README.md

metadata

language:
  - vi
  - en
license: apache-2.0
library_name: transformers
tags:
  - moe
  - mixture-of-experts
  - text-generation
  - decode-series
  - llm
  - vietnamese-llm
datasets:
  - markov-ai/computer-use-large
metrics:
  - loss
  - perplexity
model-index:
  - name: Decode-12B-MoE
    results: []

🚀 Decode-12B-MoE: High-Performance Mixture of Experts Model

Decode-12B-MoE is a Large Language Model (LLM) utilizing a Sparse Mixture of Experts (MoE) architecture with a total of 12.5 billion parameters. This model is engineered to bridge the gap between massive parameter counts and computational efficiency, activating only a fraction of its weights (~2.5B) during inference.

📌 Technical Specifications

Attribute	Value
Total Parameters	12,500,340,736 (12.5B)
Active Parameters	~2.5B per token
Architecture	Sparse MoE (Decoder-only)
Context Window	4096 tokens
Format	Bfloat16 / Float16
Training Hardware	NVIDIA Tesla T4 (Prototyping) / [Your_Main_GPU]

🛠 Training Methodology

The model was trained with advanced memory optimization techniques to ensure stability on consumer and enterprise-grade hardware:

• 8-bit Optimizer: Utilized bitsandbytes AdamW to reduce optimizer state memory footprint by 75%.
• Gradient Checkpointing: Enabled to manage activation memory for deep MoE layers.
• Dataset: Fine-tuned on a diverse corpus of Vietnamese and English text, focusing on reasoning, logic, and natural conversation.

💻 Quick Start (Usage)

To use this model, ensure you have transformers and accelerate installed.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Replace with your actual Hugging Face repo ID
model_id = "your-username/decode-12b-moe"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True # Required for custom MoE architectures
)

# Test Prompt
prompt = "Explain the concept of Quantum Computing in simple terms."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=512, 
        temperature=0.7, 
        top_p=0.9,
        do_sample=True
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))