zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-06-24 09:21:18 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

match pytoch imeplentation for loss calculation

Browse Source
This commit is contained in:
Goekdeniz-Guelmez 2025-03-11 09:00:21 +01:00
parent f1961f1b79
commit 06ff47012f
1 changed files with 2 additions and 4 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

@ -381,9 +381,7 @@ def grpo_loss(
) )
# Average over tokens # Average over tokens
sequence_sums = per_token_loss.sum(axis=1) loss = (per_token_loss * length_mask).sum() / length_mask.sum() # Matches the pytorch implementaiton
sequence_lengths = length_mask.sum(axis=1)
loss = (sequence_sums / sequence_lengths).mean()
# Calculate mean KL divergence for metrics # Calculate mean KL divergence for metrics
mean_kl = ((kl_div * length_mask).sum(axis=1) / length_mask.sum(axis=1)).mean() mean_kl = ((kl_div * length_mask).sum(axis=1) / length_mask.sum(axis=1)).mean()
@ -454,7 +452,7 @@ def grpo_loss(
mx.metal.clear_cache() mx.metal.clear_cache()
return loss, sequence_lengths.sum(), metrics return loss, length_mask.sum(axis=1).sum(), metrics
def iterate_grpo_batches(dataset, batch_size, max_seq_length, train=False): def iterate_grpo_batches(dataset, batch_size, max_seq_length, train=False):