FlashRWKV2

This repository began from MoonshotAI/FlashKDA, distributed under the MIT
License. The current model operators are implemented in TileLang. Their
numerical contracts follow the canonical RWKV-7 recurrence and training
operations published by BlinkDL/RWKV-LM, including train_temp revision
952102498e9ed367ea0a59ee64106916d474d30f, and the inference operation ordering
published by BlinkDL/Albatross.

Packed recurrent metadata retains material adapted from the vLLM project and
its RWKV integration at rwkv-rs/vllm-rwkv revision
6d683f9e49a2997e405c47edc147872c8609513b:

- csrc/validation/recurrent_metadata.cpp

That file is licensed under the Apache License, Version 2.0 and retains its
SPDX/source-revision header. The shared native validation infrastructure also
retains its Apache-2.0 headers. See LICENSES/Apache-2.0.txt.

The sampling operator contains material adapted from
Triang-jyed-driung/Rapid-Sampling revision
e0297f7830c3fa581d49ddddddba32f35ea7f733, with indexed per-request state and
penalty behavior adapted from rwkv-rs/vllm-rwkv revision
fd440426689f10e240b5761e1a7c82e4c37deb8d:

- csrc/sm80/sampling/infer_fp32_forward_varlen.cpp
- csrc/sm80/sampling/infer_fp32_forward_varlen.cu

The Rapid-Sampling contributor authorized this material for inclusion under
the MIT License. The vLLM-derived changes retain their source attribution and
MIT-compatible distribution boundary.

TileLang is an external dependency under its own MIT license. No FLA
implementation source is bundled here. Original FlashRWKV2 work is licensed
under the repository's MIT license unless a file states otherwise.
