This is pretty much your usual RWKV, but it follows the RIIR principle.
Notable tech stack uses are memmap2 and safetensors.
Yes!
This can break and if it does use --rev
or --tag
flags to specify which version to install.
cargo +nightly install --git https://github.com/mrsteyk/rwkvk-rs --features="build-binary"
Example:
cargo run --features build-binary --release -- -t ../../RWKV-LM-deepspeed/20B_tokenizer_openchatgpt.json -m ../../RWKV-LM-deepspeed/RWKV-v4neo/RWKV-4-Pile-430M-20220808-8066.rnn.bf16.safetensors "hi!"
Either download one from hf:mrsteyk/RWKV-LM-safetensors or convert yourself using convert_safetensors.py
from gh:mrsteyk/RWKV-LM-deepspeed with arguments --bf16 --rnn
. Make sure to read the next section about alignment.
Update your safetensors
version! gh:huggingface/safetensors#148 got merged!
See gh:huggingface/safetensors#178.
TL;DR you need to pad the header, no need to reorder tensors because I copy time_
tensors (for now at least).