mamba-rs 0.6.2

Mamba SSM and Mamba-3 SISO in Rust with optional CUDA GPU acceleration. Inference and training (BPTT through SSM state, AdamW), CPU + GPU paths, custom CUDA kernels, CUDA Graph capture, f32 / bf16 / f16. Opt-in deterministic training (bit-identical runs, batch-invariant inference) with a tensor-core tier that beats cuBLAS on LLM-sized models.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

name: CI

on:
  push:
    branches: [main]
  pull_request:
  schedule:
    - cron: "0 2 * * *"

permissions:
  contents: read

concurrency:
  group: ci-${{ github.event.pull_request.number || github.sha }}
  cancel-in-progress: true

env:
  CARGO_INCREMENTAL: 0
  CARGO_NET_RETRY: 10
  CARGO_TERM_COLOR: always
  RUSTFLAGS: -Dwarnings
  RUST_BACKTRACE: 1

jobs:
  lockfile:
    name: Cargo.lock up to date
    runs-on: ubuntu-latest
    timeout-minutes: 10
    steps:
      - uses: actions/checkout@v6
      - uses: dtolnay/rust-toolchain@stable
      - run: cargo check --locked

  fmt:
    name: Rustfmt
    runs-on: ubuntu-latest
    timeout-minutes: 10
    steps:
      - uses: actions/checkout@v6
      - uses: dtolnay/rust-toolchain@stable
        with:
          components: rustfmt
      - run: cargo fmt --all -- --check

  clippy:
    name: Clippy
    runs-on: ubuntu-latest
    timeout-minutes: 20
    steps:
      - uses: actions/checkout@v6
      - uses: dtolnay/rust-toolchain@stable
        with:
          components: clippy
      - uses: Swatinem/rust-cache@v2
      - run: cargo clippy --workspace --all-targets -- -D warnings

  docs:
    name: Documentation
    runs-on: ubuntu-latest
    timeout-minutes: 20
    env:
      RUSTDOCFLAGS: -Dwarnings
    steps:
      - uses: actions/checkout@v6
      - uses: dtolnay/rust-toolchain@nightly
      - uses: Swatinem/rust-cache@v2
      - run: cargo doc --no-deps --workspace

  msrv:
    name: MSRV (1.97)
    runs-on: ubuntu-latest
    timeout-minutes: 20
    steps:
      - uses: actions/checkout@v6
      - uses: dtolnay/rust-toolchain@master
        with:
          toolchain: "1.97"
      - uses: Swatinem/rust-cache@v2
      - run: cargo check --workspace

  basics:
    name: Gate
    runs-on: ubuntu-latest
    needs: [lockfile, fmt, clippy, docs, msrv, deny]
    steps:
      - run: exit 0

  test:
    name: Test (${{ matrix.os }} / ${{ matrix.rust }})
    needs: basics
    runs-on: ${{ matrix.os }}
    timeout-minutes: 30
    strategy:
      fail-fast: false
      matrix:
        include:
          - os: ubuntu-latest
            rust: stable
          - os: ubuntu-latest
            rust: nightly
          - os: macos-latest
            rust: stable
          - os: windows-latest
            rust: stable
    steps:
      - uses: actions/checkout@v6
      - uses: dtolnay/rust-toolchain@master
        with:
          toolchain: ${{ matrix.rust }}
      - uses: Swatinem/rust-cache@v2
        with:
          key: ${{ matrix.rust }}
      - run: cargo test --workspace

  cuda-check:
    # C-1 (GEMM-map audit 2026-08-01): every GPU/tier test file is
    # cfg(feature = "cuda"), so the featureless legs compile ~50 test files
    # to nothing — the whole GEMM suite could bit-rot invisibly. Hosted
    # runners have no GPU, but type-checking only needs rustc + nvcc
    # (cudarc's build script probes `nvcc --version`; execution is not
    # attempted). Running the suite on a GPU runner stays a separate,
    # deliberate decision.
    name: CUDA feature type-check
    runs-on: ubuntu-latest
    timeout-minutes: 30
    steps:
      - uses: actions/checkout@v6
      - name: Install CUDA toolkit (nvcc only, no driver)
        run: |
          sudo apt-get update -qq
          sudo apt-get install -y -qq nvidia-cuda-toolkit
          nvcc --version
      - uses: dtolnay/rust-toolchain@master
        with:
          toolchain: "1.97"
      - uses: Swatinem/rust-cache@v2
        with:
          key: cuda-check
      - run: cargo check --workspace --all-targets --features cuda
      # The live-DDP surface (dist reducer, comm, watchdog wiring) only
      # compiles with the nccl feature; without this leg it would bit-rot
      # invisibly (cudarc's nccl bindings dlopen at runtime — no libnccl
      # needed to type-check).
      - run: cargo check --workspace --all-targets --features cuda,hf,gemm-blas,nccl

  publish-check:
    name: Publish dry-run
    needs: [basics, test]
    runs-on: ubuntu-latest
    timeout-minutes: 20
    steps:
      - uses: actions/checkout@v6
      - uses: dtolnay/rust-toolchain@stable
      - uses: Swatinem/rust-cache@v2
      - run: cargo publish --dry-run

  deny:
    name: Security audit
    runs-on: ubuntu-latest
    timeout-minutes: 10
    steps:
      - uses: actions/checkout@v6
      - uses: EmbarkStudios/cargo-deny-action@v2