zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-07-19 09:31:13 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

rebase loss calculation

Browse Source
This commit is contained in:
Goekdeniz-Guelmez 2025-02-09 17:13:05 +01:00
parent a527cdb39b
commit 00712522ba
1 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
llms/mlx_lm/tuner/grpo_trainer.py
View File

@ -317,7 +317,9 @@ def grpo_loss(
per_token_loss = -((policy_ratio * advantages.reshape(-1, 1) - beta * kl_div) * length_mask)
# Average over tokens
loss = per_token_loss.sum().mean()
sequence_sums = per_token_loss.sum(axis=1)
sequence_lengths = length_mask.sum(axis=1)
loss = (sequence_sums / sequence_lengths).mean()
# Calculate mean KL divergence for metrics
mean_kl = ((kl_div * length_mask).sum(axis=1) / length_mask.sum(axis=1)).mean()
@ -343,7 +345,7 @@ def grpo_loss(
}
mx.metal.clear_cache()
return loss, length_mask.sum(axis=1).sum(), metrics
return loss, sequence_lengths.sum(), metrics
def iterate_grpo_batches(dataset, tokenizer, batch_size, max_seq_length, train=False):