maxent_grpo.training.eval

Validation helpers for the MaxEnt-GRPO training loop.

Functions

_as_tensor_1d(value[, device])	Best-effort conversion to a 1D float tensor.
_build_eval_shard(evaluation_rows, accelerator)	Return shard metadata describing which rows this rank evaluates.
_compute_eval_kl_tensor(cur_logp_sum, ...)
_compute_eval_rewards(completions, answers, ...)	Return aggregated reward scores for completions.
_compute_seed_eval_metrics(ctx)	Compute auxiliary seed-eval metrics from multi-sample generations.
_deepspeed_zero_stage(accelerator)	Return DeepSpeed ZeRO stage from Accelerate plugin state when present.
_env_flag(name[, default])
_eval_logprobs_enabled()
_eval_rank_tag()
_flatten_eval_meta(grouped_meta, expected_len)
_gather_eval_stats(accelerator, eval_scores)	Gather mean reward statistics across all ranks.
_init_eval_score_stats()
_iter_eval_batches(evaluation_rows, batch_size)	Yield prompt/answer lists for evaluation rows.
_log_eval_start(step, shard, batch_size)	Log the evaluation plan when running on the main rank.
_match_tensor_length(tensor, target_len[, ...])	Pad or truncate a 1D tensor to target_len.
_maybe_presync_vllm_for_eval(ctx, ...)	Synchronize vLLM weights across all ranks before rank-0-only eval.
_progress_log_enabled()
_run_eval_batches(shard, batch_size, ctx, step)	Generate completions for the shard rows and log periodic progress.
_score_eval_batch(ctx, prompts, completions, ...)
_seed_eval_config(evaluation_cfg)	Normalize optional seed-eval settings from EvaluationSettings.
_seed_eval_enabled(raw)	Return whether seed-eval metrics are enabled from a config value.
_tally_format_issues(completions)	Return format issue counts for a batch of completions.
_update_eval_score_stats(target, update)
_warn_eval_logprobs_unavailable(reason)
run_validation_step(step, ctx)	Generate single completions on the eval set and log mean reward.

Classes

_EvalShardInfo(rows, total_rows, ...)

Metadata describing the evaluation shard for the current rank.

maxent_grpo.training.eval.run_validation_step(step, ctx)

Generate single completions on the eval set and log mean reward.

Parameters:

• step (int) – Training step identifier passed to logging hooks.

• ctx (ValidationContext) – Validation context providing evaluation rows and handles.

Returns:

None. Logs metrics through the provided handles.

Return type:

None