Fix bitsandbytes zombie module breaking test collection on CPU runners (#7580)

* Fix bitsandbytes zombie module breaking test collection

A partially failed `import bitsandbytes` leaves the package half-imported:
CPython evicts only the parent from sys.modules and keeps every submodule it
had already loaded. The next import re-executes __init__ but every
`from .x import y` is served from cache, so the submodule attributes are never
rebound. The package imports "successfully" while `bnb.functional` is gone.

Bind the submodule via `import bitsandbytes.functional as bnb_functional`,
which reads sys.modules directly and survives that state, and import
bitsandbytes in tests/conftest.py on the real CPU path before
torch.cuda.is_available() is mocked, so the half-imported state is never
created in the first place.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

unsloth/_gpu_init.py

@@ -303,13 +303,19 @@ if DEVICE_TYPE == "cuda":
     # Try loading bitsandbytes and triton
     try:
         import bitsandbytes as bnb
+
+        # Bind the submodule by name: a half-imported bitsandbytes leaves the parent
+        # without a `functional` attribute, which would otherwise be misreported below
+        # as a CUDA linking failure. See unsloth/kernels/utils.py.
+        import bitsandbytes.functional as bnb_functional
     except:
         print(
             "Unsloth: `bitsandbytes` is not installed - 4bit QLoRA unallowed, but 16bit and full finetuning works!"
         )
         bnb = None
+        bnb_functional = None
     try:
-        cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
+        cdequantize_blockwise_fp32 = bnb_functional.lib.cdequantize_blockwise_fp32
         libcuda_dirs()
     except:
         if hasattr(os, "geteuid") and os.geteuid() == 0:
@@ -351,7 +357,7 @@ if DEVICE_TYPE == "cuda":
                         pass
                 else:
                     from triton.common.build import libcuda_dirs
-                cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
+                cdequantize_blockwise_fp32 = bnb_functional.lib.cdequantize_blockwise_fp32
                 libcuda_dirs()
             except:
                 warnings.warn(

unsloth/kernels/utils.py

@@ -136,11 +136,18 @@ def calculate_settings(
 HAS_CUDA_STREAM = False
 try:
     import bitsandbytes as bnb
+
+    # If an earlier `import bitsandbytes` died inside __init__, CPython evicts only
+    # the parent from sys.modules and keeps its submodules, so this retry re-executes
+    # __init__ without rebinding `bnb.functional`. `import x.y as z` reads sys.modules
+    # directly and survives that, plain attribute access does not.
+    import bitsandbytes.functional as bnb_functional
 except Exception:
     # device_type.py already degrades to 16bit/full finetuning when bnb is missing
     # (e.g. gfx906, whose generic wheel has no kernels). Keep the import working and
     # fail only if a 4bit path is actually entered.
     bnb = None
+    bnb_functional = None
 
 
 def _bnb_required(*args, **kwargs):
@@ -153,7 +160,7 @@ def _bnb_required(*args, **kwargs):
 if bnb is not None:
     # https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1330/files
     HAS_CUDA_STREAM = Version(bnb.__version__) > Version("0.43.3")
-    get_ptr = bnb.functional.get_ptr
+    get_ptr = bnb_functional.get_ptr
 else:
     get_ptr = _bnb_required
 
@@ -263,18 +270,18 @@ if bnb is None or not native_kernels_ready(bnb, DEVICE_TYPE):
     cgemm_4bit_inference_naive_fp16 = _bnb_required
     cgemm_4bit_inference_naive_bf16 = _bnb_required
 else:
-    cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
+    cdequantize_blockwise_fp32 = bnb_functional.lib.cdequantize_blockwise_fp32
-    cdequantize_blockwise_fp16_nf4 = bnb.functional.lib.cdequantize_blockwise_fp16_nf4
+    cdequantize_blockwise_fp16_nf4 = bnb_functional.lib.cdequantize_blockwise_fp16_nf4
-    cdequantize_blockwise_bf16_nf4 = bnb.functional.lib.cdequantize_blockwise_bf16_nf4
+    cdequantize_blockwise_bf16_nf4 = bnb_functional.lib.cdequantize_blockwise_bf16_nf4
 
     if DEVICE_TYPE == "xpu":
         # https://github.com/bitsandbytes-foundation/bitsandbytes/blob/c3b8de268fdb55a88f92feada23fc811a1e6877a/bitsandbytes/backends/xpu/ops.py#L115
         # for xpu, inference gemv using above link
-        cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemv_4bit_inference_fp16
+        cgemm_4bit_inference_naive_fp16 = bnb_functional.lib.cgemv_4bit_inference_fp16
-        cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemv_4bit_inference_bf16
+        cgemm_4bit_inference_naive_bf16 = bnb_functional.lib.cgemv_4bit_inference_bf16
     else:
-        cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemm_4bit_inference_naive_fp16
+        cgemm_4bit_inference_naive_fp16 = bnb_functional.lib.cgemm_4bit_inference_naive_fp16
-        cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemm_4bit_inference_naive_bf16
+        cgemm_4bit_inference_naive_bf16 = bnb_functional.lib.cgemm_4bit_inference_naive_bf16
 
 
 torch_device_stream = (