Models Troubleshooting

PlantHelixSeek predictions are uniform / positionally meaningless

PlantHelixSeek-CRE/-Anno remote code imports fla.ops.kda.chunk.chunk_kda from flash-linear-attention for its HelixSeekDelta (KDA) layers. Without the package the models still load and run — there is no error — but 9/39 layers silently fall back to a pure-PyTorch path that is not KDA math, and outputs become positionally uninformative (measured: Anno predicts all-O at 0.82–0.98 confidence everywhere; CRE gives p(CRE) ≈ 0.007 both inside and outside real DHS sites).

Fix: install the fla extra (included in all):

uv pip install -e '.[fla]'

With the kernels installed the same checkpoints separate cleanly (p(CRE) 0.77 in-DHS vs 0.22 non-DHS on the probe set). Install the package bare — since v0.5 its [cuda]/[rocm] extras pin their own torch and would downgrade your environment — and keep the version within 0.5.x (chunk_kda semantics are not stable across minor versions).

Mamba models on macOS (Apple Silicon)

mamba-ssm relies on Triton, which only provides pre-built wheels for Linux. As a result, the [mamba] extra cannot be installed on macOS.

Workaround: HuggingFace transformers includes a pure-PyTorch Mamba implementation (MambaModel) that works on macOS without mamba-ssm. To use it:

# Install without the [mamba] extra
uv pip install -e '.[base]'

Then load a Mamba model through transformers (cpu only) as usual. Note that this fallback path is significantly slower than the optimized mamba-ssm kernels (which require Linux + CUDA).