Repository Issue Activity (beta)

#1015 Sequences longer than ~61K tokens (and often less) silently give a wrong answer or crash

#959 [🤖][Mamba-3] Quaternion States

#1022 Mamba-3 incremental decode broken: `step()` receives `(B, 1, D)` but `_preprocess` expects 2-D — EinopsError in `.generate()` for both official 1.5B checkpoints

#1024 Mamba3 step() is non-deterministic and does not match forward()

#1019 No prebuilt wheels for sm_103 (B300) and sm_121 (GB10)

#1018 mamba3 utils.py: PTX inline-asm helpers (cos_approx/sin_approx/tanh_approx) with NVPTX 'f' constraints cannot compile on ROCm/AMDGPU

#891 PyPI wheels broken on aarch64 (DGX Spark / ARM Grace)

#1017 Mamba3 SISO kernel: 3-7.5% deviation from reference (heavy-tailed, masked by bf16 rtol=1e-1)

#1013 Training API support for initial/final recurrent state (TBPTT / chunked sequence training)

#1014 Mamba3.forward() passes a 3-D tensor to step(), which requires 2-D — decode path raises EinopsError

#380 How to understand the relationship between mamba2 and GLA?

#986 apache-tvm-ffi<=0.1.12 (after #984) is incompatible with pinned tilelang==0.1.8 — measured working ceiling is 0.1.9

#999 Publish wheels for torch 2.11 / 2.12 / 2.13 — the PyPI matrix has been pinned at 2.10 since February

#938 Mamba3-SISO exploding gradients on B200

#998 Verify evals on Papers with Code

#72 loss nan after several epochs

#716 Perplexity Discrepancy Between Slow and Fast Paths

#985 Mamba3 MIMO forward doesn't work with seq_length=1

#991 Can Selective Scan be used to quantify acceleration?

#990 Mamba-3 fails on RTX PRO 6000 (Blackwell) due to smaller per-SM shared memory (128KB vs 160KB)

#983 Installation hangs while building `selective_scan_cuda` on Google Colab (Python 3.12 / PyTorch 2.11 / CUDA 12.8)

#625 TORCH_CUDA_ARCH_LIST support

#982 Not gonna leak mamba-5

#974 Installation and Build Failure on Modern Cloud Environments (Python 3.12+ / PyTorch 2.X / Kaggle Environments)

#815 Regarding the issue of building the Mamba-ssm from source code