How konpep-dev/aether-rwkv- works
Aether is a bilingual RWKV language-model project. Its product workflow runs from synthetic-corpus creation and tokenization through model training, then interactive prompt encoding, recurrent token generation, and decoded chat output. The graph preserves the documented dataset, tokenizer, training, inference, and RWKV capabilities. Some subsystem links rely on README documentation because the excerpts are partial; unsampled implementation details are not inferred.
Main entry points: train.py Training loop · inference.py Chat inference · model.py RWKV model · tokenizer.py Byte-BPE tokenizer
Components 11
Data preparation
- Corpus generator
generate_dataset.py - Synthetic corpus
- Byte-BPE tokenizer
tokenizer.py
Model training
- Training loop
train.py - Model checkpoint
Chat inference
- Chat inference
inference.py
RWKV engine
- RWKV model
model.py - Time-mix recurrence
model.py - Channel-mix layer
model.py - WKV acceleration
wkv_cuda_kernel.py
Other components
- Chat user
Connections 12
- Corpus generator → Synthetic corpus: generates
README.md - Synthetic corpus → Training loop: training data
- Training loop → Byte-BPE tokenizer: encodes samples
train.py - Training loop → RWKV model: trains
train.py - Training loop → Model checkpoint: saves weights
train.py - Model checkpoint → Chat inference: loads weights
- Chat user → Chat inference: submits prompts
README.md - Chat inference → Byte-BPE tokenizer: encodes and decodes
inference.py - Chat inference → RWKV model: generates tokens
inference.py - RWKV model → Time-mix recurrence: applies
model.py - RWKV model → Channel-mix layer: applies
model.py - Time-mix recurrence → WKV acceleration: dispatches WKV
model.py