[nathanhubens]

Summary

Six measurement paths moved the caller's model to the measuring device in place and never moved it back, so a CUDA model came out of a CPU profiling call parked on the CPU, and any later timing of it was silently a CPU timing. Every measurement now leaves the model on the device it came in on.

exported line (master)	function	public entry points affected
speed.py:72	_forward_latencies	compute_speed, compute_speed_multi, sweep_threads, sweep_batch_sizes, sweep_latency
memory.py:53, :91	_gpu_metrics, _cpu_metrics	compute_memory, compute_memory_multi
profiling.py:181	_profile_layers	LayerProfiler.profile
roofline.py:307	RooflineAnalyzer.profile
energy.py:98	compute_energy	compute_energy_multi

Checked and left alone: compute_size (reads only), compute_compute (moves the sample to the model's device), benchmark() (delegates, restores the train/eval flag). One private context manager, _on_device(model, dev) in fasterbench.core, yields the model on dev and restores the original device in a finally (a model with no parameters or buffers stays put). Each site becomes with _on_device(model.eval(), dev) as model: and its body is re-indented, nothing else: no metric, signature or message changes. In speed, energy and roofline the new with stays nested inside the device context because a check (_ensure_device_supported) or a peak measurement runs between the two and must see the model before it moves.

Also in this commit: two prose mentions of a private package in nbs/analysis/roofline.ipynb and nbs/tutorials/roofline.ipynb are reworded; this library measures, nothing here decides.

Red → green

The seven new #| notest cells (a tiny conv net on CUDA, measured with device='cpu', every parameter asserted back on CUDA) run against origin/master:

compute_speed(device='cpu')                before=['cuda:0'] after=['cpu']  FAIL
compute_memory(device='cpu')               before=['cuda:0'] after=['cpu']  FAIL
_profile_layers(device='cpu')              before=['cuda:0'] after=['cpu']  FAIL
RooflineAnalyzer.profile(device='cpu')     before=['cuda:0'] after=['cpu']  FAIL

and on this branch:

compute_speed(device='cpu')                before=['cuda:0'] after=['cuda:0']  OK
compute_memory(device='cpu')               before=['cuda:0'] after=['cuda:0']  OK
_profile_layers(device='cpu')              before=['cuda:0'] after=['cuda:0']  OK
RooflineAnalyzer.profile(device='cpu')     before=['cuda:0'] after=['cuda:0']  OK

CI is CPU-only, so it covers the CPU→CPU contract only; the CUDA restore, which is the bug, is proven by the #| notest cells run locally with flags='notest' (all seven notebooks Success.). compute_energy has no CI coverage at all (its test would run a codecarbon tracker), stated rather than hidden. Full default suite, 17 notebooks:

- Completed .../nbs/visualization/plot.ipynb
Success.
EXIT=0

Export + clean idempotent, _modidx.py gains the one new entry, __version__ stays 0.1.0.

Second commit: the tree in the shape CI's nbdev expects

CI (fastai/workflows/nbdev3-ci, nbdev 3.3.15) last ran on this repo in July and would be red on master today, independently of this fix: nbdev-clean reserializes 12 notebooks (single-string source to a list of lines, two markdown cells shedding keys the nbformat schema forbids), and nbdev-export regenerates the __init__.py docstring from every index.ipynb cell carrying an export directive. The chore commit reserializes the notebooks (content-neutral: joined source and outputs identical per cell, checked by script), stops index.ipynb from exporting __init__ (its cells stay as plain executed cells), and makes fasterbench/__init__.py hand-maintained as in fasterai, byte-identical in its version, imports and __all__. After it, CI's clean + export leave git status --porcelain -uno empty on two passes, and so does the dev environment's older nbdev.

https://claude.ai/code/session_0177tJZDQTTjQdycXV1EZtdJ