Get nvidia fabric manager and dcgm using url

aws · Jun 17, 2024 · d782cc1 · d782cc1
1 parent 6429473
commit d782cc1
Show file tree

Hide file tree

Showing 4 changed files with 37 additions and 5 deletions.
diff --git a/...rallelcluster-platform/resources/fabric_manager/partial/_fabric_manager_install_debian.rb b/...rallelcluster-platform/resources/fabric_manager/partial/_fabric_manager_install_debian.rb
@@ -14,11 +14,19 @@
 
 action :install_package do
   # For ubuntu, CINC17 apt-package resources need full versions for `version`
+  remote_file "#{node['cluster']['sources_dir']}/#{fabric_manager_package}-#{fabric_manager_version}.deb" do
+    source "#{fabric_manager_url}"
+    mode '0644'
+    retries 3
+    retry_delay 5
+    action :create_if_missing
+  end
+
   bash "install_fabricmanager_for_ubuntu" do
     user 'root'
+    cwd node['cluster']['sources_dir']
     code <<-FABRIC_MANAGER
     set -e
-    aws s3 cp #{fabric_manager_url} #{fabric_manager_package}-#{fabric_manager_version}.deb
     dpkg -i #{fabric_manager_package}-#{fabric_manager_version}.deb && apt-mark hold #{fabric_manager_package}
     FABRIC_MANAGER
     retries 3

diff --git a/...parallelcluster-platform/resources/fabric_manager/partial/_fabric_manager_install_rhel.rb b/...parallelcluster-platform/resources/fabric_manager/partial/_fabric_manager_install_rhel.rb
@@ -13,12 +13,20 @@
 # See the License for the specific language governing permissions and limitations under the License.
 
 action :install_package do
+  remote_file "#{node['cluster']['sources_dir']}/#{fabric_manager_package}-#{fabric_manager_version}.rpm" do
+    source "#{fabric_manager_url}"
+    mode '0644'
+    retries 3
+    retry_delay 5
+    action :create_if_missing
+  end
+
   package 'yum-plugin-versionlock'
   bash "Install #{fabric_manager_package}" do
     user 'root'
+    cwd node['cluster']['sources_dir']
     code <<-FABRIC_MANAGER_INSTALL
     set -e
-    aws s3 cp #{fabric_manager_url} #{fabric_manager_package}-#{fabric_manager_version}.rpm --region #{node['cluster']['region']}
     yum install -y #{fabric_manager_package}-#{fabric_manager_version}.rpm
     yum versionlock #{fabric_manager_package}
     FABRIC_MANAGER_INSTALL
@@ -32,5 +40,5 @@ def arch_suffix
 end
 
 def fabric_manager_url
-  "#{node['cluster']['artifacts_build_url']}/nvidia_fabric/#{platform}/#{fabric_manager_package}-#{fabric_manager_version}-1.#{arch_suffix}.rpm"
+  "#{node['cluster']['artifacts_s3_url']}/nvidia_fabric/#{platform}/#{fabric_manager_package}-#{fabric_manager_version}-1.#{arch_suffix}.rpm"
 end
diff --git a/cookbooks/aws-parallelcluster-platform/resources/nvidia_dcgm/partial/_nvidia_dcgm_debian.rb b/cookbooks/aws-parallelcluster-platform/resources/nvidia_dcgm/partial/_nvidia_dcgm_debian.rb
@@ -13,11 +13,19 @@
 # See the License for the specific language governing permissions and limitations under the License.
 
 action :install_package do
+  remote_file "#{node['cluster']['sources_dir']}/#{dcgm_package}-#{package_version}.deb" do
+    source "#{dcgm_url}"
+    mode '0644'
+    retries 3
+    retry_delay 5
+    action :create_if_missing
+  end
+
   bash "Install #{dcgm_package}" do
     user 'root'
+    cwd node['cluster']['sources_dir']
     code <<-DCGM_INSTALL
     set -e
-    aws s3 cp #{dcgm_url} #{dcgm_package}-#{package_version}.deb --region #{node['cluster']['region']}
     dpkg -i #{dcgm_package}-#{package_version}.deb
     DCGM_INSTALL
     retries 3

diff --git a/cookbooks/aws-parallelcluster-platform/resources/nvidia_dcgm/partial/_nvidia_dcgm_rhel.rb b/cookbooks/aws-parallelcluster-platform/resources/nvidia_dcgm/partial/_nvidia_dcgm_rhel.rb
@@ -13,11 +13,19 @@
 # See the License for the specific language governing permissions and limitations under the License.
 
 action :install_package do
+  remote_file "#{node['cluster']['sources_dir']}/#{dcgm_package}-#{package_version}.rpm" do
+    source "#{dcgm_url}"
+    mode '0644'
+    retries 3
+    retry_delay 5
+    action :create_if_missing
+  end
+
   bash "Install #{dcgm_package}" do
     user 'root'
+    cwd node['cluster']['sources_dir']
     code <<-DCGM_INSTALL
     set -e
-    aws s3 cp #{dcgm_url} #{dcgm_package}-#{package_version}.rpm --region #{node['cluster']['region']}
     yum install -y #{dcgm_package}-#{package_version}.rpm
     DCGM_INSTALL
     retries 3