Qwen2-4x1.5B-v2.5

Qwen2-4x1.5B-v2.5 is a Mixture of Experts (MoE) made with the following models using LazyMergekit:

• cognitivecomputations/dolphin-2.9.3-qwen2-1.5b
• Replete-AI/Qwen2-1.5b-Instruct-Replete-Adapted
• M4-ai/Hercules-5.0-Qwen2-1.5B
• d-llm/Qwen2-1.5B-Instruct-orpo

🧩 Configuration

gate_mode: hidden
architecture: qwen
dtype: bfloat16
experts_per_token: 2
base_model: cognitivecomputations/dolphin-2.9.3-qwen2-1.5b
experts:
  - source_model: cognitivecomputations/dolphin-2.9.3-qwen2-1.5b
    positive_prompts:
    - "explain"
    - "describe"
    - "define"
    - "help"
    - "assist"
  - source_model: Replete-AI/Qwen2-1.5b-Instruct-Replete-Adapted
    positive_prompts:
    - "code"
    - "algorithm"
    - "programming"
    - "development"
    - "software"
    - "framework"
  - source_model: M4-ai/Hercules-5.0-Qwen2-1.5B
    positive_prompts:
    - "rewrite"
    - "paraphrase"
    - "translate"
    - "reword"
  - source_model: d-llm/Qwen2-1.5B-Instruct-orpo
    positive_prompts:
    - "summarize"
    - "shorten"
    - "condense"
    - "tldr"
shared_experts:
  - source_model: M4-ai/Hercules-5.0-Qwen2-1.5B
    positive_prompts: # required by Qwen MoE for "hidden" gate mode, otherwise not allowed
      - "assistant"
      - "chat"
    # (optional, but recommended:)
    residual_scale: 0.1

💻 Usage

!pip install -qU transformers bitsandbytes accelerate

from transformers import AutoTokenizer, pipeline
import torch

model = "djuna/Qwen2-4x1.5B-v2.5"

tokenizer = AutoTokenizer.from_pretrained(model)
generator = pipeline(
    "text-generation",
    model=model,
    model_kwargs={"torch_dtype": torch.float16, "load_in_4bit": True},
)

messages = [{"role": "user", "content": "Explain what a Mixture of Experts is in less than 100 words."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = generator(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])