Skip to content

feat: add Dynamo vLLM provider for external K8s endpoints #365

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
rwipfelnv wants to merge 4 commits into from
Closed

feat: add Dynamo vLLM provider for external K8s endpoints #365

Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
734a63c
feat: add Dynamo vLLM provider support for external endpoints
rwipfelnv Mar 18, 2026
cc6abf8
fix: address code review feedback for Dynamo provider
rwipfelnv Mar 19, 2026
cc907e9
fix: simplify Dynamo provider to minimal implementation
rwipfelnv Mar 24, 2026
6b74ffc
fix: read dynamo endpoint directly in setupInference
rwipfelnv Mar 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
124 changes: 118 additions & 6 deletions bin/lib/onboard.js
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,13 @@ async function promptOrDefault(question, envVar, defaultValue) {

// ── Helpers ──────────────────────────────────────────────────────

// Shell-safe single-quote escaping to prevent injection when interpolating
// user-supplied values into shell commands. Wraps value in single quotes
// and escapes any embedded single quotes.
function shQuote(value) {
return `'${String(value).replace(/'/g, `'\"'\"'`)}'`;
}

function step(n, total, msg) {
console.log("");
console.log(` [${n}/${total}] ${msg}`);
Expand Down Expand Up @@ -221,10 +228,10 @@ function getNonInteractiveProvider() {
const providerKey = (process.env.NEMOCLAW_PROVIDER || "").trim().toLowerCase();
if (!providerKey) return null;

const validProviders = new Set(["cloud", "ollama", "vllm", "nim"]);
const validProviders = new Set(["cloud", "ollama", "vllm", "nim", "dynamo"]);
if (!validProviders.has(providerKey)) {
console.error(` Unsupported NEMOCLAW_PROVIDER: ${providerKey}`);
console.error(" Valid values: cloud, ollama, vllm, nim");
console.error(" Valid values: cloud, ollama, vllm, nim, dynamo");
process.exit(1);
}

Expand Down Expand Up @@ -473,6 +480,90 @@ async function setupNim(sandboxName, gpu) {
const vllmRunning = !!runCapture("curl -sf http://localhost:8000/v1/models 2>/dev/null", { ignoreError: true });
const requestedProvider = isNonInteractive() ? getNonInteractiveProvider() : null;
const requestedModel = isNonInteractive() ? getNonInteractiveModel(requestedProvider || "cloud") : null;

// Auto-select only with NEMOCLAW_EXPERIMENTAL=1 (prevents silent misconfiguration)
if (EXPERIMENTAL) {
if (vllmRunning) {
console.log(" ✓ vLLM detected on localhost:8000 — using it [experimental]");
provider = "vllm-local";
model = "vllm-local";
registry.updateSandbox(sandboxName, { model, provider, nimContainer });
return { model, provider };
}
if (ollamaRunning) {
console.log(" ✓ Ollama detected on localhost:11434 — using it [experimental]");
provider = "ollama-local";
model = "nemotron-3-nano";
registry.updateSandbox(sandboxName, { model, provider, nimContainer });
return { model, provider };
}
}

// Non-interactive: honor NEMOCLAW_PROVIDER before building interactive options
if (isNonInteractive() && requestedProvider) {
const providerKey = requestedProvider;
console.log(` [non-interactive] Provider: ${providerKey}`);
if (providerKey === "dynamo") {
// Dynamo: external vLLM endpoint (e.g., K8s service)
const dynamoEndpointRaw = (process.env.NEMOCLAW_DYNAMO_ENDPOINT || "").trim();
if (!dynamoEndpointRaw) {
console.error(" NEMOCLAW_DYNAMO_ENDPOINT is required when NEMOCLAW_PROVIDER=dynamo.");
process.exit(1);
}
// Validate URL format and protocol
let parsedUrl;
try {
parsedUrl = new URL(dynamoEndpointRaw);
} catch {
console.error(` Invalid NEMOCLAW_DYNAMO_ENDPOINT URL: ${dynamoEndpointRaw}`);
process.exit(1);
}
if (!["http:", "https:"].includes(parsedUrl.protocol)) {
console.error(` NEMOCLAW_DYNAMO_ENDPOINT must use http or https: ${dynamoEndpointRaw}`);
process.exit(1);
}
// Log sanitized URL (redact credentials and query params if present)
const safeLogUrl = `${parsedUrl.protocol}//${parsedUrl.host}${parsedUrl.pathname}`;
const redacted = [];
if (parsedUrl.username || parsedUrl.password) redacted.push("credentials");
if (parsedUrl.search) redacted.push("query");
console.log(` Using Dynamo endpoint: ${safeLogUrl}${redacted.length ? ` [REDACTED: ${redacted.join(", ")}]` : ""}`);

provider = "dynamo";
model = (process.env.NEMOCLAW_DYNAMO_MODEL || "").trim() || "dynamo";
registry.updateSandbox(sandboxName, { model, provider, nimContainer, dynamoEndpoint: dynamoEndpointRaw });
return { model, provider, dynamoEndpoint: dynamoEndpointRaw };
} else if (providerKey === "ollama") {
if (!ollamaRunning) {
console.error(" Ollama is not running on localhost:11434. Start it first.");
process.exit(1);
}
provider = "ollama-local";
model = requestedModel || "nemotron-3-nano";
registry.updateSandbox(sandboxName, { model, provider, nimContainer });
return { model, provider };
} else if (providerKey === "vllm") {
if (!vllmRunning) {
console.error(" vLLM is not running on localhost:8000. Start it first.");
process.exit(1);
}
provider = "vllm-local";
model = requestedModel || "vllm-local";
registry.updateSandbox(sandboxName, { model, provider, nimContainer });
return { model, provider };
} else if (providerKey === "nim") {
if (!EXPERIMENTAL) {
console.error(" NEMOCLAW_PROVIDER=nim requires NEMOCLAW_EXPERIMENTAL=1.");
process.exit(1);
}
if (!gpu || !gpu.nimCapable) {
console.error(" Local NIM requires a compatible NVIDIA GPU.");
process.exit(1);
}
}
// "cloud" or "nim" fall through to normal flow below
}

// Build options list — only show local options with NEMOCLAW_EXPERIMENTAL=1
const options = [];
if (EXPERIMENTAL && gpu && gpu.nimCapable) {
Expand Down Expand Up @@ -642,10 +733,30 @@ async function setupNim(sandboxName, gpu) {

// ── Step 5: Inference provider ───────────────────────────────────

async function setupInference(sandboxName, model, provider) {
async function setupInference(sandboxName, model, provider, opts = {}) {
step(5, 7, "Setting up inference provider");

if (provider === "nvidia-nim") {
if (provider === "dynamo") {
// Dynamo: external vLLM endpoint (e.g., K8s service)
const dynamoEndpoint = opts.dynamoEndpoint;
if (!dynamoEndpoint) {
console.error(" Dynamo provider requires dynamoEndpoint in options.");
process.exit(1);
}
// Use shQuote for shell-safe escaping of the endpoint URL
run(
`openshell provider create --name dynamo --type openai ` +
`--credential "OPENAI_API_KEY=dummy" ` +
`--config "OPENAI_BASE_URL=${shQuote(dynamoEndpoint)}" 2>&1 || ` +
`openshell provider update dynamo --credential "OPENAI_API_KEY=dummy" ` +
`--config "OPENAI_BASE_URL=${shQuote(dynamoEndpoint)}" 2>&1 || true`,
{ ignoreError: true }
);
run(
`openshell inference set --no-verify --provider dynamo --model ${shQuote(model)} 2>/dev/null || true`,
{ ignoreError: true }
);
} else if (provider === "nvidia-nim") {
// Create nvidia-nim provider
run(
`openshell provider create --name nvidia-nim --type openai ` +
Expand Down Expand Up @@ -850,6 +961,7 @@ function printDashboard(sandboxName, model, provider) {
if (provider === "nvidia-nim") providerLabel = "NVIDIA Cloud API";
else if (provider === "vllm-local") providerLabel = "Local vLLM";
else if (provider === "ollama-local") providerLabel = "Local Ollama";
else if (provider === "dynamo") providerLabel = "Dynamo vLLM";

console.log("");
console.log(` ${"─".repeat(50)}`);
Expand Down Expand Up @@ -878,8 +990,8 @@ async function onboard(opts = {}) {
const gpu = await preflight();
await startGateway(gpu);
const sandboxName = await createSandbox(gpu);
const { model, provider } = await setupNim(sandboxName, gpu);
await setupInference(sandboxName, model, provider);
const { model, provider, dynamoEndpoint } = await setupNim(sandboxName, gpu);
await setupInference(sandboxName, model, provider, { dynamoEndpoint });
await setupOpenclaw(sandboxName, model, provider);
await setupPolicies(sandboxName);
printDashboard(sandboxName, model, provider);
Expand Down
Loading