mamba · Issues· 624 open

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

[][Mamba-3] Quaternion States

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

Mamba3 MIMO backward cannot launch on a lower-shared-memory Blackwell SKU (220 KB requested vs 99 KB limit)

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

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

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

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

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

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

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

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

How to understand the relationship between mamba2 and GLA?

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

Verify evals on Papers with Code