bump: bitsandbytes >=0.45.2,<0.47.0 & compatibility patch for bnb.functional.int8_double_quant (#20911)

* compatibility patch for `bnb.functional.int8_double_quant`
* bitsandbytes >=0.45.2,<0.47.0

---------

Co-authored-by: Copilot <[email protected]>

Author: Borda

requirements/fabric/strategies.txt

@@ -6,4 +6,4 @@
 # note: is a bug around 0.10 with `MPS_Accelerator must implement all abstract methods`
 #  shall be resolved by https://github.com/microsoft/DeepSpeed/issues/4372
 deepspeed >=0.8.2, <=0.9.3; platform_system != "Windows" and platform_system != "Darwin"  # strict
-bitsandbytes >=0.45.2,<0.45.3; platform_system != "Darwin"
+bitsandbytes >=0.45.2,<0.47.0; platform_system != "Darwin"

requirements/pytorch/extra.txt

@@ -8,4 +8,4 @@ hydra-core >=1.2.0, <1.4.0
 jsonargparse[signatures] >=4.39.0, <4.41.0
 rich >=12.3.0, <14.1.0
 tensorboardX >=2.2, <2.7.0  # min version is set by torch.onnx missing attribute
-bitsandbytes >=0.45.2,<0.45.3; platform_system != "Darwin"
+bitsandbytes >=0.45.2,<0.47.0; platform_system != "Darwin"

src/lightning/fabric/plugins/precision/bitsandbytes.py

@@ -256,9 +256,10 @@ def quantize(
             if int8params.has_fp16_weights:
                 int8params.data = B
             else:
-                CB, CBt, SCB, SCBt, _ = bnb.functional.double_quant(B)
-                del CBt
-                del SCBt
+                if hasattr(bnb.functional, "double_quant"):
+                    CB, _, SCB, _, _ = bnb.functional.double_quant(B)
+                else:  # for bitsandbytes versions ≥0.46
+                    CB, SCB = bnb.functional.int8_double_quant(B)
                 int8params.data = CB
                 setattr(int8params, "CB", CB)
                 setattr(int8params, "SCB", SCB)