Update CUDA/ROCm setup tests

[sstamenk]

Updated backend-specific setup handling and tests for CUDA/ROCm library selection. The change makes get_cuda_bnb_library_path() fail fast when the wrong override variable is used (BNB_ROCM_VERSION on CUDA or BNB_CUDA_VERSION on ROCm), improves the warning/error guidance, and refreshes the evaluator tests to use BNB_BACKEND with clearer CUDA/ROCm coverage.
Removed the test_get_rocm_bnb_library_path_rocm_override_takes_priority unit-test as it doesn't provide much utility.

This PR is a subset of the changes originally proposed in #1888

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[matthewdouglas]

Review: Update CUDA/ROCm setup tests

(Note: this review was created with the assistance of an LLM.)

Thanks for the PR! The fail-fast behavior for wrong override variables and the HIP_ENVIRONMENT → BNB_BACKEND migration are good improvements. A couple of issues to address before merging:

1. Missing path assertion in test_get_rocm_bnb_library_path_override

The diff removes the functional assertion that validates the override actually produces the correct library path:

assert get_cuda_bnb_library_path(rocm70_spec).stem == "libbitsandbytes_rocm72"

The test now only checks assert "BNB_ROCM_VERSION" in caplog.text, which verifies the warning was emitted but not that the correct library was selected. This looks like an accidental omission — please restore it.

2. test_get_cuda_bnb_library_path_override is fragile

This test sets BNB_CUDA_VERSION but doesn't clear BNB_ROCM_VERSION. With the new logic, rocm_override_value is checked first — so if BNB_ROCM_VERSION happens to be set in the environment, the test will raise RuntimeError instead of exercising the CUDA override path. Please add:

monkeypatch.delenv("BNB_ROCM_VERSION", raising=False)

(The base test test_get_cuda_bnb_library_path already does this correctly.)

Minor notes

• The dead re.sub(r"rocm\d+", ...) call in the CUDA+BNB_ROCM_VERSION error path is harmless (regex doesn't match libbitsandbytes_cuda*) but could be avoided by moving the substitution after the torch.version.cuda check.
• Consider keeping an adapted version of test_get_rocm_bnb_library_path_rocm_override_takes_priority — the "both overrides set" edge case is still reachable and worth covering, especially now that the branching logic has changed. Not blocking though.

[sstamenk]

@matthewdouglas everything should be good now, here is an overview of the changes:

Changes addressing review feedback on get_cuda_bnb_library_path and its tests

bitsandbytes/cextension.py

Restructured override logic to branch on backend first. The previous if rocm_override_value / elif cuda_override_value structure branched on which env var was set, causing two issues:

• The re.sub ran before the wrong-backend error check (harmless but dead code on the wrong backend)
• When both BNB_ROCM_VERSION and BNB_CUDA_VERSION were set, behavior was asymmetric: error on CUDA, silent ignore on ROCm

The new structure dispatches on torch.version.hip / torch.version.cuda first, then handles overrides within each backend branch:

Scenario	Behavior
Only wrong variable set	RuntimeError (unchanged)
Both set	Applies the correct one, warns about the ignored wrong one (new)
Only correct variable set	Applies override with existing warning (unchanged)
Neither set	Uses detected version (unchanged)
Other backends (e.g. XPU)	Rejects both overrides with RuntimeError (new)

tests/test_cuda_setup_evaluator.py

• Fixed test_get_cuda_bnb_library_path_override — added monkeypatch.delenv("BNB_ROCM_VERSION", raising=False) so a stray env var doesn't cause the test to hit the wrong code path
• Mirrored in test_get_rocm_bnb_library_path_override — added monkeypatch.delenv("BNB_CUDA_VERSION", raising=False) for symmetry
• Added test_get_cuda_bnb_library_path_cuda_override_takes_priority — both overrides set on CUDA: CUDA wins, ROCm override warned about
• Added test_get_rocm_bnb_library_path_rocm_override_takes_priority — both overrides set on ROCm: ROCm wins, CUDA override warned about
• Updated error match strings — reflect new messages (not a ROCm/CUDA build instead of detected for CUDA/ROCm)

Tested

ROCm tests verified passing on torch 2.9.1+rocm7.2.0 (4 passed, 4 CUDA-only skipped).

[matthewdouglas]

Thanks!