Merge branch 'main' into pyworker-sdk

.devcontainer/devcontainer.json

@@ -1,37 +0,0 @@
-// .devcontainer/devcontainer.json
-// Dev container for the Vast.ai serverless Ollama template.
-// Includes Docker-in-Docker so you can build and test images from inside the container.
-{
-  "name": "vast.ai-serverless-ollama",
-  "image": "mcr.microsoft.com/devcontainers/base:trixie",
-  "features": {
-    "ghcr.io/devcontainers/features/python:1": {
-      "installTools": true,
-      "version": "3.12"
-    },
-    "ghcr.io/devcontainers/features/docker-in-docker:3.0.1": {
-      "moby": false,
-      "version": "latest",
-      "installDockerBuildx": true,
-      "dockerDashComposeVersion": "v2"
-    }
-  },
-  "runArgs": ["--privileged"],
-  "containerEnv": {
-    "DOCKER_BUILDKIT": "1"
-  },
-  "postCreateCommand": "python3 -m pip install --user --upgrade pip && python3 -m pip install --user -r requirements.txt pyyaml",
-  "customizations": {
-    "vscode": {
-      "extensions": [
-        "ms-python.python",
-        "ms-azuretools.vscode-docker"
-      ],
-      "settings": {
-        "python.defaultInterpreterPath": "/usr/bin/python3",
-        "terminal.integrated.defaultProfile.linux": "bash",
-        "docker.showStartPage": false
-      }
-    }
-  }
-}

requirements.txt

@@ -1 +1,13 @@
-vastai-sdk
+aiohttp==3.10.1
+aiodns~=3.6.0
+pycares~=4.11.0
+anyio~=4.4
+lib~=4.0
+nltk~=3.9
+psutil~=6.0
+pycryptodome~=3.20
+Requests~=2.32
+transformers~=4.52
+utils==1.0.*
+hf_transfer>=0.1.9
+vastai-sdk>=0.3.0

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"
 
@@ -53,42 +46,6 @@ JSON
     exit 1
 }
 
-function install_vastai_sdk() {
-    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 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 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"
-    if ! uv pip install "${uv_flags[@]}" vastai; then
-        report_error_and_exit "Failed to install vastai"
-    fi
-}
-
 [ -n "$BACKEND" ] && [ -z "$HF_TOKEN" ] && report_error_and_exit "HF_TOKEN must be set when BACKEND is set!"
 [ -z "$CONTAINER_ID" ] && report_error_and_exit "CONTAINER_ID must be set!"
 [ "$BACKEND" = "comfyui" ] && [ -z "$COMFY_MODEL" ] && report_error_and_exit "For comfyui backends, COMFY_MODEL must be set!"
@@ -106,8 +63,7 @@ echo_var DEBUG_LOG
 echo_var PYWORKER_LOG
 echo_var MODEL_LOG
 
-ROTATE_MODEL_LOG="${ROTATE_MODEL_LOG:-false}"
+if [ -e "$MODEL_LOG" ]; then
-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"
@@ -128,21 +84,8 @@ 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
@@ -161,27 +104,10 @@ 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
+        if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF"); then
-            echo "Force updating to pyworker reference: $PYWORKER_REF"
+            report_error_and_exit "Failed to checkout 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
 
@@ -197,8 +123,6 @@ elif [ ! -d "$ENV_PATH" ]; then
         report_error_and_exit "Failed to install Python requirements"
     fi
 
-    install_vastai_sdk
-
     if ! touch ~/.no_auto_tmux; then
         report_error_and_exit "Failed to create ~/.no_auto_tmux"
     fi
@@ -208,44 +132,11 @@ 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
@@ -283,51 +174,16 @@ 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
+        --data-binary @/etc/instance.csr \
-    for attempt in $(seq 1 "$max_retries"); do
+        -X \
-        http_code=$(curl -sS -o /etc/instance.crt -w '%{http_code}' \
+        POST "https://console.vast.ai/api/v0/sign_cert/?instance_id=$CONTAINER_ID" > /etc/instance.crt; then
-            --header 'Content-Type: application/octet-stream' \
+        report_error_and_exit "Failed to sign SSL certificate"
-            --data-binary @/etc/instance.csr \
+    fi
-            -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
@@ -339,19 +195,19 @@ 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 +221,4 @@ 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"

workers/openai/worker.py

@@ -1,45 +1,37 @@
+import nltk
 import random
 import os
-import re
-import logging
 
 from vastai import Worker, WorkerConfig, HandlerConfig, LogActionConfig, BenchmarkConfig
 
-logging.getLogger().setLevel(logging.WARNING)  # Only show warnings and errors
+# vLLM model configuration
-
-# Ollama model configuration
 MODEL_SERVER_URL           = 'http://127.0.0.1'
-MODEL_SERVER_PORT          = 11434
+MODEL_SERVER_PORT          = 18000
-MODEL_LOG_FILE             = '/var/log/onstart.log'
+MODEL_LOG_FILE             = '/var/log/portal/vllm.log'
-MODEL_HEALTHCHECK_ENDPOINT = "/"
+MODEL_HEALTHCHECK_ENDPOINT = "/health"
 
-# Ollama-specific log messages
+# vLLM-specific log messages
-def request_parser(request):
+MODEL_LOAD_LOG_MSG = [
-    data = request
+    "Application startup complete.",
-    if request.get("input") is not None:
+]
-        data = request.get("input")
+
-    return data
+MODEL_ERROR_LOG_MSGS = [
+    "INFO exited: vllm",
+    "RuntimeError: Engine",
+    "Traceback (most recent call last):"
+]
+
+MODEL_INFO_LOG_MSGS = [
+    '"message":"Download'
+]
+
+nltk.download("words")
+WORD_LIST = nltk.corpus.words.words()
 
 
 def completions_benchmark_generator() -> dict:
-    # extract words from the python source code of the worker to create a list of words for generating prompts
-
-    WORD_LIST = []
-
-    # Try to load from perl copyright file first
-    try:
-        with open("/usr/share/doc/perl/copyright", 'r') as f:
-            source_code = f.read()
-            WORD_LIST = re.findall(r'\b\w+\b', source_code)
-    except (FileNotFoundError, IOError):
-        # Fallback to loading from python file
-        with open(__file__, 'r') as f:
-            source_code = f.read()
-            WORD_LIST = re.findall(r'\b\w+\b', source_code)
-
     prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
     model = os.environ.get("MODEL_NAME")
-
     if not model:
         raise ValueError("MODEL_NAME environment variable not set")
 
@@ -62,26 +54,24 @@ worker_config = WorkerConfig(
             route="/v1/completions",
             workload_calculator= lambda data: data.get("max_tokens", 0),
             allow_parallel_requests=True,
-            request_parser=request_parser,
+            max_queue_time=60.0,
-            max_queue_time=600.0,
             benchmark_config=BenchmarkConfig(
                 generator=completions_benchmark_generator,
-                concurrency=10,
+                concurrency=100,
-                runs=3
+                runs=2
             )
         ),
         HandlerConfig(
             route="/v1/chat/completions",
             workload_calculator= lambda data: data.get("max_tokens", 0),
             allow_parallel_requests=True,
-            request_parser=request_parser,
+            max_queue_time=60.0,
-            max_queue_time=600.0,
         )
     ],
     log_action_config=LogActionConfig(
-        on_load=["llama_server: model loaded"],
+        on_load=MODEL_LOAD_LOG_MSG,
-        on_error=["Traceback (most recent call last):","Error:"],
+        on_error=MODEL_ERROR_LOG_MSGS,
-        #on_info=["load_tensors:","llama_context:","print_info:","llama_model_loader:"]
+        on_info=MODEL_INFO_LOG_MSGS
     )
 )
 

workers/tgi/worker.py

@@ -35,7 +35,7 @@ def benchmark_generator() -> dict:
     benchmark_data = {
         "inputs": prompt,
         "parameters": {
-            "max_new_tokens": 500,
+            "max_new_tokens": 128,
             "temperature": 0.7,
             "return_full_text": False
         }
@@ -52,18 +52,17 @@ worker_config = WorkerConfig(
         HandlerConfig(
             route="/generate",
             allow_parallel_requests=True,
-            max_queue_time=600.0,
+            max_queue_time=60.0,
             benchmark_config=BenchmarkConfig(
                 generator=benchmark_generator,
-                concurrency=10,
+                concurrency=50
-                runs=3
             ),
             workload_calculator= lambda x: x["parameters"]["max_new_tokens"]
         ),
         HandlerConfig(
             route="/generate_stream",
             allow_parallel_requests=True,
-            max_queue_time=600.0,
+            max_queue_time=60.0,
             workload_calculator= lambda x: x["parameters"]["max_new_tokens"]
         )
     ],