Add devcontainer configuration for Vast.ai serverless Ollama template

* Add nltk requirement for openai worker

* pin version

.devcontainer/devcontainer.json

@@ -0,0 +1,37 @@
+// .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:3.0.0": {
+      "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
+      }
+    }
+  }
+}

README.md

@@ -9,7 +9,7 @@ This repository contains **example PyWorkers** used by Vast.ai’s default Serve
 - Optionally supports FIFO queueing when the backend cannot process concurrent requests
 - Detects readiness/failure from model logs and runs a benchmark to estimate throughput
 
-> Important: The **core PyWorker framework** (Worker, WorkerConfig, HandlerConfig, BenchmarkConfig, LogActionConfig) is provided by the **`vastai`** Python package (https://github.com/vast-ai/vast-cli). This repo focuses on *worker implementations and examples*, not the framework internals.
+> Important: The **core PyWorker framework** (Worker, WorkerConfig, HandlerConfig, BenchmarkConfig, LogActionConfig) is provided by the **`vastai` / `vastai-sdk`** Python package (https://github.com/vast-ai/vast-sdk). This repo focuses on *worker implementations and examples*, not the framework internals.
 
 ## Repository Purpose
 

lib/wheres-my-pyworker.txt

@@ -1,16 +1,16 @@
 # Where did the PyWorker code go?
-We have moved the PyWorker source code into the `vastai` Python package.
+We have moved the PyWorker source code into the `vastai-sdk` Python SDK.
 You can install it with
 ```
-pip install vastai
+pip install vastai-sdk
 ```
 
 All of the source code can be found here:
-https://github.com/vast-ai/vast-cli
+https://github.com/vast-ai/vast-sdk
 
 And can be imported from vastai.serverless.server.lib
 
-Serverless instances automatically run the start_server.sh script, which installs the vastai package.
+Serverless instances automatically run the start_server.sh script, which installs the vastai-sdk.
 This is how the PyWorker source code makes it onto your serverless instances.
 You provide a worker.py file in your PYWORKER_REPO, and the start_server.sh will 
 create and run a PyWorker according to your configuration defined in the file.

requirements.txt

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

start_server.sh

@@ -53,7 +53,7 @@ JSON
     exit 1
 }
 
-function install_vastai() {
+function install_vastai_sdk() {
     local uv_flags=()
     if [ "${USE_SYSTEM_PYTHON:-}" = "true" ]; then
         uv_flags+=(--system --break-system-packages)
@@ -77,8 +77,8 @@ function install_vastai() {
 
     if [ -n "${SDK_VERSION:-}" ]; then
         echo "Installing vastai version ${SDK_VERSION}"
-        if ! uv pip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ "${uv_flags[@]}" "vastai-sdk-vast==${SDK_VERSION}"; then
+        if ! uv pip install "${uv_flags[@]}" "vastai==${SDK_VERSION}"; then
-            report_error_and_exit "Failed to install vastai-vast==${SDK_VERSION}"
+            report_error_and_exit "Failed to install vastai==${SDK_VERSION}"
         fi
         return 0
     fi
@@ -140,7 +140,7 @@ if [ "${USE_SYSTEM_PYTHON:-}" = "true" ]; then
             fi
         fi
     fi
-    install_vastai
+    install_vastai_sdk
     touch ~/.no_auto_tmux
 elif [ ! -d "$ENV_PATH" ]; then
     echo "setting up venv"
@@ -197,7 +197,7 @@ elif [ ! -d "$ENV_PATH" ]; then
         report_error_and_exit "Failed to install Python requirements"
     fi
 
-    install_vastai
+    install_vastai_sdk
 
     if ! touch ~/.no_auto_tmux; then
         report_error_and_exit "Failed to create ~/.no_auto_tmux"
@@ -237,7 +237,7 @@ else
             fi
         fi
 
-        install_vastai
+        install_vastai_sdk
     fi
 fi
 
@@ -318,8 +318,8 @@ if [ "$IS_DEPLOYMENT" = "true" ]; then
     # 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 function which supports SDK_BRANCH/SDK_VERSION)
+    # Install SDK (uses the install_vastai_sdk function which supports SDK_BRANCH/SDK_VERSION)
-    install_vastai
+    install_vastai_sdk
     # Run deployment in serve mode
     export VAST_DEPLOYMENT_MODE=serve
     echo "Starting deployment: python3 $DEPLOY_DIR/deployment.py"
@@ -339,19 +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
@@ -365,4 +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"

workers/openai/worker.py

@@ -7,12 +7,15 @@ from vastai import Worker, WorkerConfig, HandlerConfig, LogActionConfig, Benchma
 # vLLM model configuration
 MODEL_SERVER_URL           = 'http://127.0.0.1'
 MODEL_SERVER_PORT          = 18000
-MODEL_LOG_FILE             = '/var/log/portal/vllm.log'
+MODEL_LOG_FILE             = '/var/log/onstart.log'
 MODEL_HEALTHCHECK_ENDPOINT = "/health"
 
 # vLLM-specific log messages
 MODEL_LOAD_LOG_MSG = [
     "Application startup complete.",
+    "llama runner started in",
+    "Server listening on",
+    "msg=\"Listening on",
 ]
 
 MODEL_ERROR_LOG_MSGS = [

workers/tgi/worker.py

@@ -35,7 +35,7 @@ def benchmark_generator() -> dict:
     benchmark_data = {
         "inputs": prompt,
         "parameters": {
-            "max_new_tokens": 128,
+            "max_new_tokens": 500,
             "temperature": 0.7,
             "return_full_text": False
         }