mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-07-29 14:30:40 +00:00
hotfix
This commit is contained in:
parent
e774edeb80
commit
5c5cb1863b
@ -371,20 +371,6 @@ class GRPOConsumer(BaseConsumer):
|
|||||||
kl = all_reduce_mean(torch.mean(torch.stack(kl)).to(loss.device), self.plugin).data
|
kl = all_reduce_mean(torch.mean(torch.stack(kl)).to(loss.device), self.plugin).data
|
||||||
mean_kl.append(kl)
|
mean_kl.append(kl)
|
||||||
mean_loss.append(all_reduce_mean(loss, self.plugin).data)
|
mean_loss.append(all_reduce_mean(loss, self.plugin).data)
|
||||||
mini_batch_entropies.append(
|
|
||||||
all_reduce_mean(
|
|
||||||
(
|
|
||||||
(
|
|
||||||
(
|
|
||||||
entropy_from_logits(policy_model_logits[:, -num_action:])
|
|
||||||
* action_mask_forward_micro_batch
|
|
||||||
).sum(-1)
|
|
||||||
)
|
|
||||||
/ action_mask_forward_micro_batch.sum(-1)
|
|
||||||
).detach(),
|
|
||||||
self.plugin,
|
|
||||||
)
|
|
||||||
)
|
|
||||||
else:
|
else:
|
||||||
policy_model_logits = self.policy_model(
|
policy_model_logits = self.policy_model(
|
||||||
input_ids=input_ids_forward_micro_batch,
|
input_ids=input_ids_forward_micro_batch,
|
||||||
|
Loading…
Reference in New Issue
Block a user