| import torch |
| import torch.nn.functional as F |
| import torch.distributed as dist |
| from typing import Tuple, Any |
|
|
| def solution( |
| hidden_states: torch.Tensor, |
| weight: torch.Tensor, |
| labels: torch.Tensor, |
| old_logprobs: torch.Tensor, |
| advantages: torch.Tensor, |
| ignore_index: int = -100, |
| ) -> Tuple[torch.Tensor, Any, torch.Tensor, torch.Tensor, torch.Tensor]: |
| logits = F.linear(hidden_states, weight) |
| logits_flat = logits.view(-1, logits.size(-1)) |
| labels_flat = labels.view(-1) |
| |
| per_token_ce = F.cross_entropy(logits_flat, labels_flat, ignore_index=ignore_index, reduction='none') |
| |
| new_logprobs_flat = -per_token_ce.detach() |
| old_logprobs_flat = old_logprobs.view(-1) |
| advantages_flat = advantages.view(-1) |
| |
| valid_mask = (labels_flat != ignore_index) |
| n_valid_local = valid_mask.sum().float() |
| |
| n_valid_global = n_valid_local.clone() |
| dist.all_reduce(n_valid_global, op=dist.ReduceOp.SUM) |
| n_valid_global_clamped = n_valid_global.clamp(min=1.0) |
| |
| delta = (new_logprobs_flat - old_logprobs_flat).masked_fill(~valid_mask, 0.0).clamp(min=-20.0, max=20.0) |
| ratio = torch.exp(delta) |
| |
| per_token_pg = -(ratio * advantages_flat).masked_fill(~valid_mask, 0.0) |
| local_pg_sum = per_token_pg.sum() |
| |
| global_pg_sum = local_pg_sum.clone() |
| dist.all_reduce(global_pg_sum, op=dist.ReduceOp.SUM) |
| true_pg = global_pg_sum / n_valid_global_clamped |
| |
| w = (ratio.detach() * advantages_flat).masked_fill(~valid_mask, 0.0) |
| local_surrogate_sum = (w * per_token_ce).sum() |
| |
| surrogate = local_surrogate_sum / n_valid_global_clamped |
| |
| loss = true_pg.detach() + surrogate - surrogate.detach() |
| |
| ratio_valid = ratio.masked_fill(~valid_mask, 0.0) |
| sum_ratio_local = ratio_valid.sum() |
| dist.all_reduce(sum_ratio_local, op=dist.ReduceOp.SUM) |
| ratio_mean = sum_ratio_local / n_valid_global_clamped |
| |
| ratio_for_min = ratio.masked_fill(~valid_mask, float('inf')) |
| min_ratio_local = ratio_for_min.min() if n_valid_local > 0 else torch.tensor(float('inf'), device=ratio.device) |
| dist.all_reduce(min_ratio_local, op=dist.ReduceOp.MIN) |
| |
| ratio_for_max = ratio.masked_fill(~valid_mask, float('-inf')) |
| max_ratio_local = ratio_for_max.max() if n_valid_local > 0 else torch.tensor(float('-inf'), device=ratio.device) |
| dist.all_reduce(max_ratio_local, op=dist.ReduceOp.MAX) |
| |
| k3_local = (ratio - delta - 1.0).masked_fill(~valid_mask, 0.0).sum() |
| dist.all_reduce(k3_local, op=dist.ReduceOp.SUM) |
| k3_mean = k3_local / n_valid_global_clamped |
| |
| entropy_local = per_token_ce.detach().masked_fill(~valid_mask, 0.0).sum() |
| dist.all_reduce(entropy_local, op=dist.ReduceOp.SUM) |
| entropy_mean = entropy_local / n_valid_global_clamped |
| |
| metrics = torch.stack([ratio_mean, min_ratio_local, max_ratio_local, k3_mean, entropy_mean]) |
| |
| per_token_logprobs = new_logprobs_flat.view_as(labels) |
| per_token_loss = per_token_pg.view_as(labels) |
| |
| return loss, None, per_token_logprobs, per_token_loss, metrics |
|
|
