mstar.api_server.openai.adapters

Per-model translation between OpenAI-shaped requests and mstar’s request path.

The OpenAI endpoints are model-agnostic; everything model-specific lives here. An adapter translates an OpenAI request into SubmitArgs (the arguments APIServer.submit_request expects) and declares which OpenAI surfaces the model supports. Output chunks are translated back to OpenAI shapes by the serving handlers, which are generic across models.

model_kwargs is non-standardized across models, so each adapter maps the standard OpenAI fields (temperature, top_p, max_tokens, seed, voice, modalities …) onto the keys its model actually honors (see _apply_sampling()). Knobs that aren’t OpenAI-standard (top_k, repetition_penalty, or model-namespaced keys like talker_top_p) are not first-class fields — pass them via the OpenAI client’s extra_body and they flow through verbatim as model_kwargs (see _passthrough()).

Models whose outputs have no OpenAI-standard representation (robot actions from Pi0.5, world-model latents from V-JEPA 2) intentionally have no adapter: they are served only through the native /generate endpoint and the Python SDK, and /v1/* returns 404 for them (they do not fall back to chat). New OpenAI-capable models opt in by adding an adapter and registering it in ADAPTER_REGISTRY.

Functions

flatten_messages(messages, upload_dir[, ...])	Flatten OpenAI chat messages into (text, file_paths, input_modalities, parts).
get_adapter(model_name)

Classes

BagelAdapter()	BAGEL: text chat (text out) + text-to-image / image editing.
Cosmos3Adapter()	NVIDIA Cosmos3: text-to-image and text/image-to-video generation.
OpenAIAdapter()	Base adapter.
OrpheusAdapter()	Orpheus: text-to-speech (audio out only).
Qwen3OmniAdapter()	Qwen3-Omni: multimodal chat (text, optionally + speech) and TTS.
SubmitArgs(text, file_paths, ...)
Wan22Adapter()	Wan2.2-TI2V-5B: text/image-to-video generation (video only).

class mstar.api_server.openai.adapters.BagelAdapter

Bases: OpenAIAdapter

BAGEL: text chat (text out) + text-to-image / image editing.

BAGEL’s get_sampling_config reads the model config, so per-request temperature / top_p are not honored; max_output_tokens and seed are.

chat_to_request(req, upload_dir)

Parameters:

• req (ChatCompletionRequest)

• upload_dir (Path)

Return type:

SubmitArgs

image_edit_to_request(prompt, image_path, extra_kwargs)

Parameters:

• prompt (str)

• image_path (str)

• extra_kwargs (dict)

Return type:

SubmitArgs

image_to_request(req, upload_dir)

Parameters:

• req (ImageGenerationRequest)

• upload_dir (Path)

Return type:

SubmitArgs

supports_chat: bool = True

supports_images: bool = True

class mstar.api_server.openai.adapters.Cosmos3Adapter

Bases: OpenAIAdapter

NVIDIA Cosmos3: text-to-image and text/image-to-video generation.

size (“WxH”) maps to the generation resolution; seed and any extra knobs (guidance_scale, num_inference_steps, negative_prompt, and for video num_frames / fps) pass through via extra_body.

image_to_request(req, upload_dir)

Parameters:

• req (ImageGenerationRequest)

• upload_dir (Path)

Return type:

SubmitArgs

supports_images: bool = True

supports_videos: bool = True

video_to_request(req, upload_dir)

Parameters:

• req (VideoGenerationRequest)

• upload_dir (Path)

Return type:

SubmitArgs

class mstar.api_server.openai.adapters.OpenAIAdapter

Bases: object

Base adapter. A model subclasses this and implements the surfaces it supports; an unimplemented surface raises and the endpoint returns 404. Models with no OpenAI-standard output have no adapter (see the module docstring) and are reached only via /generate / the SDK.

chat_to_request(req, upload_dir)

Parameters:

• req (ChatCompletionRequest)

• upload_dir (Path)

Return type:

SubmitArgs

image_edit_to_request(prompt, image_path, extra_kwargs)

Parameters:

• prompt (str)

• image_path (str)

• extra_kwargs (dict)

Return type:

SubmitArgs

image_to_request(req, upload_dir)

Parameters:

• req (ImageGenerationRequest)

• upload_dir (Path)

Return type:

SubmitArgs

speech_to_request(req, upload_dir)

Parameters:

• req (SpeechRequest)

• upload_dir (Path)

Return type:

SubmitArgs

supports_chat: bool = False

supports_images: bool = False

supports_realtime: bool = False

supports_speech: bool = False

supports_videos: bool = False

video_to_request(req, upload_dir)

Parameters:

• req (VideoGenerationRequest)

• upload_dir (Path)

Return type:

SubmitArgs

class mstar.api_server.openai.adapters.OrpheusAdapter

Bases: OpenAIAdapter

Orpheus: text-to-speech (audio out only). Honors temperature/top_p/seed (its get_sampling_config reads model_kwargs).

speech_to_request(req, upload_dir)

Parameters:

• req (SpeechRequest)

• upload_dir (Path)

Return type:

SubmitArgs

supports_speech: bool = True

class mstar.api_server.openai.adapters.Qwen3OmniAdapter

Bases: OpenAIAdapter

Qwen3-Omni: multimodal chat (text, optionally + speech) and TTS.

Sampling is split across three stages: the Thinker (text) takes thinker_* keys, the Talker (speech) talker_*, and the Talker’s CodePredictor (residual codec groups) code_predictor_*. Everything except temperature/top_p is not an OpenAI field (talker_top_k, talker_repetition_penalty, code_predictor_top_k, …) — pass those via extra_body.

chat_to_request(req, upload_dir)

Parameters:

• req (ChatCompletionRequest)

• upload_dir (Path)

Return type:

SubmitArgs

speech_to_request(req, upload_dir)

Parameters:

• req (SpeechRequest)

• upload_dir (Path)

Return type:

SubmitArgs

supports_chat: bool = True

supports_realtime: bool = True

supports_speech: bool = True

class mstar.api_server.openai.adapters.SubmitArgs(text: 'str | None' = None, file_paths: 'dict[str, list[str]] | None'=None, input_modalities: 'list[str]' = <factory>, output_modalities: 'list[str]' = <factory>, model_kwargs: 'dict' = <factory>, prompt_parts: 'list[PromptPart] | None' = None)

Bases: object

Parameters:

• text (str | None)

• file_paths (dict[str, list[str]] | None)

• input_modalities (list[str])

• output_modalities (list[str])

• model_kwargs (dict)

• prompt_parts (list[PromptPart] | None)

file_paths: dict[str, list[str]] | None = None

input_modalities: list[str]

model_kwargs: dict

output_modalities: list[str]

prompt_parts: list[PromptPart] | None = None

text: str | None = None

class mstar.api_server.openai.adapters.Wan22Adapter

Bases: OpenAIAdapter

Wan2.2-TI2V-5B: text/image-to-video generation (video only).

size (“WxH”) maps to the model’s width / height kwargs; seed and any extra knobs (guidance_scale, num_inference_steps, negative_prompt) pass through via extra_body. fps is a playback rate (the mp4 container rate), not a generation knob — see Wan22Model.postprocess.

supports_videos: bool = True

video_to_request(req, upload_dir)

Parameters:

• req (VideoGenerationRequest)

• upload_dir (Path)

Return type:

SubmitArgs

mstar.api_server.openai.adapters.flatten_messages(messages, upload_dir, allow_remote=True)

Flatten OpenAI chat messages into (text, file_paths, input_modalities, parts).

parts is the ordered sequence as written; the other three derive from it — text newline-joined, attachments persisted under upload_dir and grouped by modality, input_modalities the per-part modality sequence. So an attachment’s position and a repeated modality both survive. (Multi-turn role structure is flattened — a v1 simplification; the models apply their own prompt formatting in process_prompt.)

Parameters:

• messages (list)

• upload_dir (Path)

• allow_remote (bool)

Return type:

tuple[str | None, dict[str, list[str]], list[str], list[PromptPart]]

mstar.api_server.openai.adapters.get_adapter(model_name)

Parameters:

model_name (str)

Return type:

OpenAIAdapter | None