vllm.entrypoints.openai.fingerprint ¶

Build the system_fingerprint string returned by the OpenAI-compatible server.

Four modes, configured via --fingerprint-mode:

• full (default): vllm-<version>[-<parallelism>]-<hash8> — encodes server version, any non-trivial parallelism degree (tp/pp/dp/ep), and an 8-char prefix of vllm_config.compute_hash() (covers model identity, quant config, speculative, attention backend, etc.).
• hash: vllm-<version>-<hash8> — parallelism stripped.
• custom: user-provided literal via --fingerprint-value.
• none: the field is omitted (serialized as null).

get_system_fingerprint is only called at serving-class init (a handful of times per server); each subclass caches the returned string on self.system_fingerprint, so per-request cost is one attribute read.

get_system_fingerprint(vllm_config: Any) -> str | None

42
43
44
45

def get_system_fingerprint(vllm_config: Any) -> str | None:
    """Return the fingerprint for ``vllm_config`` using the mode configured by
    ``set_default_fingerprint_mode``."""
    return build_system_fingerprint(vllm_config, _DEFAULT_MODE, _CUSTOM_VALUE)

set_default_fingerprint_mode(
    mode: FingerprintMode, custom_value: str | None = None
) -> None

31
32
33
34
35
36
37
38
39

def set_default_fingerprint_mode(
    mode: FingerprintMode,
    custom_value: str | None = None,
) -> None:
    """Configure the fingerprint mode for subsequent ``get_system_fingerprint``
    calls. Called once at server startup."""
    global _DEFAULT_MODE, _CUSTOM_VALUE
    _DEFAULT_MODE = mode
    _CUSTOM_VALUE = custom_value