vllm-project/vllm · error

max_logprobs must be non-negative or -1, got {}

Error message

max_logprobs must be non-negative or -1, got {}

What it means

Thrown by Config::validate() (the frontend server Config, not RenderConfig) when the optional --max-logprobs value is less than -1. -1 is the sentinel for 'engine default/unbounded'; 0 and positive values are valid. Validation happens before engine startup, after parser-override, CORS, and TLS checks.

Source

Thrown at rust/src/server/src/config.rs:237

    pub keep_alive_timeout: Duration,
    /// Profiler mode that registers `/start_profile` and `/stop_profile`
    /// routes when present.
    pub profiler: Option<String>,
}

impl Config {
    /// Validate frontend configuration that can be checked before engine
    /// startup.
    pub fn validate(&self) -> Result<()> {
        vllm_chat::validate_parser_overrides(&self.tool_call_parser, &self.reasoning_parser)?;
        self.cors.validate()?;
        if let Some(tls) = &self.tls {
            tls.validate()?;
        }
        if let Some(max_logprobs) = self.max_logprobs
            && max_logprobs < -1
        {
            bail!(
                "max_logprobs must be non-negative or -1, got {}",
                max_logprobs
            );
        }
        if self.data_parallel_size == 0 {
            bail!("data parallel size must be at least 1");
        }
        if self.data_parallel_size > usize::from(u16::MAX) + 1 {
            bail!(
                "data parallel size ({}) exceeds the two-byte engine identity limit",
                self.data_parallel_size
            );
        }
        match &self.transport_mode {
            TransportMode::HandshakeOwner { engine_count, .. } => {
                if *engine_count != self.data_parallel_size {
                    bail!(
                        "managed frontend engine count ({engine_count}) must equal data parallel size ({})",

View on GitHub (pinned to c794754062)

Solutions

1. Use a non-negative integer to cap logprobs (e.g. --max-logprobs 20).
2. Use exactly -1 to defer to the engine default; no other negative is accepted.
3. If the value comes from an env var or template, print it before launch to catch arithmetic mistakes.

Example fix

# before
--max-logprobs -2

# after
--max-logprobs -1   # engine default
# or
--max-logprobs 20

fn valid_max_logprobs(v: Option<i32>) -> bool {
    v.is_none_or(|n| n >= -1)
}

fn is_valid_max_logprobs(v: Option<i32>) -> bool {
    v.is_none_or(|n| n >= -1)
}

When it happens

Trigger: Starting the server with --max-logprobs -2 or any smaller negative number. A Some(max_logprobs) value below -1 trips the let-chain guard and bails with the offending number in the message.

Common situations: Scripts that pass -1 for 'unlimited' but compute the value (e.g. `--max-logprobs $((0 - 2))`), or users assuming any negative means unlimited. Migrating from Python vLLM where the sentinel semantics differed.

Related errors

• max_logprobs must be non-negative or -1
• data parallel size must be at least 1
• engine start index does not fit usize
• invalid --allowed-methods value {method:?}: {e}
• invalid --allowed-headers value {header:?}: {e}

AI-assisted analysis of vllm-project/vllm@c794754062 (2026-08-14). Data as JSON: /api/errors/8b82c76b1c7227f0. Report an issue: GitHub.