BerriAI/litellm · error · ValueError

Unsupported call_type={call_type!r} for compression. Expecte

Error message

Unsupported call_type={call_type!r} for compression. Expected one of: {sorted(_SUPPORTED_CALL_TYPES)}.

What it means

_normalize_messages_for_compression in litellm's compression module only accepts call types whose payload is a list of role/content messages: completion, acompletion, and anthropic_messages. Any other call_type (e.g. embedding, image_generation, responses, transcription) is rejected up front, because compression operates on message lists.

Source

Thrown at litellm/compression/compress.py:110

            if item_type == "text":
                parts.append(str(item.get("text", "")))
            elif item_type == "tool_result":
                stack.append(item.get("content", ""))
    return " ".join(parts)


def _normalize_messages_for_compression(
    messages: list[dict],
    call_type: str,
) -> tuple[list[dict], list[dict]]:
    """
    Normalize each original message to a text-surrogate content for scoring.

    Returns:
        (normalized_messages, original_messages_copy)
    """
    if call_type not in _SUPPORTED_CALL_TYPES:
        raise ValueError(
            f"Unsupported call_type={call_type!r} for compression. Expected one of: {sorted(_SUPPORTED_CALL_TYPES)}."
        )

    original_messages: Final[list[dict[str, Any]]] = [dict(m) for m in messages]

    normalized_messages: Final[list[dict]] = []
    for msg in original_messages:
        normalized_messages.append(
            {
                **msg,
                "content": _content_to_text(msg.get("content", "")),
            }
        )
    return normalized_messages, original_messages


def _extract_last_user_message(messages: list[dict]) -> str:
    """Return the text content of the last user message."""

View on GitHub (pinned to 6c2dcb801b)

Solutions

1. Scope compression to chat call types only — do not apply it to embedding/responses/image calls
2. If configuring via router/proxy, filter by call type in the compression hook or set model_info so non-chat models skip compression
3. Pass the exact supported values: 'completion', 'acompletion', or 'anthropic_messages'

Example fix

# before — compression applied to every call
router_settings = {"compression": {"enabled": True}}  # wraps embedding calls too

# after — only compress chat completions
if call_type in ("completion", "acompletion", "anthropic_messages"):
    result = await compress_and_call(call_type, kwargs)
else:
    result = await original_call(**kwargs)

SUPPORTED = {"completion", "acompletion", "anthropic_messages"}

def compression_supported(call_type: str) -> bool:
    return call_type in SUPPORTED

When it happens

Trigger: Enabling compression and invoking litellm with a call type outside _SUPPORTED_CALL_TYPES — e.g. litellm.embedding, litellm.responses, litellm.anthropic_openshift or a router deployment whose model_type maps to an unsupported call type, with compression hooks applied.

Common situations: Turning on prompt compression globally (router/proxy settings) so it wraps every call type including embeddings/responses; misconfigured model_info.model_type on a router deployment; passing a typo'd call type string.

Related errors

• Invalid mode: {custom_auth_settings['mode']}
• 'cp4d_host' is required in litellm_params for WXO agents
• 'instance_id' is required in litellm_params for WXO agents
• 'wxo_agent_id' is required in litellm_params for WXO agents
• 'api_key' is required in litellm_params for WXO agents

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/70097ea4aec9f13a. Report an issue: GitHub.