bitsandbytes · Issues· 90 open

[BUG] NF4 Linear4bit wrong on gfx1100 (prebuilt ROCm 6.4 wheel); dequantize correct

NF4 reconstruction is worse than affine int4 below blocksize 64

Several tests fail: fixture 'device' not found

[Bug] FSDP1 wrapping changes AdamW8bit state precision through the flattened shard size

PytorchLARS.step() crashes with UnboundLocalError on its own default momentum=0

Params4bit.to() skips the CPU AVX512 packing inverse that .cuda()/.xpu() apply, so model.to(device) silently corrupts weights

AdEMAMix32bit and PagedAdEMAMix32bit allocate a single-size state1 and fail on the first step

Offer to help improve the MPS backend fallbacks

quantize_4bit fails with "invalid configuration argument" (ops.cu:54) for tensors of exactly >= 2^31 elements

CPU `gemm_4bit_forward` kernel is requested without `backend="cpu"`, so it never loads on a CUDA torch build

Wrong doc and function signature for 8-bit optim

CPU dequantize_4bit returns shape (1, n) for even-length 1-D inputs; all other backends return (n,)

Can't quantize nemotron 3.5

get_gaudi_sw_version() hangs indefinitely on Windows — module-level subprocess.run with shell pipe to grep

Adam/AdamW/LAMB/AdEMAMix: weight decay applied in wrong order in default and Triton backends, diverging from CUDA kernel