[Enhancement] Add early device validation in pipeline() to prevent downloading/loading models before crashing on typos

[AryanSharma21]

Feature request

Feature Description

Currently, if a user misspells or passes an invalid device argument when initializing transformers.pipeline() (e.g., passing device="mpx" instead of "mps" or "cuda"), the function proceeds to download and load the full model weights into RAM before validating the device.

Validating the device parameter early at the pipeline() entry point would save bandwidth, RAM/disk overhead, and developer time.

Current Internal Execution Flow & Behavior

When executing:

from transformers import pipeline

# Downloads gigabytes of model weights first... and then crashes at tensor placement
classifier = pipeline("text-classification", model="bert-base-uncased", device="mpx")

Proposed Solution
Add lightweight early validation at the very top of transformers.pipeline() before from_pretrained() or model initialization is triggered.

If device is a string, validate it against canonical device prefixes (cpu, cuda, mps, npu, xpu, hpu), integers, or torch.device instances. If invalid, fail immediately:

ValueError: Invalid device 'mpx'. Expected a valid device string (e.g., 'cpu', 'cuda', 'mps', 'npu', 'xpu'), an integer device ordinal (e.g., 0), or a torch.device instance.

Motivation

It is most likely that any user would be annoyed to see an entire model download and fail at tensor placement.

Your contribution

I am happy to open PR adding this early check and unit tests if the maintainers agree with this enhancement.

will open PR for this

[AryanSharma21]

Working on it, requesting a review.

[AryanSharma21]

Hey, @ameerhmz, I explicitly mentioned in the issue comments yesterday that I was actively working on implementing this fix and opening a PR for it.

I think it is standard practice to check the issue comments to see if the issue creator or another contributor is already working on the fix before opening a duplicate PR.

Additionally, in your PR #47881, the _validate_device function was declared but never actually called inside pipeline(), leaving it as dead code that doesn't trigger early validation.

I have opened PR #47898 with the tested implementation where validation is correctly called at the entry point of pipeline() and verified with unit tests. I request the maintainers to please review PR #47898. Thanks