BerriAI/litellm · error · VertexAIError

Upgrade vertex ai. Run `pip install "google-cloud-aiplatform

Q: What causes "Upgrade vertex ai. Run `pip install "google-cloud-aiplatform>=1.38"`" in BerriAI/litellm?

pip install -U "google-cloud-aiplatform>=1.38".

Error message

Upgrade vertex ai. Run `pip install "google-cloud-aiplatform>=1.38"`

What it means

Raised by the Vertex Model Garden handler right after the import check: it verifies the installed `vertexai` SDK exposes the `preview.language_models` API needed for OpenAI-compatible Model Garden calls. If the version predicate fails, the installed google-cloud-aiplatform is too old (< 1.38) and lacks that preview surface. Note the guard is lenient — `hasattr(vertexai, "preview")` alone short-circuits the check — so in practice this fires mainly on very old installs where `preview` is absent entirely.

Source

Thrown at litellm/llms/vertex_ai/vertex_model_garden/main.py:97

        client=None,
    ):
        """
        Handles calling Vertex AI Model Garden Models in OpenAI compatible format

        Sent to this route when `model` is in the format `vertex_ai/openai/{MODEL_ID}`
        """
        try:
            import vertexai

            from litellm.llms.openai_like.chat.handler import OpenAILikeChatHandler
        except Exception as e:
            raise VertexAIError(
                status_code=400,
                message=f"""vertexai import failed please run `pip install -U "google-cloud-aiplatform>=1.38"`. Got error: {e}""",
            )

        if not (hasattr(vertexai, "preview") or hasattr(vertexai.preview, "language_models")):
            raise VertexAIError(
                status_code=400,
                message="""Upgrade vertex ai. Run `pip install "google-cloud-aiplatform>=1.38"`""",
            )
        try:
            model = get_vertex_base_model_name(model=model)

            access_token, project_id = self._ensure_access_token(
                credentials=vertex_credentials,
                project_id=vertex_project,
                custom_llm_provider="vertex_ai",
            )

            openai_like_chat_completions: Final = OpenAILikeChatHandler()

            ## CONSTRUCT API BASE
            # Skip _check_custom_proxy: its ":verb" URL construction corrupts a
            # user-supplied api_base (e.g. Vertex MG dedicated endpoint), and
            # OpenAILikeChatHandler already appends "/chat/completions".

View on GitHub (pinned to 77b7c6c40c)

Solutions

pip install -U "google-cloud-aiplatform>=1.38".
If a hard pin blocks the upgrade, relax the constraint in requirements.txt/lockfile (check pip check for conflicts).
Verify with python -c "import vertexai; print(vertexai.preview.language_models)" that the preview API is present.
Rebuild Docker/CI images after the upgrade so stale layers don't re-inject the old SDK.

Example fix

# requirements.txt
# before
google-cloud-aiplatform==1.35.0

# after
google-cloud-aiplatform>=1.38

Defensive patterns

Strategy: validation

Validate before calling

def vertexai_version_ok() -> bool:
    try:
        import vertexai
        return hasattr(vertexai, "preview")
    except Exception:
        return False

assert vertexai_version_ok(), "Upgrade: pip install -U 'google-cloud-aiplatform>=1.38'"

Try / catch

try:
    litellm.completion(model="vertex_ai/openai/...", messages=msgs)
except VertexAIError as e:
    if "Upgrade vertex ai" in str(e):
        raise RuntimeError("Incompatible google-cloud-aiplatform version — upgrade to >=1.38") from e
    raise

Prevention

Pin google-cloud-aiplatform>=1.38,<next-major in requirements to prevent stale pins from other tools.
Run `pip check` in CI to catch dependency conflicts on google packages.
Test model-garden routes in CI against the same image versions you deploy.

When it happens

Trigger: Using a model of the form vertex_ai/openai/{MODEL_ID} with google-cloud-aiplatform pinned below 1.38 (common in lockfiles/requirements that were generated before the SDK matured), where `vertexai` imports fine but has no `preview` attribute.

Common situations: Corporate environments pinning google-cloud-aiplatform==1.35.x for other tooling; base images like old airflow/python templates shipping stale SDKs; upgrading litellm but not the google extras.

Related errors

AI-assisted analysis of BerriAI/litellm@77b7c6c40c (2026-08-18). Data as JSON: /api/errors/a578f59e67e7853c. Report an issue: GitHub.