Phi-3.5-mini-instruct-OpenVINO-INT4

Phi-3.5-mini-instruct-OpenVINO-INT4

This repository contains an optimized OpenVINO™ IR version of Microsoft's Phi-3.5-mini-instruct, quantized to INT4 precision using NNCF. This version is specifically tuned for low-latency, high-efficiency local inference on Intel hardware and compatible Windows-based workstations.

---

🐍 Python Inference (Optimum-Intel)

To run this model locally using the optimum-intel library:

from optimum.intel import OVModelForCausalLM
from transformers import AutoTokenizer

model_id = "CelesteImperia/Phi-3.5-mini-instruct-OpenVINO-INT4"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = OVModelForCausalLM.from_pretrained(model_id)

prompt = "Explain the core benefits of the Phi-3.5 architecture for edge AI."
messages = [
    {"role": "user", "content": prompt},
]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")

outputs = model.generate(input_ids, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

💻 For C# / .NET Users (LLamaSharp Implementation)

using LLama.Common;
using LLama;

// 1. Initialize the OpenVINO Model for Phi-3.5
var parameters = new ModelParams("path/to/openvino_model.xml")
{
ContextSize = 4096,
GpuLayerCount = 0 // Enforces OpenVINO execution on Intel hardware
};

// 2. Setup Executor
using var weights = LLamaWeights.LoadFromFile(parameters);
using var context = weights.CreateContext(parameters);
var executor = new StatelessExecutor(weights, parameters);

// 3. Inference Loop
var chatHistory = new ChatHistory();
chatHistory.AddMessage(AuthorRole.User, "Explain Phi-3.5's efficiency in a .NET environment.");

foreach (var text in executor.InferAsync(chatHistory, new InferenceParams { MaxTokens = 256 }))
{
Console.Write(text);
}

---

🏗️ Technical Details

• Optimization Tool: NNCF (Neural Network Compression Framework)
• Quantization: INT4 Asymmetric
• Precision: INT4
• Workstation Validation: Dual-GPU (RTX 3090 + RTX A4000)
• Infrastructure: S: NVMe Scratch / K: 12TB Warehouse

---

☕ Support the Forge

Maintaining a dual-GPU AI workstation and hosting high-bandwidth models requires significant resources. If our open-source tools power your projects, consider supporting our development:

Scan to support via UPI (India Only):

<img src="https://huggingface.co/datasets/CelesteImperia/Assets/resolve/main/QrCode.jpeg" width="200">

---

📜 License

This model is released under the MIT License.

---

Connect with the architect: Abhishek Jaiswal on LinkedIn