First release

.github/workflows/publish.yaml

@@ -0,0 +1,212 @@
+# This workflow will:
+# - Create a new Github release
+# - Build wheels for supported architectures
+# - Deploy the wheels to the Github release
+# - Release the static code to PyPi
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+name: Build wheels and deploy
+
+on:
+  create:
+    tags:
+      - v*
+
+jobs:
+
+  setup_release:
+    name: Create Release
+    runs-on: ubuntu-latest
+    steps:
+      - name: Get the tag version
+        id: extract_branch
+        run: echo ::set-output name=branch::${GITHUB_REF#refs/tags/}
+        shell: bash
+
+      - name: Create Release
+        id: create_release
+        uses: actions/create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ steps.extract_branch.outputs.branch }}
+          release_name: ${{ steps.extract_branch.outputs.branch }}
+
+  build_wheels:
+    name: Build Wheel
+    needs: setup_release
+    runs-on: ${{ matrix.os }}
+
+    strategy:
+      fail-fast: false
+      matrix:
+          # Using ubuntu-20.04 instead of 22.04 for more compatibility (glibc). Ideally we'd use the
+          # manylinux docker image, but I haven't figured out how to install CUDA on manylinux.
+          os: [ubuntu-20.04]
+          python-version: ['3.7', '3.8', '3.9', '3.10', '3.11']
+          torch-version: ['1.12.1', '1.13.1', '2.0.1', '2.1.1', '2.2.0.dev20231127']
+          cuda-version: ['11.8.0', '12.2.0']
+          # We need separate wheels that either uses C++11 ABI (-D_GLIBCXX_USE_CXX11_ABI) or not.
+          # Pytorch wheels currently don't use it, but nvcr images have Pytorch compiled with C++11 ABI.
+          # Without this we get import error (undefined symbol: _ZN3c105ErrorC2ENS_14SourceLocationESs)
+          # when building without C++11 ABI and using it on nvcr images.
+          cxx11_abi: ['FALSE', 'TRUE']
+          exclude:
+            # Pytorch <= 1.12 does not support Python 3.11
+            - torch-version: '1.12.1'
+              python-version: '3.11'
+            # Pytorch >= 2.0 only supports Python >= 3.8
+            - torch-version: '2.0.1'
+              python-version: '3.7'
+            - torch-version: '2.1.1'
+              python-version: '3.7'
+            - torch-version: '2.2.0.dev20231127'
+              python-version: '3.7'
+            # Pytorch <= 2.0 only supports CUDA <= 11.8
+            - torch-version: '1.12.1'
+              cuda-version: '12.2.0'
+            - torch-version: '1.13.1'
+              cuda-version: '12.2.0'
+            - torch-version: '2.0.1'
+              cuda-version: '12.2.0'
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Set CUDA and PyTorch versions
+        run: |
+          echo "MATRIX_CUDA_VERSION=$(echo ${{ matrix.cuda-version }} | awk -F \. {'print $1 $2'})" >> $GITHUB_ENV
+          echo "MATRIX_TORCH_VERSION=$(echo ${{ matrix.torch-version }} | awk -F \. {'print $1 "." $2'})" >> $GITHUB_ENV
+
+      - name: Free up disk space
+        if: ${{ runner.os == 'Linux' }}
+        # https://github.com/easimon/maximize-build-space/blob/master/action.yml
+        # https://github.com/easimon/maximize-build-space/tree/test-report
+        run: |
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /opt/hostedtoolcache/CodeQL
+
+      - name: Set up swap space
+        if: runner.os == 'Linux'
+        uses: pierotofy/set-swap-space@v1.0
+        with:
+          swap-size-gb: 10
+
+      - name: Install CUDA ${{ matrix.cuda-version }}
+        if: ${{ matrix.cuda-version != 'cpu' }}
+        uses: Jimver/cuda-toolkit@v0.2.11
+        id: cuda-toolkit
+        with:
+          cuda: ${{ matrix.cuda-version }}
+          linux-local-args: '["--toolkit"]'
+          # default method is "local", and we're hitting some error with caching for CUDA 11.8 and 12.1
+          # method: ${{ (matrix.cuda-version == '11.8.0' || matrix.cuda-version == '12.1.0') && 'network' || 'local' }}
+          method: 'network'
+          # We need the cuda libraries (e.g. cuSparse, cuSolver) for compiling PyTorch extensions,
+          # not just nvcc
+          # sub-packages: '["nvcc"]'
+
+      - name: Install PyTorch ${{ matrix.torch-version }}+cu${{ matrix.cuda-version }}
+        run: |
+          pip install --upgrade pip
+          # If we don't install before installing Pytorch, we get error for torch 2.0.1
+          # ERROR: Could not find a version that satisfies the requirement setuptools>=40.8.0 (from versions: none)
+          pip install lit
+          # We want to figure out the CUDA version to download pytorch
+          # e.g. we can have system CUDA version being 11.7 but if torch==1.12 then we need to download the wheel from cu116
+          # This code is ugly, maybe there's a better way to do this.
+          export TORCH_CUDA_VERSION=$(python -c "import os; minv = {'1.12': 113, '1.13': 116, '2.0': 117, '2.1': 118, '2.2': 118}[os.environ['MATRIX_TORCH_VERSION']]; maxv = {'1.12': 116, '1.13': 117, '2.0': 118, '2.1': 121, '2.2': 121}[os.environ['MATRIX_TORCH_VERSION']]; print(max(min(int(os.environ['MATRIX_CUDA_VERSION']), maxv), minv))")
+          if [[ ${{ matrix.torch-version }} == *"dev"* ]]; then
+            pip install --no-cache-dir --pre torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
+          else
+            pip install --no-cache-dir torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
+          fi
+          nvcc --version
+          python --version
+          python -c "import torch; print('PyTorch:', torch.__version__)"
+          python -c "import torch; print('CUDA:', torch.version.cuda)"
+          python -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)"
+        shell:
+          bash
+
+      - name: Build wheel
+        run: |
+          # We want setuptools >= 49.6.0 otherwise we can't compile the extension if system CUDA version is 11.7 and pytorch cuda version is 11.6
+          # https://github.com/pytorch/pytorch/blob/664058fa83f1d8eede5d66418abff6e20bd76ca8/torch/utils/cpp_extension.py#L810
+          # However this still fails so I'm using a newer version of setuptools
+          pip install setuptools==68.0.0
+          pip install ninja packaging wheel
+          export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
+          export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
+          # Limit MAX_JOBS otherwise the github runner goes OOM
+          MAX_JOBS=2 MAMBA_FORCE_BUILD="TRUE" MAMBA_FORCE_CXX11_ABI=${{ matrix.cxx11_abi}} python setup.py bdist_wheel --dist-dir=dist
+          tmpname=cu${MATRIX_CUDA_VERSION}torch${MATRIX_TORCH_VERSION}cxx11abi${{ matrix.cxx11_abi }}
+          wheel_name=$(ls dist/*whl | xargs -n 1 basename | sed "s/-/+$tmpname-/2")
+          ls dist/*whl |xargs -I {} mv {} dist/${wheel_name}
+          echo "wheel_name=${wheel_name}" >> $GITHUB_ENV
+
+      - name: Log Built Wheels
+        run: |
+          ls dist
+
+      - name: Get the tag version
+        id: extract_branch
+        run: echo ::set-output name=branch::${GITHUB_REF#refs/tags/}
+
+      - name: Get Release with tag
+        id: get_current_release
+        uses: joutvhu/get-release@v1
+        with:
+          tag_name: ${{ steps.extract_branch.outputs.branch }}
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Upload Release Asset
+        id: upload_release_asset
+        uses: actions/upload-release-asset@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          upload_url: ${{ steps.get_current_release.outputs.upload_url }}
+          asset_path: ./dist/${{env.wheel_name}}
+          asset_name: ${{env.wheel_name}}
+          asset_content_type: application/*
+
+  publish_package:
+    name: Publish package
+    needs: [build_wheels]
+
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - uses: actions/setup-python@v4
+        with:
+          python-version: '3.10'
+
+      - name: Install dependencies
+        run: |
+          pip install ninja packaging setuptools wheel twine
+          # We don't want to download anything CUDA-related here
+          pip install torch --index-url https://download.pytorch.org/whl/cpu
+
+      - name: Build core package
+        env:
+          MAMBA_SKIP_CUDA_BUILD: "TRUE"
+        run: |
+          python setup.py sdist --dist-dir=dist
+
+      - name: Deploy
+        env:
+          TWINE_USERNAME: "__token__"
+          TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
+        run: |
+          python -m twine upload dist/*

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "3rdparty/lm-evaluation-harness"]
+	path = 3rdparty/lm-evaluation-harness
+	url = https://github.com/EleutherAI/lm-evaluation-harness/

AUTHORS

@@ -0,0 +1,2 @@
+Tri Dao, tri@tridao.me
+Albert Gu, agu@andrew.cmu.edu

README.md

@@ -1,5 +1,141 @@
 # Mamba
 
-This repository contains the code for the paper [Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://arxiv.org/abs/2312.00752).
+![Mamba](assets/selection.png "Selective State Space")
+> **Mamba: Linear-Time Sequence Modeling with Selective State Spaces**\
+> Albert Gu*, Tri Dao*\
+> Paper: https://arxiv.org/abs/2312.00752
 
-The first official code release of the paper will be uploaded around noon EST, Monday Dec. 4.
+## Installation
+
+- `pip install causal-conv1d`: an efficient implemention of a simple causal Conv1d layer used inside the Mamba block.
+- `pip install mamba-ssm`: the core Mamba package.
+
+If `pip` complains about PyTorch versions, try passing `--no-build-isolation` to `pip`.
+
+Other requirements:
+- Linux
+- NVIDIA GPU
+- PyTorch 1.12+
+- CUDA 11.6+
+
+## Usage
+
+We expose several levels of interface with the Mamba model.
+
+### Selective SSM
+
+Mamba is based on a selective SSM layer, which is the focus of the paper (Section 3; Algorithm 2).
+
+Source: [ops/selective_scan_interface.py](mamba_ssm/ops/selective_scan_interface.py).
+
+### Mamba Block
+
+The main module of this repository is the Mamba architecture block wrapping the selective SSM.
+
+Source: [modules/mamba_simple.py](mamba_ssm/modules/mamba_simple.py).
+
+Usage:
+```
+from mamba_ssm import Mamba
+
+batch, length, dim = 2, 64, 16
+x = torch.randn(batch, length, dim).to("cuda")
+model = Mamba(
+    # This module uses roughly 3 * expand * d_model^2 parameters
+    d_model=dim, # Model dimension d_model
+    d_state=16,  # SSM state expansion factor
+    d_conv=4,    # Local convolution width
+    expand=2,    # Block expansion factor
+).to("cuda")
+y = model(x)
+assert y.shape == x.shape
+```
+
+### Mamba Language Model
+
+Finally, we provide an example of a complete language model: a deep sequence model backbone (with repeating Mamba blocks) + language model head.
+
+Source: [models/mixer_seq_simple.py](mamba_ssm/models/mixer_seq_simple.py).
+
+This is an example of how to integrate Mamba into an end-to-end neural network.
+This example is used in the generation scripts below.
+
+
+
+## Pretrained Models
+
+Pretrained models are uploaded to
+[HuggingFace](https://huggingface.co/state-spaces): `mamba-130m`, `mamba-370m`,
+`mamba-790m`, `mamba-1.4b`, `mamba-2.8b`.
+
+The models will be autodownloaded by the generation script below.
+
+These models were trained on the [Pile](https://huggingface.co/datasets/EleutherAI/pile), and follow the standard model dimensions described by GPT-3 and followed by many open source models:
+
+| Parameters | Layers | Model dim. | 
+|------------|--------|------------|
+| 130M       | 12     | 768        |
+| 370M       | 24     | 1024       |
+| 790M       | 24     | 1536       |
+| 1.4B       | 24     | 2048       |
+| 2.8B       | 32     | 2560       |
+
+(The layer count of Mamba should be doubled, as two Mamba blocks are needed for each "layer" (MHA block + MLP block) of a Transformer.)
+
+Note: these are base models trained only for 300B tokens, without any form of downstream modification (instruction tuning, etc.).
+Performance is expected to be comparable or better than other architectures trained on similar data, but not to match larger or fine-tuned models.
+
+
+## Evaluations
+
+To run zero-shot evaluations of models (corresponding to Table 3 of the paper),
+we use the
+[lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/big-refactor)
+library.
+
+1. Pull the `lm-evaluation-harness` repo by `git submodule update --init
+   --recursive`. We use the `big-refactor` branch.
+2. Install `lm-evaluation-harness`: `pip install -e 3rdparty/lm-evaluation-harness`
+3. Run evaluation with (more documentation at the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/big-refactor) repo):
+```
+python evals/lm_harness_eval.py --model mamba --model_args pretrained=state-spaces/mamba-130m --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande --device cuda --batch_size 64
+python evals/lm_harness_eval.py --model hf --model_args pretrained=EleutherAI/pythia-160m --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande --device cuda --batch_size 64
+```
+
+Note that the result of each task might differ from reported values by 0.1-0.3 due to noise in the evaluation process.
+
+## Inference
+
+The script [benchmarks/benchmark_generation_mamba_simple.py](benchmarks/benchmark_generation_mamba_simple.py)
+1. autoloads a model from the HuggingFace Hub,
+2. generates completions of a user-specified prompt,
+3. benchmarks the inference speed of this generation.
+
+Other configurable options include the top-p (nucleus sampling) probability, and the softmax temperature.
+
+### Examples
+
+To test generation latency (e.g. batch size = 1) with different sampling strategies:
+
+```
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba-2.8b" --prompt "My cat wrote all this CUDA code for a new language model and" --topp 0.9 --temperature 0.5
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "EleutherAI/pythia-2.8b" --prompt "My cat wrote all this CUDA code for a new language model and" --topp 0.9 --temperature 0.5
+```
+
+To test generation throughput with random prompts (e.g. large batch size):
+```
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba-2.8b" --batch 128
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "EleutherAI/pythia-2.8b" --batch 128
+```
+
+## Citation
+
+If you use this codebase, or otherwise found our work valuable, please cite Mamba:
+```
+@article{mamba,
+  title={Mamba: Linear-Time Sequence Modeling with Selective State Spaces},
+  author={Gu, Albert and Dao, Tri},
+  journal={arXiv preprint arXiv:2312.00752},
+  year={2023}
+}
+```