Merge branch 'main' into add-hacky-deployments-script

requirements.txt

@@ -1,2 +1 @@
 vastai-sdk>=0.3.0
-nltk==3.9.4 

start_server.sh

@@ -2,17 +2,10 @@
 
 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="$WORKSPACE_DIR/vast-pyworker"
-ENV_PATH="${ENV_PATH:-$WORKSPACE_DIR/worker-env}"
+ENV_PATH="$WORKSPACE_DIR/worker-env"
 DEBUG_LOG="$WORKSPACE_DIR/debug.log"
 PYWORKER_LOG="$WORKSPACE_DIR/pyworker.log"
 
@@ -54,38 +47,29 @@ JSON
 }
 
 function install_vastai_sdk() {
-    local uv_flags=()
+    # If SDK_BRANCH is set, install vastai-sdk from the vast-sdk repo at that branch/tag/commit.
-    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 from https://github.com/vast-ai/vast-cli/ @ ${SDK_BRANCH}"
+        echo "Installing vastai-sdk from https://github.com/vast-ai/vast-sdk/ @ ${SDK_BRANCH}"
-        if ! uv pip install "${uv_flags[@]}" "vastai @ git+https://github.com/vast-ai/vast-cli.git@${SDK_BRANCH}"; then
+        if ! uv pip install "vastai-sdk @ git+https://github.com/vast-ai/vast-sdk.git@${SDK_BRANCH}"; then
-            report_error_and_exit "Failed to install vastai from vast-ai/vast-cli@${SDK_BRANCH}"
+            report_error_and_exit "Failed to install vastai-sdk from vast-ai/vast-sdk@${SDK_BRANCH}"
         fi
         return 0
     fi
 
     if [ -n "${SDK_VERSION:-}" ]; then
-        echo "Installing vastai version ${SDK_VERSION}"
+        echo "Installing vastai-sdk version ${SDK_VERSION}"
-        if ! uv pip install "${uv_flags[@]}" "vastai==${SDK_VERSION}"; then
+        if ! uv pip install "vastai-sdk==${SDK_VERSION}"; then
-            report_error_and_exit "Failed to install vastai==${SDK_VERSION}"
+            report_error_and_exit "Failed to install vastai-sdk==${SDK_VERSION}"
         fi
         return 0
     fi
 
-    echo "Installing default vastai"
+    echo "Installing default vastai-sdk"
-    if ! uv pip install "${uv_flags[@]}" vastai; then
+    if ! uv pip install vastai-sdk; then
-        report_error_and_exit "Failed to install vastai"
+        report_error_and_exit "Failed to install vastai-sdk"
     fi
 }
 
@@ -128,21 +112,8 @@ if ! grep -q "VAST" /etc/environment; then
     fi
 fi
 
-if [ "${USE_SYSTEM_PYTHON:-}" = "true" ]; then
+if [ ! -d "$ENV_PATH" ]
-    echo "Using system Python: $(which python3)"
+then
-    if ! which uv > /dev/null 2>&1; then
-        if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then
-            report_error_and_exit "Failed to install uv package manager"
-        fi
-        if [[ -f ~/.local/bin/env ]]; then
-            if ! source ~/.local/bin/env; then
-                report_error_and_exit "Failed to source uv environment"
-            fi
-        fi
-    fi
-    install_vastai_sdk
-    touch ~/.no_auto_tmux
-elif [ ! -d "$ENV_PATH" ]; then
     echo "setting up venv"
     if ! which uv; then
         if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then
@@ -161,29 +132,12 @@ elif [ ! -d "$ENV_PATH" ]; then
         if ! git clone "${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 ${PYWORKER_REF:-} ]]; then
-        if [ "$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
 
     if ! uv venv --python-preference only-managed "$ENV_PATH" -p 3.10; then
         report_error_and_exit "Failed to create virtual environment"
@@ -208,44 +162,11 @@ else
             report_error_and_exit "Failed to source uv environment"
         fi
     fi
-    if ! source "$ENV_PATH/bin/activate"; then
+    if ! source "$WORKSPACE_DIR/worker-env/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
@@ -283,23 +204,12 @@ EOF
         report_error_and_exit "Failed to generate SSL certificate request"
     fi
 
-    max_retries=5
+    if ! curl --header 'Content-Type: application/octet-stream' \
-    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")
+        -X \
-        if [ "$http_code" -ge 200 ] && [ "$http_code" -lt 300 ]; then
+        POST "https://console.vast.ai/api/v0/sign_cert/?instance_id=$CONTAINER_ID" > /etc/instance.crt; then
-            break
+        report_error_and_exit "Failed to sign SSL certificate"
     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
@@ -317,13 +227,63 @@ if [ "$IS_DEPLOYMENT" = "true" ]; then
     # Download deployment code, retrying until the blob is available on S3.
     # The s3_key exists in the DB as soon as the deployment is created, but the
     # actual upload may still be in flight from the client side.
+    echo "Downloading deployment code..."
+    RETRY=0
+    while true; do
+        DOWNLOAD_RESPONSE=$(curl -sS \
+            -H "Authorization: Bearer $CONTAINER_API_KEY" \
+            "${VAST_API_BASE}/api/v0/deployment/${DEPLOYMENT_ID}/download_url/")
+        DOWNLOAD_URL=$(python3 -c "
+import sys, json
+try:
+    d = json.load(sys.stdin)
+    print(d.get('download_url') or '')
+except: print('')
+" <<< "$DOWNLOAD_RESPONSE")
+
+        if [ -z "$DOWNLOAD_URL" ] || [ "$DOWNLOAD_URL" = "None" ]; then
+            RETRY=$((RETRY + 1))
+            echo "No download URL yet (attempt $RETRY), retrying in 10s... response: $DOWNLOAD_RESPONSE"
+            sleep 10
+            continue
+        fi
+
+        # Got a URL — try the actual S3 download
+        HTTP_CODE=$(curl -sS -L -o "$DEPLOY_DIR/deployment.tar.gz" -w "%{http_code}" "$DOWNLOAD_URL")
+        if [ "$HTTP_CODE" = "200" ]; then
+            break
+        fi
+
+        RETRY=$((RETRY + 1))
+        echo "S3 download returned HTTP $HTTP_CODE (attempt $RETRY), blob not yet uploaded. Retrying in 10s..."
+        rm -f "$DEPLOY_DIR/deployment.tar.gz"
+        sleep 10
+    done
+
+    cd "$DEPLOY_DIR" && tar xzf deployment.tar.gz
+    echo "Deployment code extracted."
+
+    # Source secrets if present
+    if [ -f "$DEPLOY_DIR/.secrets" ]; then
+        echo "Sourcing secrets..."
+        source "$DEPLOY_DIR/.secrets"
+    fi
+
+    # Run on_start.sh to completion if present
+    if [ -f "$DEPLOY_DIR/on_start.sh" ]; then
+        echo "Running on_start.sh..."
+        chmod +x "$DEPLOY_DIR/on_start.sh"
+        bash "$DEPLOY_DIR/on_start.sh"
+        echo "on_start.sh completed."
+    fi
 
     # 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 $DEPLOY_DIR/deployment.py"
-    serve-vast-deployment
+    python3 "$DEPLOY_DIR/deployment.py"
     exit $?
 fi
 # ─── End SDK Deployment Mode ───────────────────────────────────────────
@@ -339,19 +299,19 @@ set +e
 PY_STATUS=1
 
 if [ -f "$SERVER_DIR/worker.py" ]; then
-    echo "Running worker.py"
+    echo "trying worker.py"
     python3 -m "worker" |& tee -a "$PYWORKER_LOG"
     PY_STATUS=${PIPESTATUS[0]}
 fi
 
 if [ "${PY_STATUS}" -ne 0 ] && [ -f "$SERVER_DIR/workers/$BACKEND/worker.py" ]; then
-    echo "Running workers.${BACKEND}.worker"
+    echo "trying 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/$BACKEND/server.py" ]; then
-    echo "Running workers.${BACKEND}.server"
+    echo "trying workers.${BACKEND}.server"
     python3 -m "workers.${BACKEND}.server" |& tee -a "$PYWORKER_LOG"
     PY_STATUS=${PIPESTATUS[0]}
 fi
@@ -365,4 +325,4 @@ if [ "${PY_STATUS}" -ne 0 ]; then
     report_error_and_exit "PyWorker exited with status ${PY_STATUS}"
 fi
 
-echo "PyWorker bootstrap complete"
+echo "launching PyWorker server done"