#5429 (cb15a7a5) tightened three production-path branches to
DEVICE_TYPE == "cuda" and torch.cuda.is_available() and added a
new else: SUPPORTS_BFLOAT16 = False arm to let `import unsloth.trainer`
survive on a CPU-only CI host. We already ship the package on Intel
XPU / AMD HIP / NVIDIA CUDA and don't want any extra branching in
those hot paths.
Move the entire CPU-CI handling to one place -- the top of
unsloth/device_type.get_device_type() -- so the UNSLOTH_ALLOW_CPU=1
sentinel short-circuits detection and returns "cuda" before any
torch probe runs. Every downstream DEVICE_TYPE == "cuda" branch
then behaves identically to a real CUDA host, with no additional
checks. The two existing duplicate UNSLOTH_ALLOW_CPU returns
later in the function are dropped (the new top-of-function check
covers both).
Revert the three call-site changes:
- unsloth/_gpu_init.py:212 -> back to `if DEVICE_TYPE == "cuda":`
- unsloth/_gpu_init.py:247 -> back to `if DEVICE_TYPE == "cuda":`
- unsloth/models/_utils.py:1207 -> back to `if DEVICE_TYPE == "cuda":`
- unsloth/_gpu_init.py: drop the new `else: SUPPORTS_BFLOAT16 = False`
branch (dead under the top-of-function short-circuit).
Keep the two env-var gates that are needed for zoo's drift detectors
to inspect pristine TRL source (no behavioural change on production
hosts that never set UNSLOTH_ALLOW_CPU):
- unsloth/_gpu_init.py: `if env != "1": _patch_trl_trainer()`
- unsloth/models/rl.py:PatchFastRL: `if env == "1": return`
Verified:
- CUDA_VISIBLE_DEVICES=5 python -c "import unsloth.trainer" produces
UnslothSFTTrainer.__init__ (TRL still patched on real hosts).
- UNSLOTH_ALLOW_CPU=1 + aggressive cuda spoof import succeeds and
trl.SFTTrainer.__init__.__qualname__ stays SFTTrainer.__init__.
- pytest tests/_zoo_compiler_cache_shim.py -> 5 passed, 1 skipped.
77e0929735