Repository Issue Activity (beta)

#2075 AdEMAMix32bit and PagedAdEMAMix32bit allocate a single-size state1 and fail on the first step

#2011 Offer to help improve the MPS backend fallbacks

#2067 quantize_4bit fails with "invalid configuration argument" (ops.cu:54) for tensors of exactly >= 2^31 elements

#2064 CPU `gemm_4bit_forward` kernel is requested without `backend="cpu"`, so it never loads on a CUDA torch build

#1261 Wrong doc and function signature for 8-bit optim

#2047 CPU dequantize_4bit returns shape (1, n) for even-length 1-D inputs; all other backends return (n,)

#2045 Can't quantize nemotron 3.5

#2042 Triton 4-bit quantization stores absmax past the end of the tensor unless the block count is a multiple of 8

#2041 get_gaudi_sw_version() hangs indefinitely on Windows — module-level subprocess.run with shell pipe to grep

#2010 Adam/AdamW/LAMB/AdEMAMix: weight decay applied in wrong order in default and Triton backends, diverging from CUDA kernel

#2034 `MatMul4Bit` keeps `B` and `quant_state` on `ctx` instead of `save_for_backward`, so gradient checkpointing cannot recompute them

#2025 Triton FP4 dequantization has a truncated constant, disagreeing with CUDA/CPU by 7 ULP

#2027 `gemm_4bit`: blocksize-alignment warning emitted on every call, and now also during training (new in 0.50.0)

#2021 Target-wide `-mavx512*` puts EVEX in the scalar dequant fallbacks; source builds SIGILL on non-AVX-512 x86_64

#1785 Support quantizing tensors when numel() > INT_MAX

#1126 Linux binaries should ship with appropriate RPATH

#1868 Question: intentional FP16-only path for int8_vectorwise_quant / LLM.int8 activation quant? (BF16 support + removing casts)

#2014 ROCm 8.4 binary not found (latest ROCm is 7.14)

#1849 Failed to quant MoE models with fused expert weights in transformers v5

#2005 Memory leak: nn.parametrize.replace_parameter_4bit + HF gradient_checkpointing_enable() — dequant cache not released on early-stopped recompute

#1991 Prebuilt XPU wheel links `libsycl.so.8`; fails to load on oneAPI 2026 (`libsycl.so.9`)

#1995 quantize_blockwise / dequantize_blockwise produce incorrect results for non-contiguous CPU inputs (CPU follow-up to #1690)

#1992 Lion applies coupled (L2) weight decay instead of decoupled in 32-bit CUDA, Triton, and the default backend

#1955 [ROCm] gfx1201 (Navi 48) support - compilation guide and known issues

#1988 pip wheel ships top-level tests/ directory; shadows consumers tests/ package and breaks python -m unittest discovery