nemo_rl.models.generation.vllm.captured_chat_logprobs#

Avoid constructing OpenAI logprob objects discarded by terminal capture.

Module Contents#

Classes#

_ChatLogprobsBuilder

CapturedChatLogprobsMixin

Scope a formatting-only bypass to one serving instance and request task.

Data#

_G_BYPASS_OWNER

API#

nemo_rl.models.generation.vllm.captured_chat_logprobs._G_BYPASS_OWNER: contextvars.ContextVar[tuple[object, asyncio.Task[Any] | None, int] | None]#

‘ContextVar(…)’

class nemo_rl.models.generation.vllm.captured_chat_logprobs._ChatLogprobsBuilder#

Bases: typing.Protocol

_create_chat_logprobs(
*args: Any,
**kwargs: Any,
) → Any#
class nemo_rl.models.generation.vllm.captured_chat_logprobs.CapturedChatLogprobsMixin#

Scope a formatting-only bypass to one serving instance and request task.

Place before vLLM’s OpenAIServingChat in the MRO. Engine sampling parameters and the raw-array attachment/validation path remain unchanged.

_bypass_chat_logprobs(
*,
enabled: bool,
) → collections.abc.Iterator[None]#

Enable only around full-response formatting for admitted raw capture.

_create_chat_logprobs(
*args: Any,
**kwargs: Any,
) → Any#

Delegate normally, or omit objects that capture will strip after PUT.