Issues

• LLM.int8(): int8 CUDA ops ignore input strides, so Linear8bitLt silently returns wrong rows for non-contiguous inputs like hidden_states[:, 0]

  bitsandbytes-foundation/bitsandbytes#2130 · Filter by author Mazukiri opened on Nov 12, 2026
  1

• gemm_4bit (via matmul_4bit / MatMul4Bit) and dequantize_4bit + F.linear differ by up to 1 ulp for M <= 256 on sm_120 (RTX 5070 Laptop), fp16 and bf16 - intended, and can the path be pinned?

  bitsandbytes-foundation/bitsandbytes#2121 · Filter by author MakazhanAlpamys opened on Nov 6, 2026

• CPU quantize_blockwise stores 0.0 as a neighbouring code, so Adam8bit on CPU moves weights with zero gradient

  bitsandbytes-foundation/bitsandbytes#2118 · Filter by author damienen opened on Nov 2, 2026
  1

• LAMB / LARS never update a parameter whose norm is zero (zero-initialised biases, LoRA B matrices, zero-init residual gains)

  bitsandbytes-foundation/bitsandbytes#2116 · Filter by author Andreas-Hum opened on Nov 1, 2026
  2

• MatMul8bitLt backward pairs row-quantized grad_output with column stats, so weight gradients are wrong

  bitsandbytes-foundation/bitsandbytes#2112 · Filter by author egeozkoc opened on Oct 28, 2026

• k-bit linear bias gradient is invalid for 1-D inputs (and wrong rank for >2-D)

  bitsandbytes-foundation/bitsandbytes#2110 · Filter by author anh-nguyenn opened on Oct 26, 2026

• Linear8bitLt CPU/XPU backward crashes for non-2D inputs in MatMul8bitFp

  bitsandbytes-foundation/bitsandbytes#2108 · Filter by author anh-nguyenn opened on Oct 26, 2026

• dequantize_4bit reads out of bounds when quant_state.shape exceeds the packed tensor

  bitsandbytes-foundation/bitsandbytes#2106 · Filter by author tritsystem opened on Oct 25, 2026

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

  bitsandbytes-foundation/bitsandbytes#2087 · Filter by author Yxmura opened on Sep 13, 2026
  1

• NF4 reconstruction is worse than affine int4 below blocksize 64

  bitsandbytes-foundation/bitsandbytes#2086 · Filter by author Asadkhan282 opened on Sep 11, 2026
  3

• Several tests fail: fixture 'device' not found

  bitsandbytes-foundation/bitsandbytes#2084 · Filter by author yurivict opened on Sep 11, 2026
  1

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

  bitsandbytes-foundation/bitsandbytes#2080 · Filter by author gss10282025 opened on Sep 10, 2026
  1

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

  bitsandbytes-foundation/bitsandbytes#2079 · Filter by author caiotheodoro opened on Sep 9, 2026

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

  bitsandbytes-foundation/bitsandbytes#2078 · Filter by author caiotheodoro opened on Sep 9, 2026

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

  bitsandbytes-foundation/bitsandbytes#2075 · Filter by author caiotheodoro opened on Sep 8, 2026
  1

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

  bitsandbytes-foundation/bitsandbytes#2067 · Filter by author TomMoeras opened on Sep 3, 2026
  2

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

  bitsandbytes-foundation/bitsandbytes#2064 · Filter by author jjjsood opened on Aug 30, 2026
  1

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

  bitsandbytes-foundation/bitsandbytes#2047 · Filter by author 2sumtech opened on Aug 20, 2026
  1

• Can't quantize nemotron 3.5

  bitsandbytes-foundation/bitsandbytes#2045 · Filter by author supersimple33 opened on Aug 19, 2026

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

  DuplicateProposing to CloseWindows
  Bug

  bitsandbytes-foundation/bitsandbytes#2041 · Filter by author Mcic1980 opened on Aug 15, 2026
  1

• MatMul4Bit keeps B and quant_state on ctx instead of save_for_backward, so gradient checkpointing cannot recompute them

  bitsandbytes-foundation/bitsandbytes#2034 · Filter by author MakazhanAlpamys opened on Aug 7, 2026

• gemm_4bit: blocksize-alignment warning emitted on every call, and now also during training (new in 0.50.0)

  bitsandbytes-foundation/bitsandbytes#2027 · Filter by author albertvillanova opened on Jul 29, 2026
  1

• Target-wide -mavx512* puts EVEX in the scalar dequant fallbacks; source builds SIGILL on non-AVX-512 x86_64

  BuildLinuxx64 CPU
  Bug

  bitsandbytes-foundation/bitsandbytes#2021 · Filter by author pjordanandrsn opened on Jul 27, 2026
  4

• Offer to help improve the MPS backend fallbacks

  bitsandbytes-foundation/bitsandbytes#2011 · Filter by author egeozkoc opened on Jul 17, 2026
  7

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

  CUDAOptimizers
  Bug

  bitsandbytes-foundation/bitsandbytes#2010 · Filter by author ErenAta16 opened on Jul 17, 2026 · v0.51.0
  7