You can deploy this model and start earning money today!

OpenMOSE/
HRWKV7-Reka-Flash3-Preview

knowledge-distillation

Views

No views yet

HRWKV7-Reka-Flash3-Preview

I'm simply exploring the possibility of linearizing existing Transformer models. It's still far from perfect, but I hope you'll bear with me as I continue this journey. :)

Paper and Project Details

This model is part of the research presented in the paper RADLADS: Rapid Attention Distillation to Linear Attention Decoders at Scale.

The main codebase for the RADLADS project can be found at: https://github.com/recursal/RADLADS-paper

Model Description

HRWKV7-Reka-Flash3-Preview is an experimental hybrid architecture model that combines RWKV v7's linear attention mechanism with Group Query Attention (GQA) layers. Built upon the Reka-flash3 21B foundation, this model replaces most Transformer attention blocks with RWKV blocks while strategically maintaining some GQA layers to enhance performance on specific tasks.

Developed by: OpenMOSE
Model type: Hybrid Linear-Attention Language Model
Language(s): Multilingual (inherited from Reka-flash3 21B)
License: Apache-2.0
Base Model: Reka-flash3 21B(https://huggingface.co/RekaAI/reka-flash-3)
Year: 2025

Architecture Specifications

Architecture: RWKV v7 based "hxa079" Architecture + Group Query Attention Hybrid
Total Layers: 44 layers (L44D6114)
- 38 RWKV layers (with Rope)
- 6 GQA layers (No Rope, No Position Embeddings)
Hidden Dimension: 6144
Training Context Window: 4096 tokens
Inference Context Window 32768+
Training Strategy Following RADLADS method based knowledge distillation

Technical Innovation

RWKV "hxa079" Architecture

The model implements several key improvements over standard RWKV architectures:

Token Shift Removal: In order to effectively inherit the teacher model weights, we removed the residual connection one token ago.
GroupNorm Removal: Helps improve training stability issues
k_first Introduction: Experimentally adopted the approach of residually connecting k layers in layer 0.

Hybrid Design Benefits

Linear Attention Inference: RWKV blocks enable O(1) memory complexity during inference, and the hybrid approach reduces the KVCache to 1/7 of full GQA.
Enhanced Needle Tasks: Strategic placement of GQA layers significantly improves performance on needle-in-haystack retrieval tasks, addressing a known limitation of pure linear attention models
Implicit Position Encoding: Interestingly, the model achieves better performance when RoPE (Rotary Position Embedding) is not applied to GQA layers, suggesting that RWKV blocks provide implicit positional encoding capabilities

Intended Use

This is an experimental research model designed to explore hybrid architectures combining linear and quadratic attention mechanisms. It is intended for:

Research into efficient attention mechanisms
Benchmarking hybrid architecture performance
Exploring linear attention limitations and solutions
Academic and industrial R&D purposes

Limitations

Experimental Status: This model is in experimental stages and may exhibit unexpected behaviors
Context Window: Limited to 4096 tokens during training, though RWKV architecture theoretically supports longer sequences
Performance Variability: As a hybrid model, performance may vary significantly across different task types

Training Details

Training Context Window: 4096 tokens
Training GPU AMD MI300X x 1(takes 68hrs)
Training Strategy 8bit MLP Quant, frozen emb,mlp,head, Deepspeed Stage1
Base Model Initialization: Weights initialized from Reka-flash3 21B
Architecture Conversion: Transformer attention blocks systematically replaced with RWKV blocks, except for 6 strategically placed GQA layers

Evaluation

Performance evaluation is ongoing. The model shows promising results in:

Maintaining base model capabilities while achieving linear attention efficiency
Significantly improved needle-in-haystack task performance compared to pure RWKV architectures
Competitive performance on standard language modeling benchmarks

Usage with Hugging Face Transformers

This model can be loaded and used with the transformers library. Ensure you have transformers installed: pip install transformers. When loading, remember to set trust_remote_code=True because of the custom architecture.

python

from transformers import pipeline, AutoTokenizer
import torch

model_name = "OpenMOSE/HRWKV7-Reka-Flash3-Preview" # Replace with the actual model ID if different
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
pipe = pipeline(
    "text-generation",
    model_name,
    tokenizer=tokenizer,
    torch_dtype=torch.bfloat16, # or torch.float16 depending on your GPU and model precision
    device_map="auto",
    trust_remote_code=True,
)

text = "The quick brown fox jumps over the lazy "
result = pipe(text, max_new_tokens=20, do_sample=True, top_p=0.9, temperature=0.7)[0]["generated_text"]
print(result)

Run with RWKV-Infer (as provided by original authors)

RWKV-Infer now support hxa079

bash

curl http://127.0.0.1:9000/loadmodel -X POST -H "Content-Type: application/json" -d '{"model_filename":"/home/client/Projects/llm/hxa079-reka-flash3-stage2-hybrid.pth","model_viewname":"RWKV HXA079 L38T6 Reka Flash3","model_strategy":"int8","adapter_filename":"","adapter_mode":"", "template":"rekaflash3", "endtoken":"
 <sep>","default_temperature":"0.2", "default_top_p":"0.3", "rope_theta":"8000000.0", "rms_norm_eps":"1e-5"}'

Thank you for Big help :)

SmerkyG Inspired by RADLADS (https://arxiv.org/abs/2505.03005)

Training Code

https://github.com/OpenMOSE/RWKVInside (still buggy)

Model Card Contact

OpenMOSE - 2025

Note: This is an experimental model. Performance characteristics and behaviors may differ from both pure RWKV and standard Transformer architectures. Users should thoroughly evaluate the model for their specific use cases.

Citation

If you use this code or find our work valuable, please consider citing RADLADS:

bibtex

@misc{goldstein2025radladsrapidattentiondistillation,
      title={RADLADS: Rapid Attention Distillation to Linear Attention Decoders at Scale}, 
      author={Daniel Goldstein and Eric Alcaide and Janna Lu and Eugene Cheah},
      year={2025},
      eprint={2505.03005},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.03005}, 
}