[akanyaani]

Fine-tuning on a laptop, and a clearer statement of which backbones are supported.

• Device choice. DECISIONFORMER_DEVICE (cuda, mps, cpu) when set; otherwise CUDA, then Apple silicon (MPS) for single-process runs, then CPU. Loading a model for inference follows the same order. MPS and CPU train in float32. Tests pin the CPU.
• Training prompts that don't fit have the end of their state truncated by default (data.overflow: truncate_state_end), so one long row no longer stops a run. Loaded and served models still refuse rather than truncate, as the Decision Index requires.
• New recipe configs/recipes/local_smollm2.yaml: SmolLM2-135M-Instruct on deciformer-15k. It uses micro-batches of 4, a 4,096-token budget per micro-batch, gradient checkpointing, and checkpoints every 100 steps.
• README:
  • Backbones are the architectures built into transformers; models that need trust_remote_code are not supported.
  • The backbone table is ordered dense, mixture of experts, hybrid, then encoders.
  • New section: "On a laptop (Mac or Linux)".
• The training log names the device and precision.

Testing

All on Linux (Python 3.12, CPU only, GPU hidden):

• Full suite: 265 passed, 12 skipped.
• Laptop recipe forced onto the CPU (DECISIONFORMER_DEVICE=cpu): trained, evaluated, exported, reloaded and answered a request. Peak memory 3.7 GB, about 15 s per step on 6 cores.
• MPS device selection is covered by a unit test with a mocked device; no run on Apple hardware.