Is this page helpful?

core.inference.apis.serve_config#

Module Contents#

Classes#

ServeConfig

Programmatic configuration for MegatronAsyncLLM.serve(...).

API#

class core.inference.apis.serve_config.ServeConfig#

Programmatic configuration for MegatronAsyncLLM.serve(...).

This dataclass also serves as the future source of truth for a megatron serve CLI. It controls only the HTTP serving surface; engine construction and coordinator addressing are configured separately via the MegatronLLM / MegatronAsyncLLM constructor.

host: str#

‘0.0.0.0’

HTTP bind host for the OpenAI-compatible frontend.

Distinct from the MegatronLLM / MegatronAsyncLLM constructor’s coordinator_host argument: coordinator_host is the internal/routable address used for coordinator ZMQ traffic, whereas host is the externally-visible interface where the HTTP server accepts client connections.

port: int#

5000

HTTP bind port for the OpenAI-compatible frontend.

parsers: list[str]#

‘field(…)’

Response parser names to enable on the HTTP frontend.

Examples include ["json", "tool_use"]. Values are passed through to the underlying text-generation server unchanged.

verbose: bool#

False

Whether the HTTP frontend should log per-request detail.

frontend_replicas: int#

4

Number of HTTP frontend processes spawned on the primary rank.

The default of 4 matches the existing start_text_gen_server default of num_replicas=4.

sock: socket.socket | None#

None

Pre-bound listening socket to serve on instead of binding host:port.

Must already be bound to a real port; when set, host / port are not used for binding.

default_temperature: float#

1.0

Default temperature value when an HTTP request omits temperature.

default_top_p: float#

1.0

Default top-p value when an HTTP request omits top_p.

default_top_k: int#

0

Default top-k value when an HTTP request omits top_k.

eval_mode: bool#

False

Use evaluation defaults instead of RL-oriented response behavior.

In evaluation mode, chat requests default prevent_retokenization to false, avoiding transmission of prompt token IDs. Individual requests can still opt in by setting prevent_retokenization or return_tokenized_data.