maxent_grpo.training.rollout.vllm_adapter

vLLM-focused helpers split away from the local generation path.

Functions

_broadcast_object_list(accelerator, payload, *)	Broadcast python objects even when Accelerate lacks the helper.
_env_flag(name[, default])
_env_int(name)
_gather_object_list(accelerator, value)	Gather Python lists across ranks with graceful Accelerate fallbacks.
_import_vllm_client_cls([import_fn])	Return TRL's VLLMClient using the provided import fn (defaults to optional_import).
_is_peft_model_safe(target)	Return True if accelerate.utils reports that the model uses PEFT adapters.
_loopback_host(base_url)
_optional_import(module_name)	Import a module if available without triggering import errors.
_progress_log_enabled()
_resolve_vllm_group_port()
_scatter_object(accelerator, input_list, *)	Scatter python objects from src to all ranks.
_temporary_env(overrides)
_use_vllm_collective()
_vllm_client_nccl_overrides(base_url)
_zero3_gather_factory(accelerator)
broadcast_object_list(accelerator, payload, *)	Public alias for broadcasting Python objects across ranks.
gather_object_list(accelerator, value)	Public alias for gathering Python objects across ranks.
scatter_object(accelerator, input_list, *[, src])	Public alias for scattering Python objects across ranks.

Classes

VLLMGenerationMixin(ctx)

All vLLM-specific plumbing extracted from the main generator.

class maxent_grpo.training.rollout.vllm_adapter.VLLMGenerationMixin(ctx)

Bases: object

All vLLM-specific plumbing extracted from the main generator.

Parameters:

ctx (GenerationContext)

ctx: GenerationContext

generate(prompts, num_samples, per_prompt_counts=None)

Produce completions, preferring vLLM when configured.

Parameters:

• prompts (List[str])

• num_samples (int)

• per_prompt_counts (List[int] | None)

Return type:

Tuple[List[List[str]], List[List[VLLMLogprobResult | None]] | None]

maxent_grpo.training.rollout.vllm_adapter.broadcast_object_list(accelerator, payload, *, src=0)

Public alias for broadcasting Python objects across ranks.

Parameters:

• accelerator (Any)

• payload (List[Any])

• src (int)

Return type:

None

maxent_grpo.training.rollout.vllm_adapter.gather_object_list(accelerator, value)

Public alias for gathering Python objects across ranks.

Parameters:

• accelerator (Any)

• value (List[Any])

Return type:

List[List[Any]]

maxent_grpo.training.rollout.vllm_adapter.scatter_object(accelerator, input_list, *, src=0)

Public alias for scattering Python objects across ranks.

Parameters:

• accelerator (Any)

• input_list (List[Any] | None)

• src (int)

Return type:

Any