maxent_grpo.training.generation.vllm_distributed

Distributed helpers used by the vLLM generation helper.

Functions

_current_torch()

Return torch, preferring the vLLM module shim when patched in tests.

_gather_object_list(accelerator, value)

Gather python lists across ranks with Accelerate/torch fallbacks.

_scatter_object(accelerator, input_list, *)

Scatter python objects from src rank to every other process.

Classes

VLLMDistributedMixin()

Split out scatter/gather helpers from the vLLM helper.
class maxent_grpo.training.generation.vllm_distributed.VLLMDistributedMixin[source]

Bases: object

Split out scatter/gather helpers from the vLLM helper.

ctx: Any