FlashRWKV2

FlashRWKV2 is a high-performance CUDA operator library for RWKV-7. Its public inference API exposes the highest semantic fusion islands used by the model; kernel composition and shape-dependent dispatch remain internal. 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
• A CUDA GPU binary-compatible with the native SM90 or SM120 backends

The runtime and reproducible source-build contracts currently pin PyTorch 2.13.0. Native builds always compile separate SM90 and SM120 private extensions; the active extension is selected at runtime. A backend suffix is its minimum native cubin target, not an exact-device allowlist: _C_sm90 serves compatible SM9.x devices and _C_sm120 serves compatible SM12.x devices whose minor compute capability is at least the compiled target.

Installation

The current alpha publishes a prebuilt CPython 3.12 Linux x86_64 wheel. Its currently validated product is the RTX PRO 6000 at SM120; binary compatibility does not by itself constitute a product-level correctness or CUDA Graph claim. 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 both CUDA extensions on the target machine:

python -m pip install --pre FlashRWKV2

For development from a checkout:

git clone https://github.com/rwkv-rs/FlashRWKV2.git
cd FlashRWKV2
uv sync
./.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.

The release-specific WKV7 measurements, evidence identities, acceptance boundaries, and non-target advisory regressions are recorded in the FlashRWKV2 0.1.0a9 WKV7 performance report.

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.

TMix inference benchmarks are owned by the public wkv_prepare and readout modules. Native-private Linear, activation, VRes, gate and sparse helpers do not have standalone public benchmark entry points.

License

FlashRWKV2 is distributed under the MIT License.