FlashRWKV2

FlashRWKV2 is a high-performance CUDA operator library for RWKV-7. It provides composable inference and training operators; complete models, schedulers, and training frameworks remain the responsibility of downstream projects.

Requirements

• Python 3.10 or newer
• An NVIDIA CUDA build environment and a supported CUDA device
• uv, using the repository-local ./.venv
• SM120 or newer for the current native build

The runtime and reproducible source-build contracts currently pin PyTorch 2.13.0. When the build cannot detect a local GPU, set TORCH_CUDA_ARCH_LIST=12.0 explicitly.

Installation

The current alpha publishes a prebuilt CPython 3.12 Linux x86_64 wheel for SM120 GPUs. The wheel requires glibc 2.38 or newer, PyTorch 2.13.0, and a CUDA 13 runtime supplied through PyTorch's dependencies:

python -m pip install --pre FlashRWKV2

Other Python or platform combinations install from the source distribution and build the CUDA extension on the target machine:

TORCH_CUDA_ARCH_LIST=12.0 \
  python -m pip install --pre FlashRWKV2

For development from a checkout:

git clone https://github.com/rwkv-rs/FlashRWKV2.git
cd FlashRWKV2
uv sync
TORCH_CUDA_ARCH_LIST=12.0 \
  ./.venv/bin/python -m pip install -v --no-build-isolation -e .

Kernel API

See the Kernel API reference for the complete public operator surface and tensor contracts.

Tests

./.venv/bin/python -m pytest -q

CUDA tests require a successfully built flashrwkv2._C extension and a supported GPU.

Benchmarks

Operator benchmarks live in benchmarks/. For example, run the WKV7 recurrent correctness benchmark with:

./.venv/bin/python -m benchmarks.tmix.wkv7.bench \
  --shapes h32d64 \
  --dtype bfloat16 \
  --correctness-only \
  --output /tmp/flashrwkv2-wkv7-correctness.json

These benchmarks measure individual operators. They do not report or infer model-level latency.

For Albatross-compatible TMix low-rank inference, varlen means that the operator consumes packed token rows; it does not mean that one fused kernel is used for every row count. The public composite callers automatically use the canonical fused rank-in window at M<=7, the fused rank-out/value-residual window at M<=4, and the canonical large-row linear dispatcher otherwise. Callers may provide both the original checkpoint layout and a runtime layout. Both layouts must be prepared outside the timed forward region and retained for the lifetime of the inference weights; FlashRWKV2 never transposes or copies a missing layout during dispatch.

The existing low-rank operator benchmark reports correctness and latency as JSON without adding a model-level benchmark:

./.venv/bin/python benchmarks/tmix/linear/bench.py \
  --operator projection-group --layout both --channels 4096

License

FlashRWKV2 is distributed under the MIT License.