MLX engine for training + gradio (Apple Silicon) - #10

Merged

Cortexelus merged 19 commits into

mps-supportfrom

mlx-engine

Jul 16, 2026

Commits on Jul 15, 2026

Add MLX engine selector (torch|mlx) for training + gradio
“Cortexelus”
committed
5a536c4
mlx-engine: transform underfit demo block to the MLX trainer's --demo-config list
“Cortexelus”
committed
50b0122
mlx-engine: map optimizer/scheduler/effective-length + adapter_type default
“Cortexelus”
committed
21eb9ba
mlx-engine: --dit-weights is optional (trainer auto-downloads the base npz)
“Cortexelus”
committed
e92f973
dashboard: platform-aware device UI (Apple Silicon MPS/MLX, not just CUDA)
“Cortexelus”
committed
68558cf
dashboard: engine-aware base-model gate + shorten MPS label
“Cortexelus”
committed
90628e8
dashboard: MLX dataset encoding on Apple (torch-free)
“Cortexelus”
committed
4492c40
dashboard: MLX encode writes into latent_dir, not output_dir
“Cortexelus”
committed
25a37fd
dashboard: pass exclude-file to the MLX encoder
“Cortexelus”
committed
08f90f4
mlx-engine: pass ARC demos, forward grad-clip, inherit trainer stdout
“Cortexelus”
committed
c1669ee
dashboard: fix Apple/MLX gaps in gradio launch, device panel, resume, model gate
“Cortexelus”
committed
b7f6dfd
dashboard/installer: on-demand + install-time MLX model download
“Cortexelus”
committed
f8e606f
mlx-engine: pass demo `duration` through to the MLX trainer
“Cortexelus”
committed
08c0003
dashboard: don't offer MPS without a torch backend; fix greyed-silent gate
“Cortexelus”
committed
ecc15e1
dashboard: pass base_model to the MLX gradio launch (--pretrained-name)
“Cortexelus”
committed
fba6445
dashboard: default MLX demos to ARC-only (skip the slow RF demos)
“Cortexelus”
committed
0549280
dashboard: SA3-medium "Use SAME-S for demos (faster)" toggle
“Cortexelus”
committed
d8a6a91
dashboard: relabel demo presets One/Two/Four on MLX (ARC-only counts)
“Cortexelus”
committed
899dfc8
dashboard: hide dotfile scratch from the checkpoint lists
“Cortexelus”
committed
de117ea

Navigation Menu

Uh oh!

MLX engine for training + gradio (Apple Silicon) - #10

MLX engine for training + gradio (Apple Silicon)#10
Cortexelus merged 19 commits into
mps-supportfrom
mlx-engine

Commits on Jul 15, 2026

Add MLX engine selector (torch|mlx) for training + gradio

mlx-engine: transform underfit demo block to the MLX trainer's --demo-config list

mlx-engine: map optimizer/scheduler/effective-length + adapter_type default

mlx-engine: --dit-weights is optional (trainer auto-downloads the base npz)

dashboard: platform-aware device UI (Apple Silicon MPS/MLX, not just CUDA)

dashboard: engine-aware base-model gate + shorten MPS label

dashboard: MLX dataset encoding on Apple (torch-free)

dashboard: MLX encode writes into latent_dir, not output_dir

dashboard: pass exclude-file to the MLX encoder

mlx-engine: pass ARC demos, forward grad-clip, inherit trainer stdout

dashboard: fix Apple/MLX gaps in gradio launch, device panel, resume, model gate

dashboard/installer: on-demand + install-time MLX model download

mlx-engine: pass demo `duration` through to the MLX trainer

dashboard: don't offer MPS without a torch backend; fix greyed-silent gate

dashboard: pass base_model to the MLX gradio launch (--pretrained-name)

dashboard: default MLX demos to ARC-only (skip the slow RF demos)

dashboard: SA3-medium "Use SAME-S for demos (faster)" toggle

dashboard: relabel demo presets One/Two/Four on MLX (ARC-only counts)

dashboard: hide dotfile scratch from the checkpoint lists