Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Studio: round the MLX warmup and eval ratios up like transformers (#16827)
Vineeth Sai Varikuntlad90dae6c1b75
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.pyindex c88a21c8e15..93b501df1e1 100644--- a/studio/backend/core/training/worker.py+++ b/studio/backend/core/training/worker.py@@ -2913,9 +2913,9 @@ def _configure_mlx_training_schedule()trainer.args.max_steps = max_stepsif warmup_ratio is not None:- trainer.args.warmup_steps = int(round(warmup_ratio * max_steps))+ trainer.args.warmup_steps = math.ceil(warmup_ratio * max_steps)if eval_steps_ratio is not None:- trainer.args.eval_steps = max(1, int(eval_steps_ratio * max_steps))+ trainer.args.eval_steps = max(1, math.ceil(eval_steps_ratio * max_steps))return max_steps, int(trainer.args.eval_steps)diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.pyindex 30fc8512b02..36e6aacbfbf 100644--- a/studio/backend/tests/test_mlx_training_worker_config.py+++ b/studio/backend/tests/test_mlx_training_worker_config.py@@ -591,6 +591,23 @@ def test_mlx_epoch_steps_divide_by_world_size_only_when_derived():assert steps(7, 16, 2, 2, 3, 2) == 7+# Expected = transformers 4.57.6 get_warmup_steps / TrainerState.compute_steps.+ "ratio, max_steps, expected",+ [(0.05, 10, 1), (0.03, 15, 1), (0.1, 25, 3), (0.1, 45, 5), (0.15, 30, 5), (0.29, 100, 29)],+)+def test_mlx_ratio_schedule_rounds_up_like_transformers(ratio, max_steps, expected):+ trainer = types.SimpleNamespace(+ distributed_world_size = 1,+ args = types.SimpleNamespace(max_steps = 0, warmup_steps = 0, eval_steps = 0),+ )+ _worker._configure_mlx_training_schedule(+ trainer, max_steps, None, 1, 1, 1, warmup_ratio = ratio, eval_steps_ratio = ratio+ )+ assert trainer.args.warmup_steps == expected+ assert trainer.args.eval_steps == expected++@pytest.mark.parametrize("rank", [0, 1])def test_mlx_final_save_is_rank0_owned(rank):trainer = _FakeRankTrainer(rank = rank, world_size = 2)