Test that we can successfully quantize a large tensor. Note that the following limitations apply: - On CUDA/XPU/ROCm, the maximum num

Test that AdEMAMix can save/load state_dict and continue training without NaN. Regression test for https://github.com/bitsandbytes-foundation/bit

Test parametrization with different block sizes to verify flexibility.

Test that we can successfully quantize a large tensor. Note that the following limitations apply: - On CUDA/XPU/ROCm, the maximum num

Test that proper errors are raised for invalid inputs.

Integration test: FSDP get_model_state_dict with cpu_offload on a 4-bit model (#1405). Launches a single-GPU FSDP process via torchrun to exercis

Regression test: offset tensors not in float32 must still produce correct results.

Test that quantized parameters have proper gradient behavior.

Regression test for #1904: __getattr__ on Params4bit causes graph breaks under torch.compile. Activation checkpointing replays the forward pass d

Test moving a Linear8bitLt layer between CPU and an accelerator device.

Lion must use *decoupled* weight decay (p *= 1 - lr*wd), matching the Lion paper and the lion_pytorch reference. Regression test for a couple

Test parametrization with MoE-style parameter shape

Test realistic MoE forward computation with quantized expert weights.

Test applying parametrization to multiple parameters in the same module.

BNB_CUDA_VERSION overrides path selection.

Test that override_config works when called after register_parameters (issue #1269).

Override value must be digits only (e.g. '124'), not dotted or alphanumeric.

Test the Bnb4bitParametrization forward method directly.

Test that Params4bit proxies QuantState attributes for FSDP state_dict traversal (#1405). PyTorch's FSDP state_dict machinery traverses FQN paths

Test that torch.chunk and torch.split preserve Params4bit subclass for FSDP2 compatibility.

Test applying parametrization to already quantized parameters.

Test that quantization state is properly preserved and accessible.

Test that quantized weights survive a flatten-chunk-reassemble roundtrip. Non-uint8 quant_storage exists so that FSDP can shard quantized tensors

Test quantize/dequantize roundtrip when n_elements is not divisible by blocksize.

End-to-end test: quantize non-contiguous, dequantize, compare with contiguous path.