fix: TG013 downloads and device restores, and preprocessors are not models

[HimanshuJanbandhu]

Finishes the triage. TG013 was the one rule left without a verdict, and TG002 had two causes still standing.

TG013 — 3 false of 10, two narrow causes

A download is not a redundant upload. _device_argument accepted a "cpu" literal as a destination, so this was reported:

tensor = torch.arange(1, 1_000_000, device="cuda")...
for i in range(100):
    cpu_tensor = tensor.to("cpu", non_blocking=True)

That's pinmem_nonblock.py — a tutorial whose entire subject is measuring transfer behaviour, copying to the host, on a tensor already on the device. Wrong twice over.

Restoring the device after a deliberate .cpu() is required. fast_neural_style does:

transformer.eval().cpu()
torch.save(transformer.state_dict(), ckpt_model_path)
transformer.to(device).train()      # we said: hoist this out of the loop

Hoisting it would leave the model on the host for the rest of training.

The other 7 are true — the mild device= advice the rule exists to give (a host factory or torch.tensor(python_list) inside a loop), plus one genuinely hoistable constant, self.STOP_TOKENS_TENSOR.to(self._device) in torchtune.

TG002 — from_pretrained made a tokenizer a model

from_pretrained sits in MODEL_WRAPPERS and matched before anything could object, so AutoTokenizer.from_pretrained(...) was a model and this was a forward pass:

dataset.map(lambda x: tokenizer(x["question"]), batched=True)

We reported a missing no_grad around tokenisation. Feature extractors and image processors the same. The exclusion is by class name and runs first; a guard test asserts AutoModelForCausalLM.from_pretrained still counts as a model.

Result

	errors	warnings	notes
before	13	49	207
after	11	46	207

Five removed, zero new findings.

Where the triage ends up

Every rule now has a verdict. Remaining: TG004 207 (verified accurate, notes), TG008 31, TG003 3 (true, fixed upstream in pytorch/examples#1424), TG013 7, TG001 10, TG002 6.

Two items are left as decisions rather than fixes, recorded in the TODO:

• TG008 (31) — factually right, but mostly synthetic-data example scripts where seeding changes nothing observable. Same profile that moved TG004 to note under RFC 0003, so it deserves the same test: is this code defective, or merely untuned?
• Two TG001 causes with one finding each — the label naming heuristic on non-tensor lists, and container element types. One finding is not yet a pattern.

Testing

447 tests, 4 new — two per fix, each including the guard for the case the fix must not silence.

Stacks on #49 and #50.