maxent_grpo.training.scoring_common

Scoring helpers extracted from the MaxEnt-GRPO training loop.

Functions

_as_context_manager(value)	Return value as a context manager when possible, otherwise a no-op.
_autocast_context(accelerator, device)	Return the right autocast context for the current accelerator/device.
_coerce_optional_int(value)	Return value coerced to int when possible, else None.
_coerce_shape(value)
_describe_embedding_module(module, name)	Return a human-friendly summary of an embedding module.
_dist_all(dist, flag)	Return True if flag is True on all ranks (best-effort).
_dist_any(dist, flag)	Return True if flag is True on any rank (best-effort).
_dist_collective_ready(torch_mod)	Return a dist module when initialized, otherwise None.
_get_config_value(config, key[, default])	Return a config value from either Mapping or object-style configs.
_get_embedding_vocab_size(model, config)	Return the vocab size exposed by the model's embedding weights.
_maybe_long_tensor(value, torch_mod)	Return a tensor cast to long when the stub lacks long.
_model_has_non2d_embeddings(model)	Return True when any known embedding weight is not 2-D.
_prefetch_iterator(iterator, buffer_size)	Yield from iterator while prefetching up to buffer_size slices.
_progress_log_enabled()
_refresh_torch()	Return the active torch module.
_resolve_dtype(dtype)	Normalize dtype objects coming from various stubs.
_score_slice_log_enabled()
_size_hint(tensor_obj, dim)	Return tensor.size(dim) with fallbacks for numpy-backed stubs.
_to_numpy_array(obj)	Return a numpy view of obj for stub compatibility.
_weight_is_stub_tensor(weight)	Return True for tensor-like stubs used in tests.
_weight_is_two_dimensional(weight)	Return True if the provided weight exposes a 2-D shape.

Classes

_DistModuleLike(*args, **kwargs)	Minimal distributed API needed for best-effort gathers.
_LongDTypeProxy(target)	Lightweight wrapper that compares equal to torch.long in stubs.
_PadTokenGuard(targets, value)	Context manager that temporarily clamps padding attributes.
_TorchModuleLike(*args, **kwargs)