              #- name: NCCL_IGNORE_DISABLED_P2P
+              #  value: "1"
+              - name: HF_TOKEN
+                value: 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
+              command:
+                - bash
+                - -c
+                - "torchrun --nproc_per_node=8 --nnodes=2 examples/finetuning.py --num_epochs=1 --batch_size_training=3 --enable_fsdp --pure_bf16 --model_name meta-llama/Llama-2-7b-hf --output_dir ."
+              volumeMounts:
+                - name: shmem
+                  mountPath: /dev/shm
+root@cb9511473ccc:/eks/deployment/distributed-training/pytorch/pytorchjob/fsdp# cat Dockerfile.llama2-efa
+FROM nvidia/cuda:12.2.2-devel-ubuntu22.04
+
+ARG EFA_INSTALLER_VERSION=1.29.1
+ARG AWS_OFI_NCCL_VERSION=v1.7.3-aws
+ARG NCCL_TESTS_VERSION=master
+ARG NCCL_VERSION=2.18.5
+
+RUN apt-get update -y
+RUN apt-get remove -y --allow-change-held-packages \
+    libmlx5-1 ibverbs-utils libibverbs-dev libibverbs1 libnccl2 libnccl-dev
+
+RUN rm -rf /opt/hpcx \
+    && rm -rf /usr/local/mpi \
+    && rm -f /etc/ld.so.conf.d/hpcx.conf \
+    && ldconfig
+ENV OPAL_PREFIX=
+
+RUN DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
+    git \
+    gcc \
+    vim \
+    kmod \
+    openssh-client \
+    openssh-server \
+    build-essential \
+    curl \
+    autoconf \
+    libtool \
+    gdb \
+    automake \
+    python3-distutils \
+    cmake \
+    apt-utils \
+    devscripts \
+    debhelper \
+    libsubunit-dev \
+    check \
+    pkg-config
+
+RUN mkdir -p /var/run/sshd
+RUN sed -i 's/[ #]\(.*StrictHostKeyChecking \).*/ \1no/g' /etc/ssh/ssh_config && \
+    echo "    UserKnownHostsFile /dev/null" >> /etc/ssh/ssh_config && \
+    sed -i 's/#\(StrictModes \).*/\1no/g' /etc/ssh/sshd_config
+ENV LD_LIBRARY_PATH /usr/local/cuda/extras/CUPTI/lib64:/opt/amazon/openmpi/lib:/opt/nccl/build/lib:/opt/amazon/efa/lib:/opt/aws-ofi-nccl/install/lib:/usr/local/lib:$LD_LIBRARY_PATH
+ENV PATH /opt/amazon/openmpi/bin/:/opt/amazon/efa/bin:/usr/bin:/usr/local/bin:$PATH
+
+RUN curl https://bootstrap.pypa.io/get-pip.py -o /tmp/get-pip.py \
+    && python3 /tmp/get-pip.py \
+    && pip3 install awscli pynvml
+
+#################################################
+## Install NVIDIA GDRCopy
+#RUN git clone https://github.com/NVIDIA/gdrcopy.git /opt/gdrcopy \
+#    && cd /opt/gdrcopy \
+#    && make lib_install install \
+#    && cd /opt/gdrcopy/tests \
+#    && make \
+#    && mv copylat copybw sanity apiperf /usr/bin/
+
+#################################################
+## Install EFA installer
+RUN cd $HOME \
+    && curl -O https://efa-installer.amazonaws.com/aws-efa-installer-${EFA_INSTALLER_VERSION}.tar.gz \
+    && tar -xf $HOME/aws-efa-installer-${EFA_INSTALLER_VERSION}.tar.gz \
+    && cd aws-efa-installer \
+    && ./efa_installer.sh -y -g -d --skip-kmod --skip-limit-conf --no-verify \
+    && rm -rf $HOME/aws-efa-installer
+
+###################################################
+## Install NCCL
+RUN git clone https://github.com/NVIDIA/nccl -b v${NCCL_VERSION}-1 /opt/nccl \
+    && cd /opt/nccl \
+    && make -j $(nproc) src.build CUDA_HOME=/usr/local/cuda \
+    NVCC_GENCODE="-gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_86,code=sm_86 -gencode=arch=compute_90,code=sm_90"
+
+###################################################
+## Install AWS-OFI-NCCL plugin
+RUN apt-get install libtool autoconf cmake nasm unzip pigz parallel nfs-common build-essential hwloc libhwloc-dev libjemalloc2 libnuma-dev numactl libjemalloc-dev preload htop iftop liblapack-dev libgfortran5 ipcalc wget curl devscripts debhelper check libsubunit-dev fakeroot pkg-config dkms -y
+RUN export OPAL_PREFIX="" \
+    && git clone https://github.com/aws/aws-ofi-nccl.git /opt/aws-ofi-nccl \
+    && cd /opt/aws-ofi-nccl \
+    && git checkout ${AWS_OFI_NCCL_VERSION} \
+    && ./autogen.sh \
+    && ./configure --prefix=/opt/aws-ofi-nccl/install \
+    --with-libfabric=/opt/amazon/efa/ \
+    --with-cuda=/usr/local/cuda \
+    --with-nccl=/opt/nccl/build \
+    --with-mpi=/opt/amazon/openmpi/ \
+    && make -j $(nproc) && make install
+###################################################
+## Install fsdp
+
+RUN mkdir -p /workspace/
+
+WORKDIR /workspace
+
+#RUN git clone -b flop_counter https://github.com/facebookresearch/llama-recipes.git
+#RUN git clone -b flop_counter_gc https://github.com/facebookresearch/llama-recipes.git
+RUN git clone https://github.com/facebookresearch/llama-recipes.git
+
+WORKDIR /workspace/llama-recipes
+
+RUN pip3 install -U pip setuptools
+
+RUN pip3 install fsspec==2023.1.0
+RUN pip3 install huggingface_hub==0.17.0
+RUN pip3 install -r requirements.txt
+
+RUN pip3 install -e .
+
+RUN pip3 install tabulate
+
+RUN pip3 install protobuf
+
+RUN pip3 install python-etcd
+
+#RUN pip3 uninstall -y torch
+#RUN pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
+
+ENV PYTHONPATH="${PYTHONPATH}:/workspace/llama-recipes/src"
diff --git a/4.validation_and_observability/5.nsight/EKS/custom_values.yaml b/4.validation_and_observability/5.nsight/EKS/custom_values.yaml
@@ -0,0 +1,35 @@
+# If we dont specify the Nsight image, 2024.2 version is used by default. 
+# Will use 2024.4 version which is planned to be released by 5/24/2024
+devtoolBinariesImage:
+  image: ${REGISTRY}.dkr.ecr.${REGION}.amazonaws.com/nsight-systems-cli:2024.4.1-ubuntu22.04
+  imagePullPolicy: Always
+
+# Assuming EKS cluster has a FSx for Lustre filesystem mounted on it. Nsight reports will be saved in /fsx_shared
+profile:
+  volumes:
+    [
+      {
+        "name": "nsys-output-volume",
+        "persistentVolumeClaim": { "claimName": "fsx-pvc" }
+      }
+    ]
+  volumeMounts:
+    [
+      {
+        "name": "nsys-output-volume",
+        "mountPath": "/fsx_shared"
+      }
+    ]
+
+  # CLI options: https://docs.nvidia.com/nsight-systems/UserGuide/index.html#cli-command-switches
+  # delay and duration values in secs
+
+  # Use %{} to include environment variables in the Nsight report filename
+
+  # The arguments for the Nsight Systems. The placeholders will be replaced with the actual values.
+  devtoolArgs: "profile --force-overwrite true --trace nvtx,cuda  --delay 150 --duration 60 \
+  -o /fsx_shared/fsdp/auto_{PROCESS_NAME}_%{POD_FULLNAME}_%{CONTAINER_NAME}_{TIMESTAMP}_{UID}.nsys-rep"
+
+  injectionMatch: "^/usr/bin/python3 /usr/local/bin/torchrun.*$"
+  #injectionMatch: "^.*torchrun.*$"
+
diff --git a/4.validation_and_observability/5.nsight/EKS/fsdp.yaml b/4.validation_and_observability/5.nsight/EKS/fsdp.yaml
@@ -0,0 +1,77 @@
+apiVersion: "kubeflow.org/v1"
+kind: PyTorchJob
+metadata:
+  name: fsdp
+spec:
+  elasticPolicy:
+    rdzvBackend: etcd
+    rdzvHost: etcd
+    rdzvPort: 2379
+    minReplicas: 1
+    maxReplicas: 96
+    maxRestarts: 100
+    #metrics:
+    #  - type: Resource
+    #    resource:
+    #      name: cpu
+    #      target:
+    #        type: Utilization
+    #        averageUtilization: 80
+  pytorchReplicaSpecs:
+    Worker:
+      replicas: 2
+      restartPolicy: OnFailure
+      template:
+        metadata:
+          labels:
+            app: fsdp
+            nvidia-devtools-sidecar-injector: enabled
+        spec:
+          volumes:
+            - name: shmem
+              #emptyDir:
+              #  medium: Memory
+              hostPath:
+                path: /dev/shm
+          #nodeSelector:
+          #  node.kubernetes.io/instance-type: "p5.48xlarge"
+          containers:
+            - name: pytorch
+              image: 159553542841.dkr.ecr.us-west-2.amazonaws.com/fsdp:llama2-efa-main-02-13
+              imagePullPolicy: Always
+              resources:
+                requests:
+                  nvidia.com/gpu:
+                  vpc.amazonaws.com/efa: 4
+                limits:
+                  nvidia.com/gpu:
+                  vpc.amazonaws.com/efa: 4
+              env:
+              # for P5 FI_* should be commented out
+              #- name: LOGLEVEL
+              #  value: "DEBUG"
+              - name: FI_PROVIDER
+                value: efa
+              - name: FI_EFA_USE_DEVICE_RDMA
+                value: "1"
+              - name: FI_EFA_FORK_SAFE
+                value: "1"
+              - name: FI_LOG_LEVEL
+                value: "1"
+              - name: FI_EFA_ENABLE_SHM_TRANSFER
+                value: "1"
+             #- name: NCCL_DEBUG
+             #   value: "INFO"
+              - name: NCCL_ASYNC_ERROR_HANDLING
+                value: "1"
+              #- name: NCCL_IGNORE_DISABLED_P2P
+              #  value: "1"
+              - name: HF_TOKEN
+                value: <HF_token>
+              command:
+                - bash
+                - -c
+                - "torchrun --nproc_per_node=8 --nnodes=2 examples/finetuning.py --num_epochs=1 --batch_size_training=3 --enable_fsdp --pure_bf16 --model_name meta-llama/Llama-2-7b-hf --output_dir ."
+              volumeMounts:
+                - name: shmem
+                  mountPath: /dev/shm
diff --git a/4.validation_and_observability/5.nsight/EKS/fsdp_eks_report_screenshot.png b/4.validation_and_observability/5.nsight/EKS/fsdp_eks_report_screenshot.png
diff --git a/4.validation_and_observability/5.nsight/EKS/install-injector b/4.validation_and_observability/5.nsight/EKS/install-injector
@@ -0,0 +1,4 @@
+#!/bin/bash -x
+
+helm install -f custom_values.yaml \
+    devtools-sidecar-injector https://helm.ngc.nvidia.com/nvidia/devtools/charts/devtools-sidecar-injector-1.0.0.tgz
diff --git a/4.validation_and_observability/5.nsight/EKS/label-namespace b/4.validation_and_observability/5.nsight/EKS/label-namespace
@@ -0,0 +1,2 @@
+#!/bin/bash -x
+kubectl label namespaces ${example-ns} nvidia-devtools-sidecar-injector=enabled --overwrite=true
diff --git a/4.validation_and_observability/5.nsight/EKS/move_report b/4.validation_and_observability/5.nsight/EKS/move_report
@@ -0,0 +1,10 @@
+#!/bin/bash -x
+
+# kubectl cp -n <namespace> <pod-name>:<path> <destination-on-local-system>
+
+FILE=auto_python3_default_fsdp-worker-1_pytorch_1715996702335_5a061871.nsys-rep
+
+kubectl cp fsx-share-test:fsx_shared/fsdp/$FILE /eks/deployment/distributed-training/pytorch/pytorchjob/fsdp/$FILE
+
+aws s3 cp $FILE s3://${S3_BUCKET}
+
diff --git a/4.validation_and_observability/5.nsight/EKS/uniinstall-injector b/4.validation_and_observability/5.nsight/EKS/uniinstall-injector
@@ -0,0 +1,17 @@
+#!/bin/bash -x
+
+
+helm uninstall devtools-sidecar-injector
+
+kubectl delete namespace nvidia-devtools-sidecar-injector
+
+kubectl delete mutatingwebhookconfigurations sidecar-injector-webhook
+kubectl delete mutatingwebhookconfiguration nvidia-devtools-sidecar-injector-webhook
+
+kubectl delete cm -n example-ns nvidia-devtools-sidecar-injector
+kubectl delete cm -n example-ns nvidia-devtools-sidecar-injector-custom
+
+kubectl delete cm nvidia-devtools-sidecar-injector
+kubectl delete cm nvidia-devtools-sidecar-injector-custom
+
+#kubectl get all --all-namespaces -l nvidia-devtools-sidecar-injector=enabled -o custom-columns=:.metadata.name,NS:.metadata.namespace,KIND:.kind --no-headers | while read name namespace >