rwkv7-g1-st

RWKV-7 G1 — Native Transformers Safetensors

This repository contains native Transformers Safetensors conversions of canonical RWKV-7 G1 checkpoints from BlinkDL/rwkv7-g1 and selected temporary training checkpoints from BlinkDL/temp-latest-training-models. Each checkpoint is a self-contained subfolder with model weights, configuration, the native RWKV World tokenizer, and chat prompt template. No remote model code is required.

Checkpoints

The repository root intentionally does not contain a default model checkpoint. Select one of the subfolders explicitly.

G1l temporary training checkpoints

These are preview training checkpoints, retaining the original source filenames as subfolder names. Source revision: 624363d666cf64e8fbeb1b2e029aef78ab1dec08. Exact source-to-Safetensors comparisons and native reload checks passed for every checkpoint; each subfolder records its source and shard SHA256 values in PROVENANCE.md.

The context value is deployment metadata matching the existing serving limit. The temporary source filenames do not declare a training context length.

Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer


repo_id = "rwkv-rs/rwkv7-g1-st"
subfolder = "rwkv7-g1j-2.9b-20260831-ctx16384"

tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    subfolder=subfolder,
    trust_remote_code=False,
)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    subfolder=subfolder,
    dtype=torch.float16,
    trust_remote_code=False,
)

RWKV-7 product computation is delegated to the published FlashRWKV2 operator package. Unsupported hardware or a missing native operator fails closed; there is no repository-local PyTorch, FLA, or CUDA-kernel fallback.

Artifact contract

• Public model identity: model_type="rwkv", RwkvForCausalLM.
• Architecture generation: architecture_version="rwkv7".
• Model vocabulary: 65,536 entries.
• Tokenizer vocabulary: IDs 0 through 65,529.
• BOS and EOS share token ID 0; PAD and UNK are intentionally undefined.
• Weight dtype: bfloat16.
• Each subfolder contains config.json, three generation configurations, Safetensors, tokenizer files, chat_template.jinja, and PROVENANCE.md.
• The root rwkv_vocab_v20230424.json is retained as the immutable canonical source vocabulary asset used to build the native tokenizer artifacts.

The included chat template is a prompt interface. It does not claim that these base checkpoints are safety aligned.

License

The converted model weights are distributed under Apache-2.0. FlashRWKV2 is a separate MIT-licensed operator package.