diff --git a/scripts/start_server.sh b/scripts/start_server.sh index 129031c7..946e08b9 100644 --- a/scripts/start_server.sh +++ b/scripts/start_server.sh @@ -2,10 +2,17 @@ set -e -o pipefail +# Check for force update flag +FORCE_UPDATE=false +if [ -f "/.force_update" ]; then + echo "Force update flag detected at /.force_update" + FORCE_UPDATE=true +fi + WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}" SERVER_DIR="$WORKSPACE_DIR/vast-pyworker" -ENV_PATH="$WORKSPACE_DIR/worker-env" +ENV_PATH="${ENV_PATH:-$WORKSPACE_DIR/worker-env}" DEBUG_LOG="$WORKSPACE_DIR/debug.log" PYWORKER_LOG="$WORKSPACE_DIR/pyworker.log" @@ -47,29 +54,38 @@ JSON } function install_vastai_sdk() { - # If SDK_BRANCH is set, install vastai-sdk from the vast-sdk repo at that branch/tag/commit. + local uv_flags=() + if [ "${USE_SYSTEM_PYTHON:-}" = "true" ]; then + uv_flags+=(--system --break-system-packages) + fi + if [ "$FORCE_UPDATE" = true ]; then + uv_flags+=(--force-reinstall) + echo "Force reinstalling vastai" + fi + + # If SDK_BRANCH is set, install vastai from the vast-cli repo at that branch/tag/commit. if [ -n "${SDK_BRANCH:-}" ]; then if [ -n "${SDK_VERSION:-}" ]; then echo "WARNING: Both SDK_BRANCH and SDK_VERSION are set; using SDK_BRANCH=${SDK_BRANCH}" fi - echo "Installing vastai-sdk from https://github.com/vast-ai/vast-sdk/ @ ${SDK_BRANCH}" - if ! uv pip install "vastai-sdk @ git+https://github.com/vast-ai/vast-sdk.git@${SDK_BRANCH}"; then - report_error_and_exit "Failed to install vastai-sdk from vast-ai/vast-sdk@${SDK_BRANCH}" + echo "Installing vastai from https://github.com/vast-ai/vast-cli/ @ ${SDK_BRANCH}" + if ! uv pip install "${uv_flags[@]}" "vastai @ git+https://github.com/vast-ai/vast-cli.git@${SDK_BRANCH}"; then + report_error_and_exit "Failed to install vastai from vast-ai/vast-cli@${SDK_BRANCH}" fi return 0 fi if [ -n "${SDK_VERSION:-}" ]; then - echo "Installing vastai-sdk version ${SDK_VERSION}" - if ! uv pip install "vastai-sdk==${SDK_VERSION}"; then - report_error_and_exit "Failed to install vastai-sdk==${SDK_VERSION}" + echo "Installing vastai version ${SDK_VERSION}" + if ! uv pip install "${uv_flags[@]}" "vastai==${SDK_VERSION}"; then + report_error_and_exit "Failed to install vastai==${SDK_VERSION}" fi return 0 fi - echo "Installing default vastai-sdk" - if ! uv pip install vastai-sdk; then - report_error_and_exit "Failed to install vastai-sdk" + echo "Installing default vastai" + if ! uv pip install "${uv_flags[@]}" vastai; then + report_error_and_exit "Failed to install vastai" fi } @@ -90,7 +106,8 @@ echo_var DEBUG_LOG echo_var PYWORKER_LOG echo_var MODEL_LOG -if [ -e "$MODEL_LOG" ]; then +ROTATE_MODEL_LOG="${ROTATE_MODEL_LOG:-false}" +if [ "$ROTATE_MODEL_LOG" = "true" ] && [ -e "$MODEL_LOG" ]; then echo "Rotating model log at $MODEL_LOG to $MODEL_LOG.old" if ! cat "$MODEL_LOG" >> "$MODEL_LOG.old"; then report_error_and_exit "Failed to rotate model log" @@ -111,8 +128,21 @@ if ! grep -q "VAST" /etc/environment; then fi fi -if [ ! -d "$ENV_PATH" ] -then +if [ "${USE_SYSTEM_PYTHON:-}" = "true" ]; then + echo "Using system Python: $(which python3)" + if ! which uv > /dev/null 2>&1; then + if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then + report_error_and_exit "Failed to install uv package manager" + fi + if [[ -f ~/.local/bin/env ]]; then + if ! source ~/.local/bin/env; then + report_error_and_exit "Failed to source uv environment" + fi + fi + fi + install_vastai_sdk + touch ~/.no_auto_tmux +elif [ ! -d "$ENV_PATH" ]; then echo "setting up venv" if ! which uv; then if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then @@ -131,10 +161,27 @@ then if ! git clone "${PYWORKER_REPO:-https://github.com/vast-ai/pyworker}" "$SERVER_DIR"; then report_error_and_exit "Failed to clone pyworker repository" fi + elif [ "$FORCE_UPDATE" = true ]; then + echo "Force updating pyworker repository" + if ! (cd "$SERVER_DIR" && git fetch --all); then + report_error_and_exit "Failed to fetch pyworker repository updates" + fi fi if [[ -n ${PYWORKER_REF:-} ]]; then - if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF"); then - report_error_and_exit "Failed to checkout pyworker reference: $PYWORKER_REF" + if [ "$FORCE_UPDATE" = true ]; then + echo "Force updating to pyworker reference: $PYWORKER_REF" + if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF" && git pull); then + report_error_and_exit "Failed to force update pyworker reference: $PYWORKER_REF" + fi + else + if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF"); then + report_error_and_exit "Failed to checkout pyworker reference: $PYWORKER_REF" + fi + fi + elif [ "$FORCE_UPDATE" = true ]; then + echo "Force updating pyworker to latest" + if ! (cd "$SERVER_DIR" && git pull); then + report_error_and_exit "Failed to pull latest pyworker changes" fi fi @@ -161,11 +208,44 @@ else report_error_and_exit "Failed to source uv environment" fi fi - if ! source "$WORKSPACE_DIR/worker-env/bin/activate"; then + if ! source "$ENV_PATH/bin/activate"; then report_error_and_exit "Failed to activate existing virtual environment" fi echo "environment activated" echo "venv: $VIRTUAL_ENV" + + # Handle force update for existing environment + if [ "$FORCE_UPDATE" = true ]; then + echo "Performing force update on existing environment" + + if [[ -d $SERVER_DIR ]]; then + echo "Force updating pyworker repository" + if ! (cd "$SERVER_DIR" && git fetch --all); then + report_error_and_exit "Failed to fetch pyworker repository updates" + fi + + if [[ -n ${PYWORKER_REF:-} ]]; then + echo "Force updating to pyworker reference: $PYWORKER_REF" + if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF" && git pull); then + report_error_and_exit "Failed to force update pyworker reference: $PYWORKER_REF" + fi + else + echo "Force updating pyworker to latest" + if ! (cd "$SERVER_DIR" && git pull); then + report_error_and_exit "Failed to pull latest pyworker changes" + fi + fi + fi + + install_vastai_sdk + fi +fi + +# Remove force update flag after successful update +if [ "$FORCE_UPDATE" = true ]; then + echo "Removing force update flag" + rm -f "/.force_update" + echo "Force update completed successfully" fi if [ "$USE_SSL" = true ]; then @@ -203,16 +283,51 @@ EOF report_error_and_exit "Failed to generate SSL certificate request" fi - if ! curl --header 'Content-Type: application/octet-stream' \ - --data-binary @/etc/instance.csr \ - -X \ - POST "https://console.vast.ai/api/v0/sign_cert/?instance_id=$CONTAINER_ID" > /etc/instance.crt; then - report_error_and_exit "Failed to sign SSL certificate" - fi + max_retries=5 + retry_delay=2 + for attempt in $(seq 1 "$max_retries"); do + http_code=$(curl -sS -o /etc/instance.crt -w '%{http_code}' \ + --header 'Content-Type: application/octet-stream' \ + --data-binary @/etc/instance.csr \ + -X POST "https://console.vast.ai/api/v0/sign_cert/?instance_id=$CONTAINER_ID") + if [ "$http_code" -ge 200 ] && [ "$http_code" -lt 300 ]; then + break + fi + echo "SSL cert signing attempt $attempt/$max_retries failed (HTTP $http_code)" + if [ "$attempt" -eq "$max_retries" ]; then + report_error_and_exit "Failed to sign SSL certificate after $max_retries attempts (HTTP $http_code)" + fi + sleep "$retry_delay" + retry_delay=$((retry_delay * 2)) + done fi export REPORT_ADDR WORKER_PORT USE_SSL UNSECURED +# ─── SDK Deployment Mode ─────────────────────────────────────────────── +if [ "$IS_DEPLOYMENT" = "true" ]; then + echo "=== SDK Deployment Mode ===" + echo "DEPLOYMENT_ID: $DEPLOYMENT_ID" + + DEPLOY_DIR="/workspace/deployment" + mkdir -p "$DEPLOY_DIR" + + VAST_API_BASE="${VAST_API_BASE:-https://console.vast.ai}" + + # Download deployment code, retrying until the blob is available on S3. + # The s3_key exists in the DB as soon as the deployment is created, but the + # actual upload may still be in flight from the client side. + + # Install SDK (uses the install_vastai_sdk function which supports SDK_BRANCH/SDK_VERSION) + install_vastai_sdk + # Run deployment in serve mode + export VAST_DEPLOYMENT_MODE=serve + echo "Starting deployment: python3 $DEPLOY_DIR/deployment.py" + serve-vast-deployment + exit $? +fi +# ─── End SDK Deployment Mode ─────────────────────────────────────────── + if ! cd "$SERVER_DIR"; then report_error_and_exit "Failed to cd into SERVER_DIR: $SERVER_DIR" fi @@ -224,19 +339,19 @@ set +e PY_STATUS=1 if [ -f "$SERVER_DIR/worker.py" ]; then - echo "trying worker.py" + echo "Running worker.py" python3 -m "worker" |& tee -a "$PYWORKER_LOG" PY_STATUS=${PIPESTATUS[0]} fi if [ "${PY_STATUS}" -ne 0 ] && [ -f "$SERVER_DIR/workers/$BACKEND/worker.py" ]; then - echo "trying workers.${BACKEND}.worker" + echo "Running workers.${BACKEND}.worker" python3 -m "workers.${BACKEND}.worker" |& tee -a "$PYWORKER_LOG" PY_STATUS=${PIPESTATUS[0]} fi if [ "${PY_STATUS}" -ne 0 ] && [ -f "$SERVER_DIR/workers/$BACKEND/server.py" ]; then - echo "trying workers.${BACKEND}.server" + echo "Running workers.${BACKEND}.server" python3 -m "workers.${BACKEND}.server" |& tee -a "$PYWORKER_LOG" PY_STATUS=${PIPESTATUS[0]} fi @@ -250,4 +365,4 @@ if [ "${PY_STATUS}" -ne 0 ]; then report_error_and_exit "PyWorker exited with status ${PY_STATUS}" fi -echo "launching PyWorker server done" +echo "PyWorker bootstrap complete" \ No newline at end of file diff --git a/vastai/serverless/server/lib/metrics.py b/vastai/serverless/server/lib/metrics.py index 5feee0ea..71b77074 100644 --- a/vastai/serverless/server/lib/metrics.py +++ b/vastai/serverless/server/lib/metrics.py @@ -157,6 +157,7 @@ def _model_loaded(self, max_throughput: float) -> None: ) self.system_metrics.model_is_loaded = True self.model_metrics.max_throughput = max_throughput + self.update_pending = True def _model_errored(self, error_msg: str) -> None: self.model_metrics.set_errored(error_msg)