mlflow/mlflow · error · AIGatewayException

Cannot set both 'temperature' and 'top_p' parameters.

Error message

Cannot set both 'temperature' and 'top_p' parameters.

What it means

Anthropic's chat API rejects combining 'temperature' and 'top_p'; the MLflow gateway's Anthropic provider enforces this client-side and raises a 422 if both keys are present in the chat payload. Only one sampling knob may be set. Remove 'top_p' (keep 'temperature') to resolve.

Source

Thrown at mlflow/gateway/providers/anthropic.py:124

    for value in schema.values():
        match value:
            case dict():
                _enforce_strict_schema(value)
            case list():
                for item in value:
                    if isinstance(item, dict):
                        _enforce_strict_schema(item)


class AnthropicAdapter(ProviderAdapter):
    @classmethod
    def chat_to_model(cls, payload, config):
        key_mapping = {"stop": "stop_sequences"}
        payload["model"] = config.model.name
        payload = rename_payload_keys(payload, key_mapping)

        if "top_p" in payload and "temperature" in payload:
            raise AIGatewayException(
                status_code=422, detail="Cannot set both 'temperature' and 'top_p' parameters."
            )

        max_completion_tokens = payload.pop("max_completion_tokens", None)
        max_tokens = payload.get("max_tokens") or max_completion_tokens
        if max_tokens is None:
            max_tokens = MLFLOW_AI_GATEWAY_ANTHROPIC_DEFAULT_MAX_TOKENS
        if max_tokens > MLFLOW_AI_GATEWAY_ANTHROPIC_MAXIMUM_MAX_TOKENS:
            raise AIGatewayException(
                status_code=422,
                detail="Invalid value for max_tokens: cannot exceed "
                f"{MLFLOW_AI_GATEWAY_ANTHROPIC_MAXIMUM_MAX_TOKENS}.",
            )
        payload["max_tokens"] = max_tokens

        if payload.pop("n", 1) != 1:
            raise AIGatewayException(
                status_code=422,

View on GitHub (pinned to 6a27f2decc)

Solutions

1. Remove 'top_p' from the payload and keep only 'temperature'.
2. Alternatively remove 'temperature' and keep only 'top_p'.
3. Strip one of the keys in the caller before sending if your framework always sets both.

Example fix

// before
payload = {"messages": msgs, "temperature": 0.7, "top_p": 0.9}
// after
payload = {"messages": msgs, "temperature": 0.7}

def sanitize_anthropic_chat(payload):
    if "temperature" in payload and "top_p" in payload:
        payload.pop("top_p", None)
    return payload
sanitize_anthropic_chat(payload)

try:
    resp = gateway.chat(endpoint="anthropic-chat", payload=payload)
except MlflowException as e:
    if "temperature" in str(e) and "top_p" in str(e):
        payload.pop("top_p", None)
        resp = gateway.chat(endpoint="anthropic-chat", payload=payload)
    else:
        raise

When it happens

Trigger: POSTing a chat request to an Anthropic gateway endpoint where both 'temperature' and 'top_p' keys exist in the payload after stop->stop_sequences renaming.

Common situations: Copying payload defaults from an OpenAI request where both were set; merging config dictionaries that each contribute one of the two keys; framework defaults that always emit both.

Related errors

• 'n' must be '1' for the Anthropic provider. Received value:
• Only function calling tool is supported, but received tool t
• Invalid value for max_tokens: cannot exceed {MLFLOW_AI_GATEW
• Cannot set both 'temperature' and 'top_p' parameters. Please
• Setting the 'stream' parameter to 'true' is not supported wi

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/8012ff8cefb701c3. Report an issue: GitHub.