[deep1blue]

What does this PR do?

Fixes two issues in src/lightning/fabric/plugins/precision/bitsandbytes.py:

1. In-place $O(N)$ filtering in _ignore_missing_weights_hook:
   Previously, _ignore_missing_weights_hook iterated over reversed(incompatible_keys.missing_keys) and called incompatible_keys.missing_keys.remove(key) on each "*weight" entry:

   for key in reversed(incompatible_keys.missing_keys):
       if key.endswith("weight"):
           incompatible_keys.missing_keys.remove(key)

   Because list.remove(key) always searches linearly from index 0, removing elements from the end of missing_keys requires scanning the entire prefix on every removal ($O(N^2)$ time). Replacing this with in-place slice assignment (incompatible_keys.missing_keys[:] = [key for key in incompatible_keys.missing_keys if not key.endswith("weight")]) preserves the missing_keys list object identity in $O(N)$ time (reducing 10,000-key filtering from 401.2 ms to 0.49 ms, an ~825x speedup).

2. Correct Int8Params / torch.int8 guards in _Linear8bitLt.to_empty() and _Linear8bitLt.reset_parameters():
   _Linear8bitLt.to_empty() and _Linear8bitLt.reset_parameters() checked self.weight.dtype == torch.uint8 and isinstance(self.weight, bnb.nn.Params4bit) (which are the 4-bit invariants from _Linear4bit), whereas _Linear8bitLt uses bnb.nn.Int8Params and torch.int8 storage (as checked in _Linear8bitLt.quantize_()). As a result:

   • Calling to_empty() on an already-quantized _Linear8bitLt layer (dtype == torch.int8) bypassed the NotImplementedError guard.
   • In _Linear8bitLt.reset_parameters(), linear_init_finished = isinstance(self.weight, bnb.nn.Params4bit) was always False, bypassing both the quantized NotImplementedError guard (causing kaiming_uniform_ to fail with RuntimeError: "normal_kernel_cpu" not implemented for 'Char') and the post-init re-quantization / _replace_param branch.

Validation

pytest tests/tests_fabric/plugins/precision/test_bitsandbytes.py -k test_ignore_missing_weights_hook_and_linear8bit_guards

Acknowledgments

Credit and thanks to @carmocca (Carlos Mocholí, #18655, #19150), @awaelchli (Adrian Wälchli, #19886, #20121), @Borda (Jirka Borovec, #20911), @matthewdouglas (Matthew Douglas, #20956), and @bhimrazy (Bhimraj Yadav, #21858) for designing and maintaining the BitsandbytesPrecision plugin in Lightning Fabric.