Remove deprecated APIs: research module, non-blockwise optimizers, and legacy quantization functions

Remove all remaining deprecated code that has been emitting FutureWarning
since v0.45.0 (December 2024). Two prior cleanup rounds (v0.47.0, v0.49.0)
already removed the easier items; this finishes the job.

- Delete quantize(), dequantize(), quantize_no_absmax(),
  dequantize_no_absmax(), optimizer_update_8bit(), percentile_clipping(),
  and the str2optimizer8bit dispatch table from functional.py
- Remove the non-blockwise 8-bit optimizer path from Optimizer2State and
  Optimizer1State; LAMB/LARS now use blockwise quantization
- Remove percentile_clipping and block_wise parameters from all ~33
  optimizer class constructors
- Delete bitsandbytes/research/ (FP8 matmul, SwitchBack)
- Delete bitsandbytes/nn/triton_based_modules.py, SwitchBackLinearBnb,
  and the orphaned bitsandbytes/triton/ kernel directory
- Remove dead MatmulLtState fields (CxB, CxBt, formatB, _tile_indices)
- Delete test_deprecated.py, test_triton.py; clean test_autograd.py,
  test_optim.py, test_functional.py
- Remove benchmarking/switchback/ and update docs

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: TimDettmers

benchmarking/switchback/README.md

@@ -9,7 +9,7 @@
 
 import torch
 
-from . import _ops, research, utils
+from . import _ops, utils
 from .autograd._functions import (
     MatmulLtState,
     matmul,

benchmarking/switchback/info_a100_py2.jsonl

@@ -53,16 +53,12 @@ def get_current_outlier_idx(self):
 
 @dataclass
 class MatmulLtState:
-    _tile_indices: Optional[torch.Tensor] = None  # TODO: remove
-
     force_no_igemmlt: bool = False
 
     CB: Optional[torch.Tensor] = None
-    CxB: Optional[torch.Tensor] = None  # TODO: Deprecate/remove
     SB: Optional[torch.Tensor] = None
     SCB: Optional[torch.Tensor] = None
 
-    CxBt: Optional[torch.Tensor] = None  # TODO: Deprecate/remove
     SBt: Optional[torch.Tensor] = None
     CBt: Optional[torch.Tensor] = None
 
@@ -75,22 +71,15 @@ class MatmulLtState:
     is_training = True
     has_fp16_weights = True
     use_pool = False
-    formatB = "row"  # TODO: Deprecate/remove
 
     def reset_grads(self):
         self.CB = None
-        self.CxB = None
         self.SB = None
         self.SCB = None
 
-        self.CxBt = None
         self.SBt = None
         self.CBt = None
 
-    @property
-    def tile_indices(self):
-        raise ValueError("tile_indices is no longer supported.")
-
 
 class MatMul8bitLt(torch.autograd.Function):
     @staticmethod
@@ -293,7 +282,6 @@ def backward(ctx, grad_output):
 
 class MatMul4Bit(torch.autograd.Function):
     # forward is the same, but we added the fallback for pre-turing GPUs
-    # backward is mostly the same, but adds one extra clause (see "elif state.CxB is not None")
 
     @staticmethod
     def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState] = None):

benchmarking/switchback/make_plot_with_jsonl.py

@@ -4,9 +4,8 @@
 import torch
 
 try:
-    import triton.language as tl  # noqa: F401
-
     import triton  # noqa: F401
+    import triton.language as tl  # noqa: F401
 
     triton_available = True
 except ImportError: