[tivris]

Description

Fixes #1945.

Without device=, boot kept cfg.device at get_device() for an hf_model with no hf_device_map, but it never moves that model. When the weights were on another device, calls that move their inputs to cfg.device failed, including bridge("text"), run_with_hooks on text and generate.

get_device() is now called only when boot loads the model itself. Without device=, an hf_model with no hf_device_map gets cfg.device from the existing fallback that reads its first parameter, as in build_bridge_from_module. An explicit device= and models with an hf_device_map are unchanged.

test_boot_device.py boots a tiny from_config GPT-NeoX built on the meta device through hf_model=, with get_device patched to return "cuda", and checks that cfg.device is "meta", which neither get_device() nor a CPU default would give. It fails on dev (cfg.device is "cuda") and passes with the fix. On MPS, the issue's example and its CPU variant with TRANSFORMERLENS_ALLOW_MPS=1 now run.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Looks good! I'll merge once CI passes

[jlarson4]

Update: there is a bug with the Qwen3.5 testing model we use for CI that is preventing Full Code Coverage from passing, I will patch dev and push the patch to your branch when it is verified.