fix: make TransformersRunnable work on CPU-only machines (#4376)

[jissen706]

What does this PR address?

Fixes #4376.

Instantiating a runnable for a torch pretrained transformers model (bentoml.transformers.save_model(name, pretrained_model) → bento_model.to_runnable()()) crashed in two independent ways:

1. CUDA default tensor type set on CPU-only machines (the bug reported in bug: TransformersRunnable.__init__ sets torch default tensor type to torch.cuda.FloatTensor even if gpu is not available #4376): TransformersRunnable.__init__ called torch.set_default_tensor_type("torch.cuda.FloatTensor") unconditionally for torch-framework pretrained models — even when CUDA_VISIBLE_DEVICES was unset/-1 and the model had just been placed on CPU. On machines without CUDA this raises TypeError: type torch.cuda.FloatTensor not available. The call is now gated on the same GPU-assignment condition used for device placement, matching how common/pytorch.py and detectron.py gate the identical call on torch.cuda.is_available().

2. Stale generation methods in default signatures: make_default_signatures records generate, greedy_search, beam_search, etc. for every PreTrainedModel subclass, but on transformers >= 4.50 these methods moved to GenerationMixin (and the individual search methods were removed entirely), so runnable init crashed at getattr(self.model, method_name) for every torch pretrained model, GPU or not. Default signatures are now filtered to methods the class actually provides, and signature methods missing at load time (e.g. models saved with an older transformers version) are skipped with a warning instead of crashing.

Testing

Added two regression tests in tests/integration/frameworks/test_transformers_unit.py using a tiny offline BertForSequenceClassification built from config (same pattern as the existing custom-pipeline test — no network needed):

• test_pretrained_runnable_init_without_gpu: before this fix it fails with the exact TypeError from the issue; now it verifies the model lands on CPU and the default tensor type stays a CPU type.
• test_pretrained_runnable_skips_unavailable_signature_methods: saves a model with a stale greedy_search signature and verifies runnable init warns and skips it instead of crashing.

Verified locally on macOS (CPU-only, torch 2.13, transformers 4.57): both new tests pass, tests/unit passes (298 passed, 5 skipped), and the remaining failures in test_transformers_unit.py are pre-existing on main (transformers 4.57 API incompatibilities, unrelated to this change).

Before submitting:

• Does the Pull Request follow Conventional Commits specification naming?
• Does the code follow BentoML's code style, both make format and make lint script have passed (documentation)?
• Did your changes require updates to the documentation? Not needed — behavior fix only.
• Did you write tests to cover your changes?

🤖 Generated with Claude Code