Commits

History for

Commits on Sep 9, 2026

fix(mlx): honour training.max_grad_norm instead of dropping it silently (#750)
Srinivasan8888
authored
c7c6314
feat(mlx): report validation loss from on_val_loss_report (#739)
Srinivasan8888
authored
6b999eb
fix(mlx): wire warmup, scheduler, weight decay and optimizer choice (#734)
Srinivasan8888
authored
765f5b4

Commits on Sep 8, 2026

fix(mlx): honour train_on_responses_only with a real per-token mask (#733)
Srinivasan8888
authored
f71aef0

Commits on Sep 7, 2026

Commits on Sep 6, 2026

fix(trainer): support lora.r=0 as full fine-tuning on task=embedding (#705)

kok-o
and
claude
authored
9b48deb
fix(mlx): wire gradient_accumulation_steps into the MLX SFT TrainingArgs (#696)

AmirF194
and
claude
authored
3365f89
fix(distill): apply the response-only label mask and causal shift to the ULD loss (#711)

AmirF194
and
claude
authored
c5c0f32
fix(distill): refuse wasserstein and topk_align on tokenizers that disagree (#704)

AmirF194
and
claude
authored
a1eab0f
fix(trainer): put SFT packing on SFTConfig for TRL 0.29 (#709)

jagadeepmamidi
and
claude
authored
991ab8f
fix(trainer): expose model and args on the embedding trainer wrapper (#708)

jagadeepmamidi
and
claude
authored
ab2412e
fix(bench): measure the GEMM ceiling in the card's resolved stream dtype (#648)

Samearth17
and
claude
authored
f14bf42

Commits on Sep 5, 2026

feat(mlx): drive the live display, tracker and SSE stream from the MLX trainer (#665)
Srinivasan8888
authored
87b479c

Commits on Sep 3, 2026

Commits on Sep 1, 2026

Commits on Aug 28, 2026

Commits on Aug 27, 2026

Commits on Aug 26, 2026

fix(train): enable BF16 on validated MPS trainers (#564)
Amix29
authored
4b2947b

Commits on Aug 25, 2026

Commits on Aug 24, 2026

Commits on Aug 23, 2026