# Generate FlashInfer Source Files

```bash
# 1. Clone FlashInfer (pin version with --branch)
git clone --depth 1 --branch v0.2.0 https://github.com/flashinfer-ai/flashinfer flashinfer

# 2. Apply patch (if needed)
cd flashinfer && git apply ../generated.patch

# 3. Generate AOT files
export CUDA_HOME=/usr/local/cuda-12.6 && export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.7 8.9 9.0" && uv run --with torch python -m flashinfer.aot

# 4. Copy generated files to csrc directory
cp -r build/aot/generated ../csrc/generated
```

now manually comment out the `TORCH_LIBRARY_FRAGMENT` in the generated files

specifically in the activation `csrc/generated/gelu_and_mul.cu` and etc since those are included in the current build.

### build and test

```bash
nix develop -L .#test --command python tests/simple_test.py
```

```txt
GELU and multiply operation completed. Output shape: torch.Size([128, 2048])
Output tensor sample: tensor([[ 7.8613e-02, -4.7656e-01, -4.8637e-03,  6.2073e-02, -3.8745e-01],
        [ 1.5686e-01, -5.0964e-03, -3.8981e-04,  1.8945e+00, -7.3792e-02],
...
```