skypilot-nightly 1.0.0.dev20250905__py3-none-any.whl → 1.0.0.dev20251210__py3-none-any.whl

sky/templates/kubernetes-ray.yml.j2

@@ -33,14 +33,11 @@ provider:
     networking_mode: {{k8s_networking_mode}}
 
     # We use internal IPs since we set up a port-forward between the kubernetes
-    # cluster and the local machine, or directly use NodePort to reach the
-    # head node.
+    # cluster and the local machine.
     use_internal_ips: true
 
     timeout: {{timeout}}
 
-    ssh_jump_image: {{k8s_ssh_jump_image}}
-
     # Namespace used to host SkyPilot system components, such as fuse device
     # manager.
     skypilot_system_namespace: {{k8s_skypilot_system_namespace}}
@@ -49,6 +46,10 @@ provider:
     # Used to set up the necessary permissions and sidecars.
     fuse_device_required: {{k8s_fuse_device_required}}
 
+    {% if ephemeral_volume_mounts %}
+    ephemeral_volume_specs: {{ephemeral_volume_mounts | tojson}}
+    {% endif %}
+
     # ServiceAccount created by the autoscaler for the head node pod that it
     # runs in. If this field isn't provided, the head pod config below must
     # contain a user-created service account with the proper permissions.
@@ -212,7 +213,9 @@ provider:
         metadata:
           labels:
             parent: skypilot
+            # TODO (kyuds): remove this label for v0.12.0, as skypilot-cluster label is deprecated in favor of skypilot-cluster-name.
             skypilot-cluster: {{cluster_name_on_cloud}}
+            skypilot-cluster-name: {{cluster_name_on_cloud}}
             skypilot-user: {{ user }}
           name: {{cluster_name_on_cloud}}-head-ssh
         spec:
@@ -230,7 +233,9 @@ provider:
         metadata:
             labels:
               parent: skypilot
+              # TODO (kyuds): remove this label for v0.12.0, as skypilot-cluster label is deprecated in favor of skypilot-cluster-name.
               skypilot-cluster: {{cluster_name_on_cloud}}
+              skypilot-cluster-name: {{cluster_name_on_cloud}}
               skypilot-user: {{ user }}
             # NOTE: If you're running multiple Ray clusters with services
             # on one Kubernetes cluster, they must have unique service
@@ -250,7 +255,9 @@ provider:
         metadata:
           labels:
             parent: skypilot
+            # TODO (kyuds): remove this label for v0.12.0, as skypilot-cluster label is deprecated in favor of skypilot-cluster-name.
             skypilot-cluster: {{cluster_name_on_cloud}}
+            skypilot-cluster-name: {{cluster_name_on_cloud}}
             skypilot-user: {{ user }}
           name: {{cluster_name_on_cloud}}-worker{{ worker_id }}
         spec:
@@ -275,9 +282,8 @@ available_node_types:
         labels:
             parent: skypilot
             # component will be set for the head node pod to be the same as the head node service selector above if a
+            # TODO (kyuds): remove this label for v0.12.0, as skypilot-cluster label is deprecated in favor of skypilot-cluster-name.
             skypilot-cluster: {{cluster_name_on_cloud}}
-            # Identifies the SSH jump pod used by this pod. Used in life cycle management of the ssh jump pod.
-            skypilot-ssh-jump: {{k8s_ssh_jump_name}}
             skypilot-user: {{ user }}
             # Custom tags for the pods
             {%- for label_key, label_value in labels.items() %}
@@ -444,9 +450,6 @@ available_node_types:
         # object store. If you do not provide this, Ray will fall back to
         # /tmp which cause slowdowns if is not a shared memory volume.
         volumes:
-        - name: secret-volume
-          secret:
-            secretName: {{k8s_ssh_key_secret_name}}
         - name: dshm
           emptyDir:
             medium: Memory
@@ -510,6 +513,24 @@ available_node_types:
               valueFrom:
                 fieldRef:
                   fieldPath: metadata.labels['ray-node-type']
+            - name: SKYPILOT_POD_CPU_CORE_LIMIT
+              valueFrom:
+                resourceFieldRef:
+                  containerName: ray-node
+                  resource: requests.cpu
+            - name: SKYPILOT_POD_MEMORY_BYTES_LIMIT
+              valueFrom:
+                resourceFieldRef:
+                  containerName: ray-node
+                  resource: requests.memory
+            # Disable Ray memory monitor to prevent Ray's memory manager 
+            # from interfering with kubernetes resource manager.
+            # If ray memory monitor is enabled, the ray memory monitor kills
+            # the running job is the job uses more than 95% of allocated memory,
+            # even if the job is not misbehaving or using its full allocated memory.
+            # This behavior does not give a chance for k8s scheduler to evict the pod.
+            - name: RAY_memory_monitor_refresh_ms
+              value: "0"
             {% for key, value in k8s_env_vars.items() if k8s_env_vars is not none %}
             - name: {{ key }}
               value: {{ value }}
@@ -630,12 +651,17 @@ available_node_types:
           command: ["/bin/bash", "-c", "--"]
           args:
             - |
-              # For backwards compatibility, we put a marker file in the pod
-              # to indicate that the pod is running with the changes introduced
-              # in project nimbus: https://github.com/skypilot-org/skypilot/pull/4393
-              # TODO: Remove this marker file and it's usage in setup_commands
-              # after v0.10.0 release.
-              touch /tmp/skypilot_is_nimbus
+              # Set -x to print the commands and their arguments as they are executed.
+              # Useful for debugging.
+              set -x
+
+              # Execute user-provided post-provision runcmd
+              # before any of the SkyPilot setup commands.
+              {%- if runcmd %}
+              {%- for cmd in runcmd %}
+              {{cmd}}
+              {%- endfor %}
+              {%- endif %}
 
               # Helper function to conditionally use sudo
               # TODO(zhwu): consolidate the two prefix_cmd and sudo replacements
@@ -647,15 +673,125 @@ available_node_types:
               # STEP 1: Run apt update, install missing packages, and set up ssh.
               (
                 (
-                # For backwards compatibility, we put a marker file in the pod
-                # to indicate that the apt ssh setup step will write a completion
-                # marker file (/tmp/apt_ssh_setup_complete) to the pod.
-                # TODO: Remove this marker file and its usage in setup_commands
-                # after v0.11.0 release.
-                touch /tmp/apt_ssh_setup_started
-
-                DEBIAN_FRONTEND=noninteractive $(prefix_cmd) apt-get update > /tmp/apt-update.log 2>&1 || \
-                echo "Warning: apt-get update failed. Continuing anyway..." >> /tmp/apt-update.log
+                # Helper: run apt-get update with retries
+                apt_update_with_retries() {
+                  # do not fail the whole shell; we handle return codes
+                  set +e
+                  local log=/tmp/apt-update.log
+                  local tries=3
+                  local delay=1
+                  local i
+                  for i in $(seq 1 $tries); do
+                    DEBIAN_FRONTEND=noninteractive $(prefix_cmd) apt-get update >> "$log" 2>&1 && { set -e; return 0; }
+                    echo "apt-get update attempt $i/$tries failed; retrying in ${delay}s" >> "$log"
+                    sleep $delay
+                    delay=$((delay * 2))
+                  done
+                  set -e
+                  return 1
+                }
+                apt_install_with_retries() {
+                  local packages="$@"
+                  [ -z "$packages" ] && return 0
+                  set +e
+                  local log=/tmp/apt-update.log
+                  local tries=3
+                  local delay=1
+                  local i
+                  for i in $(seq 1 $tries); do
+                    DEBIAN_FRONTEND=noninteractive $(prefix_cmd) apt-get install -y -o Dpkg::Options::="--force-confdef" -o Dpkg::Options::="--force-confold" $packages && { set -e; return 0; }
+                    echo "apt-get install failed for: $packages (attempt $i/$tries). Running -f install and retrying..." >> "$log"
+                    DEBIAN_FRONTEND=noninteractive $(prefix_cmd) apt-get -f install -y >> "$log" 2>&1 || true
+                    DEBIAN_FRONTEND=noninteractive $(prefix_cmd) apt-get clean >> "$log" 2>&1 || true
+                    sleep $delay
+                    delay=$((delay * 2))
+                  done
+                  set -e
+                  return 1
+                }
+                apt_update_install_with_retries() {
+                  apt_update_with_retries
+                  apt_install_with_retries "$@"
+                }
+                backup_dir=/etc/apt/sources.list.backup_skypilot
+                backup_source() {
+                  $(prefix_cmd) mkdir -p "$backup_dir"
+                  if [ -f /etc/apt/sources.list ] && [ ! -f "$backup_dir/sources.list" ]; then
+                    $(prefix_cmd) cp -a /etc/apt/sources.list "$backup_dir/sources.list" || true
+                  fi
+                }
+                restore_source() {
+                  if [ -f "$backup_dir/sources.list" ]; then
+                    $(prefix_cmd) cp -a "$backup_dir/sources.list" /etc/apt/sources.list || true
+                  fi
+                }
+                update_apt_sources() {
+                  local host=$1
+                  local apt_file=$2
+                  $(prefix_cmd) sed -i -E "s|https?://[a-zA-Z0-9.-]+\.ubuntu\.com/ubuntu|http://$host/ubuntu|g" $apt_file
+                }
+                # Helper: install packages across mirrors with retries
+                apt_install_with_mirrors() {
+                  local required=$1; shift
+                  local packages="$@"
+                  [ -z "$packages" ] && return 0
+                  set +e
+                  # Install packages with default sources first
+                  local log=/tmp/apt-update.log
+                  echo "$(date +%Y-%m-%d\ %H:%M:%S) Installing packages: $packages" >> "$log"
+                  restore_source
+                  apt_update_install_with_retries $packages >> "$log" 2>&1 && { set -e; return 0; }
+                  echo "Install failed with default sources: $packages" >> "$log"
+                  # Detect distro (ubuntu/debian)
+                  local APT_OS="unknown"
+                  if [ -f /etc/os-release ]; then
+                    . /etc/os-release
+                    case "$ID" in
+                      debian) APT_OS="debian" ;;
+                      ubuntu) APT_OS="ubuntu" ;;
+                      *)
+                        if [ -n "$ID_LIKE" ]; then
+                          case " $ID $ID_LIKE " in
+                            *ubuntu*) APT_OS="ubuntu" ;;
+                            *debian*) APT_OS="debian" ;;
+                          esac
+                        fi
+                        ;;
+                    esac
+                  fi
+                  # Build mirror candidates
+                  # deb.debian.org is a CDN endpoint, if one backend goes down,
+                  # the CDN automatically fails over to another mirror,
+                  # so we only retry for ubuntu here.
+                  if [ "$APT_OS" = "ubuntu" ]; then
+                    # Backup current sources once
+                    backup_source
+                    # Selected from https://launchpad.net/ubuntu/+archivemirrors
+                    # and results from apt-select
+                    local MIRROR_CANDIDATES="mirrors.wikimedia.org mirror.umd.edu"
+                    for host in $MIRROR_CANDIDATES; do
+                      echo "Trying APT mirror ($APT_OS): $host" >> "$log"
+                      if [ -f /etc/apt/sources.list ]; then
+                        update_apt_sources $host /etc/apt/sources.list
+                      else
+                        echo "Error: /etc/apt/sources.list not found" >> "$log"
+                        break
+                      fi
+                      apt_update_install_with_retries $packages >> "$log" 2>&1 && { set -e; return 0; }
+                      echo "Install failed with mirror ($APT_OS): $host" >> "$log"
+                      # Restore to default sources
+                      restore_source
+                    done
+                  fi
+                  set -e
+                  if [ "$required" = "1" ]; then
+                    echo "Error: required package install failed across all mirrors: $packages" >> "$log"
+                    return 1
+                  else
+                    echo "Optional package install failed across all mirrors: $packages; skipping." >> "$log"
+                    return 0
+                  fi
+                }
                 # Install both fuse2 and fuse3 for compatibility for all possible fuse adapters in advance,
                 # so that both fusemount and fusermount3 can be masked before enabling SSH access.
                 PACKAGES="rsync curl wget netcat gcc patch pciutils fuse fuse3 openssh-server";
@@ -682,7 +818,7 @@ available_node_types:
                 done;
                 if [ ! -z "$INSTALL_FIRST" ]; then
                   echo "Installing core packages: $INSTALL_FIRST";
-                  DEBIAN_FRONTEND=noninteractive $(prefix_cmd) apt-get install -y -o Dpkg::Options::="--force-confdef" -o Dpkg::Options::="--force-confold" $INSTALL_FIRST;
+                  apt_install_with_mirrors 1 $INSTALL_FIRST || { echo "Error: core package installation failed." >> /tmp/apt-update.log; exit 1; }
                 fi;
                 # SSH and other packages are not necessary, so we disable set -e
                 set +e
@@ -706,7 +842,8 @@ available_node_types:
                 fi
                 $(prefix_cmd) cp -p "$FUSERMOUNT_PATH" "${FUSERMOUNT_PATH}-original"
                 $(prefix_cmd) ln -sf {{k8s_fusermount_shared_dir}}/fusermount-shim "$FUSERMOUNT_PATH"
-                FUSERMOUNT3_PATH=$(which fusermount3)
+                # "|| true" because fusermount3 is not always available
+                FUSERMOUNT3_PATH=$(which fusermount3) || true
                 if [ -z "$FUSERMOUNT3_PATH" ]; then
                     FUSERMOUNT3_PATH="${FUSERMOUNT_PATH}3"
                 fi
@@ -748,18 +885,23 @@ available_node_types:
                 $(prefix_cmd) mkdir -p ~/.ssh;
                 $(prefix_cmd) chown -R $(whoami) ~/.ssh;
                 $(prefix_cmd) chmod 700 ~/.ssh;
-                $(prefix_cmd) cat /etc/secret-volume/ssh-publickey* > ~/.ssh/authorized_keys;
+                $(prefix_cmd) cat > ~/.ssh/authorized_keys <<'SKYPILOT_SSH_KEY_EOF'
+              skypilot:ssh_public_key_content
+              SKYPILOT_SSH_KEY_EOF
                 $(prefix_cmd) chmod 644 ~/.ssh/authorized_keys;
                 $(prefix_cmd) service ssh restart;
                 $(prefix_cmd) sed -i "s/mesg n/tty -s \&\& mesg n/" ~/.profile;
 
                 touch /tmp/apt_ssh_setup_complete
                 echo "=== SSH setup completed ==="
-                ) > /tmp/${STEPS[0]}.log 2>&1 || {
-                  echo "Error: ${STEPS[0]} failed. Continuing anyway..." > /tmp/${STEPS[0]}.failed
+                ) > /tmp/${STEPS[0]}.log 2>&1
+                if [ "$?" -ne "0" ]; then
+                  {
+                  echo "Error: ${STEPS[0]} failed. Continuing anyway..." > /tmp/${STEPS[0]}.failed 2>&1
                   cat /tmp/${STEPS[0]}.log
                   exit 1
-                }
+                  }
+                fi
               ) &
 
               # STEP 2: Install conda, ray and skypilot (for dependencies); start
@@ -777,15 +919,20 @@ available_node_types:
                 {{ conda_installation_commands }}
                 {{ ray_installation_commands }}
 
-                VIRTUAL_ENV=~/skypilot-runtime ~/.local/bin/uv pip install skypilot[kubernetes,remote]
+                # set UV_SYSTEM_PYTHON to false in case the user provided docker image set it to true.
+                # unset PYTHONPATH in case the user provided docker image set it.
+                VIRTUAL_ENV=~/skypilot-runtime UV_SYSTEM_PYTHON=false env -u PYTHONPATH ~/.local/bin/uv pip install skypilot[kubernetes,remote]
                 # Wait for `patch` package to be installed before applying ray patches
                 until dpkg -l | grep -q "^ii  patch "; do
                   sleep 0.1
                   echo "Waiting for patch package to be installed..."
                 done
                 # Apply Ray patches for progress bar fix
-                ~/.local/bin/uv pip list | grep "ray " | grep 2.9.3 2>&1 > /dev/null && {
-                  VIRTUAL_ENV=~/skypilot-runtime python -c "from sky.skylet.ray_patches import patch; patch()" || exit 1;
+                # set UV_SYSTEM_PYTHON to false in case the user provided docker image set it to true.
+                # unset PYTHONPATH in case the user provided docker image set it.
+                # ~/.sky/python_path is seeded by conda_installation_commands
+                VIRTUAL_ENV=~/skypilot-runtime UV_SYSTEM_PYTHON=false env -u PYTHONPATH ~/.local/bin/uv pip list | grep "ray " | grep 2.9.3 2>&1 > /dev/null && {
+                  env -u PYTHONPATH $(cat ~/.sky/python_path) -c "from sky.skylet.ray_patches import patch; patch()" || exit 1;
                 }
                 touch /tmp/ray_skypilot_installation_complete
                 echo "=== Ray and skypilot installation completed ==="
@@ -814,11 +961,14 @@ available_node_types:
                   set +e
                   {{ ray_worker_start_command }}
                 fi
-                ) > /tmp/${STEPS[1]}.log 2>&1 || {
-                  echo "Error: ${STEPS[1]} failed. Continuing anyway..." > /tmp/${STEPS[1]}.failed
+                ) > /tmp/${STEPS[1]}.log 2>&1
+                if [ "$?" -ne "0" ]; then
+                  {
+                  echo "Error: ${STEPS[1]} failed. Continuing anyway..." > /tmp/${STEPS[1]}.failed 2>&1
                   cat /tmp/${STEPS[1]}.log
                   exit 1
-                }
+                  }
+                fi
               ) &
 
 
@@ -836,11 +986,14 @@ available_node_types:
                   fi;
                 fi;
                 export -p > ~/container_env_var.sh && $(prefix_cmd) mv ~/container_env_var.sh /etc/profile.d/container_env_var.sh
-                ) > /tmp/${STEPS[2]}.log 2>&1 || {
-                  echo "Error: ${STEPS[2]} failed. Continuing anyway..." > /tmp/${STEPS[2]}.failed
+                ) > /tmp/${STEPS[2]}.log 2>&1
+                if [ "$?" -ne "0" ]; then
+                  {
+                  echo "Error: ${STEPS[2]} failed. Continuing anyway..." > /tmp/${STEPS[2]}.failed 2>&1
                   cat /tmp/${STEPS[2]}.log
                   exit 1
-                }
+                  }
+                fi
               ) &
 
               function mylsof { p=$(for pid in /proc/{0..9}*; do i=$(basename "$pid"); for file in "$pid"/fd/*; do link=$(readlink -e "$file"); if [ "$link" = "$1" ]; then echo "$i"; fi; done; done); echo "$p"; };
@@ -927,7 +1080,7 @@ available_node_types:
                 # Also, skip the jobs that are waiting to be scheduled as those does not have a controller process running.
                 # For SkyServe, this will be None and every service will be recovered. This is because SkyServe
                 # will delete the service from the database after it is terminated so everything in the database is running.
-                ALL_IN_PROGRESS_JOBS=$({{sky_python_cmd}} -c "from sky.jobs import state; jobs = state.get_managed_jobs(); print(' '.join({str(job['job_id']) for job in jobs if job['schedule_state'] not in [state.ManagedJobScheduleState.DONE, state.ManagedJobScheduleState.WAITING]}) if jobs else None)")
+                ALL_IN_PROGRESS_JOBS=$({{sky_python_cmd}} -c "from sky.jobs import state; jobs, _ = state.get_managed_jobs_with_filters(fields=['job_id', 'schedule_state']); print(' '.join({str(job['job_id']) for job in jobs if job['schedule_state'] not in [state.ManagedJobScheduleState.DONE, state.ManagedJobScheduleState.WAITING]}) if jobs else None)")
                 if [ "$ALL_IN_PROGRESS_JOBS" != "None" ]; then
                   read -ra ALL_IN_PROGRESS_JOBS_SEQ <<< "$ALL_IN_PROGRESS_JOBS"
                 fi
@@ -957,6 +1110,8 @@ available_node_types:
 
               touch {{k8s_high_availability_deployment_volume_mount_path}}/k8s_container_ready
               {% endif %}
+              # Set +x to stop printing the commands and their arguments as they are executed.
+              set +x
 
               trap : TERM INT; log_tail || sleep infinity & wait
 
@@ -970,9 +1125,6 @@ available_node_types:
           # object store. If you do not provide this, Ray will fall back to
           # /tmp which cause slowdowns if is not a shared memory volume.
           volumeMounts:
-          - name: secret-volume
-            readOnly: true
-            mountPath: "/etc/secret-volume"
           - mountPath: /dev/shm
             name: dshm
           {% if k8s_enable_gpudirect_tcpx %}
@@ -1204,24 +1356,21 @@ setup_commands:
     start_epoch=$(date +%s);
 
     # Wait for SSH setup to complete before proceeding
-    if [ -f /tmp/apt_ssh_setup_started ]; then
-      echo "=== Logs for asynchronous SSH setup ===";
-      [ -f /tmp/apt_ssh_setup_complete ] && cat /tmp/${STEPS[0]}.log ||
-      { tail -f -n +1 /tmp/${STEPS[0]}.log & TAIL_PID=$!; echo "Tail PID: $TAIL_PID"; until [ -f /tmp/apt_ssh_setup_complete ]; do sleep 0.5; done; kill $TAIL_PID || true; };
-      [ -f /tmp/${STEPS[0]}.failed ] && { echo "Error: ${STEPS[0]} failed. Exiting."; exit 1; } || true;
-    fi
+    echo "=== Logs for asynchronous SSH setup ===";
+    ([ -f /tmp/apt_ssh_setup_complete ]|| [ -f /tmp/${STEPS[0]}.failed ]) && cat /tmp/${STEPS[0]}.log ||
+    { tail -f -n +1 /tmp/${STEPS[0]}.log & TAIL_PID=$!; echo "Tail PID: $TAIL_PID"; sleep 0.5; until [ -f /tmp/apt_ssh_setup_complete ] || [ -f /tmp/${STEPS[0]}.failed ]; do sleep 0.5; done; kill $TAIL_PID || true; };
+    [ -f /tmp/${STEPS[0]}.failed ] && { echo "Error: ${STEPS[0]} failed. Exiting."; exit 1; } || true;
 
     echo "=== Logs for asynchronous ray and skypilot installation ===";
-    if [ -f /tmp/skypilot_is_nimbus ]; then
-      echo "=== Logs for asynchronous ray and skypilot installation ===";
-      [ -f /tmp/ray_skypilot_installation_complete ] && cat /tmp/${STEPS[1]}.log ||
-      { tail -f -n +1 /tmp/${STEPS[1]}.log & TAIL_PID=$!; echo "Tail PID: $TAIL_PID"; until [ -f /tmp/ray_skypilot_installation_complete ]; do sleep 0.5; done; kill $TAIL_PID || true; };
-      [ -f /tmp/${STEPS[1]}.failed ] && { echo "Error: ${STEPS[1]} failed. Exiting."; exit 1; } || true;
-    fi
+    ([ -f /tmp/ray_skypilot_installation_complete ]|| [ -f /tmp/${STEPS[1]}.failed ]) && cat /tmp/${STEPS[1]}.log ||
+    { tail -f -n +1 /tmp/${STEPS[1]}.log & TAIL_PID=$!; echo "Tail PID: $TAIL_PID"; sleep 0.5; until [ -f /tmp/ray_skypilot_installation_complete ] || [ -f /tmp/${STEPS[1]}.failed ]; do sleep 0.5; done; kill $TAIL_PID || true; };
+    [ -f /tmp/${STEPS[1]}.failed ] && { echo "Error: ${STEPS[1]} failed. Exiting."; exit 1; } || true;
+
     end_epoch=$(date +%s);
     echo "=== Ray and skypilot dependencies installation completed in $(($end_epoch - $start_epoch)) secs ===";
     start_epoch=$(date +%s);
     {{ skypilot_wheel_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
     end_epoch=$(date +%s);
     echo "=== Skypilot wheel installation completed in $(($end_epoch - $start_epoch)) secs ===";
     start_epoch=$(date +%s);

sky/templates/lambda-ray.yml.j2

@@ -91,6 +91,7 @@ setup_commands:
     rm ~/.local/bin/pip ~/.local/bin/pip3 ~/.local/bin/pip3.8 ~/.local/bin/pip3.10;
     {{ conda_installation_commands }}
     {{ ray_skypilot_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
     touch ~/.sudo_as_admin_successful;
     sudo bash -c 'rm -rf /etc/security/limits.d; echo "* soft nofile 1048576" >> /etc/security/limits.conf; echo "* hard nofile 1048576" >> /etc/security/limits.conf';
     sudo grep -e '^DefaultTasksMax' /etc/systemd/system.conf || (sudo bash -c 'echo "DefaultTasksMax=infinity" >> /etc/systemd/system.conf'); sudo systemctl set-property user-$(id -u $(whoami)).slice TasksMax=infinity; sudo systemctl daemon-reload;

sky/templates/nebius-ray.yml.j2

@@ -10,6 +10,7 @@ provider:
   module: sky.provision.nebius
   region: "{{region}}"
   use_internal_ips: {{use_internal_ips}}
+  use_static_ip_address: {{ use_static_ip_address }}
 
 {%- if docker_image is not none %}
 docker:
@@ -150,11 +151,13 @@ setup_commands:
     mkdir -p ~/.ssh; touch ~/.ssh/config;
     {{ conda_installation_commands }}
     {{ ray_skypilot_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
   {%- if env_vars is defined %}
     {%- for env_var, env_value in env_vars.items() %}
     echo '{{env_var}}={{env_value}}' | sudo tee -a /etc/environment;
     {%- endfor %}
   {%- endif %}
+    IP=$(hostname -I | awk '{print $1}'); echo "$IP $(hostname)" | sudo tee -a /etc/hosts;
     sudo bash -c 'rm -rf /etc/security/limits.d; echo "* soft nofile 1048576" >> /etc/security/limits.conf; echo "* hard nofile 1048576" >> /etc/security/limits.conf';
     sudo grep -e '^DefaultTasksMax' /etc/systemd/system.conf || (sudo bash -c 'echo "DefaultTasksMax=infinity" >> /etc/systemd/system.conf'); sudo systemctl set-property user-$(id -u $(whoami)).slice TasksMax=infinity; sudo systemctl daemon-reload;
     mkdir -p ~/.ssh; (grep -Pzo -q "Host \*\n  StrictHostKeyChecking no\n  IdentityFile ~/.ssh/sky-cluster-key\n  IdentityFile ~/.ssh/id_rsa" ~/.ssh/config) || printf "Host *\n  StrictHostKeyChecking no\n  IdentityFile ~/.ssh/sky-cluster-key\n  IdentityFile ~/.ssh/id_rsa\n" >> ~/.ssh/config;

sky/templates/oci-ray.yml.j2

@@ -85,6 +85,7 @@ setup_commands:
     mkdir -p ~/.ssh; touch ~/.ssh/config;
     {{ conda_installation_commands }}
     {{ ray_skypilot_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
     touch ~/.sudo_as_admin_successful;
     sudo bash -c 'rm -rf /etc/security/limits.d; echo "* soft nofile 1048576" >> /etc/security/limits.conf; echo "* hard nofile 1048576" >> /etc/security/limits.conf';
     sudo grep -e '^DefaultTasksMax' /etc/systemd/system.conf || (sudo bash -c 'echo "DefaultTasksMax=infinity" >> /etc/systemd/system.conf'); sudo systemctl set-property user-$(id -u $(whoami)).slice TasksMax=infinity; sudo systemctl daemon-reload;

sky/templates/paperspace-ray.yml.j2

@@ -87,6 +87,7 @@ setup_commands:
     mkdir -p ~/.ssh; touch ~/.ssh/config;
     {{ conda_installation_commands }}
     {{ ray_skypilot_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
     sudo bash -c 'rm -rf /etc/security/limits.d; echo "* soft nofile 1048576" >> /etc/security/limits.conf; echo "* hard nofile 1048576" >> /etc/security/limits.conf';
     sudo grep -e '^DefaultTasksMax' /etc/systemd/system.conf || (sudo bash -c 'echo "DefaultTasksMax=infinity" >> /etc/systemd/system.conf'); sudo systemctl set-property user-$(id -u $(whoami)).slice TasksMax=infinity; sudo systemctl daemon-reload;
     mkdir -p ~/.ssh; (grep -Pzo -q "Host \*\n  StrictHostKeyChecking no\n  IdentityFile ~/.ssh/sky-cluster-key\n  IdentityFile ~/.ssh/id_rsa" ~/.ssh/config) || printf "Host *\n  StrictHostKeyChecking no\n  IdentityFile ~/.ssh/sky-cluster-key\n  IdentityFile ~/.ssh/id_rsa\n" >> ~/.ssh/config;

sky/templates/primeintellect-ray.yml.j2

@@ -0,0 +1,72 @@
+cluster_name: {{cluster_name_on_cloud}}
+
+# The maximum number of workers nodes to launch in addition to the head node.
+max_workers: {{num_nodes - 1}}
+upscaling_speed: {{num_nodes - 1}}
+idle_timeout_minutes: 60
+
+provider:
+  type: external
+  module: sky.provision.primeintellect
+  region: "{{region}}"
+  zones: "{{zones}}"
+
+auth:
+  ssh_user: skypilot:ssh_user
+  ssh_private_key: {{ssh_private_key}}
+
+available_node_types:
+  ray_head_default:
+    resources: {}
+    node_config:
+      InstanceType: {{instance_type}}
+      DiskSize: {{disk_size}}
+      ImageId: {{image_id}}
+      PublicKey: |-
+        skypilot:ssh_public_key_content
+
+head_node_type: ray_head_default
+
+# Format: `REMOTE_PATH : LOCAL_PATH`
+file_mounts: {
+  "{{sky_ray_yaml_remote_path}}": "{{sky_ray_yaml_local_path}}",
+  "{{sky_remote_path}}/{{sky_wheel_hash}}": "{{sky_local_path}}",
+{%- for remote_path, local_path in credentials.items() %}
+  "{{remote_path}}": "{{local_path}}",
+  "~/.ssh/sky-cluster-key": "{{ssh_private_key}}",
+{%- endfor %}
+}
+
+rsync_exclude: []
+
+initialization_commands: []
+
+# List of shell commands to run to set up nodes.
+# NOTE: these are very performance-sensitive. Each new item opens/closes an SSH
+# connection, which is expensive. Try your best to co-locate commands into fewer
+# items!
+#
+# Increment the following for catching performance bugs easier:
+#   current num items (num SSH connections): 1
+setup_commands:
+  # Disable unattended-upgrades and handle apt-get locks
+  # Install patch utility for Ray
+  # Install conda and Ray
+  # Set system limits for Ray performance (nofile and TasksMax)
+  - {%- for initial_setup_command in initial_setup_commands %}
+    {{ initial_setup_command }}
+    {%- endfor %}
+    sudo systemctl stop unattended-upgrades || true;
+    sudo systemctl disable unattended-upgrades || true;
+    sudo sed -i 's/Unattended-Upgrade "1"/Unattended-Upgrade "0"/g' /etc/apt/apt.conf.d/20auto-upgrades || true;
+    sudo kill -9 `sudo lsof /var/lib/dpkg/lock-frontend | awk '{print $2}' | tail -n 1` || true;
+    sudo pkill -9 apt-get;
+    sudo pkill -9 dpkg;
+    sudo dpkg --configure -a;
+    which patch > /dev/null || sudo apt install -y patch;
+    {{ conda_installation_commands }}
+    {{ ray_skypilot_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
+    sudo bash -c 'rm -rf /etc/security/limits.d; echo "* soft nofile 1048576" >> /etc/security/limits.conf; echo "* hard nofile 1048576" >> /etc/security/limits.conf';
+    sudo grep -e '^DefaultTasksMax' /etc/systemd/system.conf || (sudo bash -c 'echo "DefaultTasksMax=infinity" >> /etc/systemd/system.conf'); sudo systemctl set-property user-$(id -u $(whoami)).slice TasksMax=infinity; sudo systemctl daemon-reload;
+    {{ ssh_max_sessions_config }}

sky/templates/runpod-ray.yml.j2

@@ -93,6 +93,7 @@ setup_commands:
     mkdir -p ~/.ssh; touch ~/.ssh/config;
     {{ conda_installation_commands }}
     {{ ray_skypilot_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
     touch ~/.sudo_as_admin_successful;
     sudo bash -c 'rm -rf /etc/security/limits.d; echo "* soft nofile 1048576" >> /etc/security/limits.conf; echo "* hard nofile 1048576" >> /etc/security/limits.conf';
     sudo grep -e '^DefaultTasksMax' /etc/systemd/system.conf || (sudo bash -c 'echo "DefaultTasksMax=infinity" >> /etc/systemd/system.conf'); sudo systemctl set-property user-$(id -u $(whoami)).slice TasksMax=infinity; sudo systemctl daemon-reload;

sky/templates/scp-ray.yml.j2

@@ -56,6 +56,7 @@ setup_commands:
   - mkdir -p ~/.ssh; touch ~/.ssh/config;
     {{ conda_installation_commands }}
     {{ ray_skypilot_installation_commands }}
+    {{ copy_skypilot_templates_commands }}
     sudo bash -c 'rm -rf /etc/security/limits.d; echo "* soft nofile 1048576" >> /etc/security/limits.conf; echo "* hard nofile 1048576" >> /etc/security/limits.conf';
     sudo grep -e '^DefaultTasksMax' /etc/systemd/system.conf || (sudo bash -c 'echo "DefaultTasksMax=infinity" >> /etc/systemd/system.conf'); sudo systemctl set-property user-$(id -u $(whoami)).slice TasksMax=infinity; sudo systemctl daemon-reload;
     mkdir -p ~/.ssh; (grep -Pzo -q "Host \*\n  StrictHostKeyChecking no\n  IdentityFile ~/.ssh/sky-cluster-key\n  IdentityFile ~/.ssh/id_rsa" ~/.ssh/config) || printf "Host *\n  StrictHostKeyChecking no\n  IdentityFile ~/.ssh/sky-cluster-key\n  IdentityFile ~/.ssh/id_rsa\n" >> ~/.ssh/config;