4. Optimize training performance: torch.compile, AMP, persistent workers

[musicalplatypus]

Summary

Adds performance optimization flags to the training pipeline, improving training throughput on both GPU and CPU.

Changes

1. torch.compile support — Wraps models with torch.compile() when available (PyTorch 2.0+), enabling fused kernels and graph optimizations
2. Native AMP (Automatic Mixed Precision) — Adds --native-amp flag for bfloat16/float16 autocast with GradScaler on CUDA
3. Persistent DataLoader workers — Enables persistent_workers=True to avoid worker process restart overhead between epochs
4. Eval efficiency — Streamlined evaluation loops to reduce redundant computation
5. Config pipeline threading — compile_model and native_amp flags threaded through the modelmaker YAML config → params → training argv pipeline

Files Changed (9 files)

• tinyml-tinyverse/tinyml_tinyverse/common/utils/utils.py — core optimizations
• tinyml-tinyverse/tinyml_tinyverse/references/common/train_base.py — shared training base
• tinyml-tinyverse/tinyml_tinyverse/references/*/train.py — per-task integration
• tinyml-modelmaker/.../timeseries/params.py — config parameter definitions
• tinyml-modelmaker/.../timeseries_base.py — argv construction

Testing

• Benchmarked on classification and forecasting tasks
• torch.compile provides ~15-30% speedup on supported hardware
• No regression when flags are disabled (default behavior unchanged)

[musicalplatypus]

Re: CI failure — The test failures (ModuleNotFoundError: No module named 'tinyml_tinyverse') are pre-existing and not introduced by this PR.

Checking out upstream/main directly produces the identical error — the Tier 1 tests fail at collection time before any of this PR's code runs. The pip install -e tinyml-tinyverse step in the workflow doesn't make the package importable in the test runner's environment. Happy to help investigate the CI setup if useful.

[musicalplatypus]

Re: CI failures — The test failures on this PR are pre-existing and not introduced by these changes. The root cause is a CI configuration bug tracked in PR #17: `pip install -e tinyml-tinyverse` silently fails during the install step (due to a `ti_mcu_nnc` version mismatch and wrong install order), so `tinyml_tinyverse` is never installed, and the Tier 1 tests fail at import time before any code in this PR runs. Checking out `upstream/main` directly produces the identical `ModuleNotFoundError`.

PR #17 fixes the CI setup. Once merged into `main`, the workflow will be correct and CI should pass for all open PRs.

Also: pushed a fix for `output.squeeze()` → `output.squeeze(-1)` in `evaluate_classification`. The bare `squeeze()` would collapse the batch dimension when the last evaluation batch contains exactly one sample (any dataset whose size isn't a multiple of `batch_size`), producing wrong shapes for loss and accuracy. `squeeze(-1)` only removes the class dimension for binary models that output `[N, 1]`.