vivacious-cloud/llama-3.2-3b-vivacious-lora

Llama 3.2 3B Instruct — Fine-Tuned via Vivacious Cloud

Trained on wholesale spot GPUs via autonomous multi-cloud routing.
76% cheaper than AWS & RunPod. Zero PyTorch boilerplate. Zero CUDA setup.

---

🎯 The Developer Flow: Try Demo → Understand → Run on Vivacious Cloud

1. Try the Demo: Test drive the VRAM calculation and 12-cloud spot arbitrage in our Hugging Face Space Simulator.
2. Understand the Savings: See how autonomous multi-cloud routing cuts training spend by up to 76% with guaranteed preflight OOM protection.
3. Run on Vivacious Cloud: Deploy your fine-tuning run with zero PyTorch code, zero CUDA drivers, and zero idle costs at vivaciouscloud.com.

---

💻 Python Inference: How to Load This Adapter

You can load and run this LoRA adapter directly using Hugging Face transformers and peft:

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

# 1. Load adapter configuration
adapter_id = "vivacious-cloud/llama-3.2-3b-vivacious-lora"
config = PeftConfig.from_pretrained(adapter_id)

# 2. Load base model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
base_model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# 3. Merge LoRA adapter weights
model = PeftModel.from_pretrained(base_model, adapter_id)

# 4. Generate text
prompt = "Explain how Vivacious Cloud stops CUDA OOM errors before instance allocation."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📊 Training Cost & Efficiency Comparison

Platform / Setup	Hourly Rate	Total Cost	Setup Time	Spot Preemption Defense
AWS EC2 (g5.2xlarge)	$3.06 / hr	$18.36	2+ hours (DevOps, quotas, drivers)	Run dies / progress lost
RunPod (A5000)	$1.89 / hr	$11.34	45 mins (Docker, template config)	Run dies / progress lost
Vast.ai (Spot)	$0.95 / hr	$5.70	30 mins (SSH keys, untrusted hosts)	Run dies / progress lost
Vivacious Cloud (Autonomous)	$0.42 / hr	₹240 (~$2.90)	0 mins (One terminal command)	Auto-checkpoint & migrate

---

💡 Intended Use & Limitations

• Intended Use: Conversational assistance, AI systems engineering instruction, and parameter-efficient fine-tuning demonstration.
• Limitations: Inherits base model limitations of Llama 3.2 3B. Recommended to run with temperature 0.7 for creative outputs and 0.2 for factual extraction.

---

💻 Fine-Tune Your Own Model in 3 Steps

# 1. Install CLI
curl -fsSL https://vivaciouscloud.com/install.sh | sh

# 2. Prepare dataset locally (₹0 cost)
vivacious prepare ./my-dataset.jsonl

# 3. Check preflight permit & deploy to cheapest GPU alive
vivacious permit <workspace> my-run --model unsloth/Llama-3.2-3B-Instruct --method lora
vivacious deploy <workspace>

---

🌐 Links

• 🚀 Official Web Console: https://vivaciouscloud.com
• 🎮 Interactive Playground: Vivacious Terminal Simulator
• 📦 CLI Repository: https://github.com/Viavcious-cloud/vivacious-cli
• 💬 Community Telegram: https://t.me/iEverYours