zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-09-01 04:14:38 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

reduction moved to CPU in case of distributed training (#1200)

Browse Source
This commit is contained in:
Ivan Fioravanti
2025-01-15 02:20:42 +01:00
committed by GitHub
parent c117af83b8
commit 6ae6c72c2e
2 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
llms/mlx_lm/tuner/trainer.py
View File

@@ -159,8 +159,8 @@ def evaluate(
ntokens += toks
mx.eval(all_losses, ntokens)
all_losses = mx.distributed.all_sum(all_losses)
ntokens = mx.distributed.all_sum(ntokens)
all_losses = mx.distributed.all_sum(all_losses, stream=mx.cpu)
ntokens = mx.distributed.all_sum(ntokens, stream=mx.cpu)
return (all_losses / ntokens).item()
@@ -272,9 +272,9 @@ def train(
if it % args.steps_per_report == 0 or it == args.iters:
stop = time.perf_counter()
train_loss = mx.distributed.all_sum(losses).item()
train_loss = mx.distributed.all_sum(losses, stream=mx.cpu).item()
train_loss /= steps * mx.distributed.init().size()
n_tokens = mx.distributed.all_sum(n_tokens).item()
n_tokens = mx.distributed.all_sum(n_tokens, stream=mx.cpu).item()
learning_rate = optimizer.learning_rate.item()
it_sec = args.steps_per_report / (stop - start)
tokens_sec = float(n_tokens) / (stop - start)