---
library_name: ctranslate2
license: mit
base_model:
  - microsoft/phi-4
base_model_relation: quantized
tags:
  - ctranslate2
  - phi-4
  - chat
---

Ctranslate2 convesion of Phi-4

# Example Usage

<details><summary>Non-Streaming Example:</summary>

```python
import ctranslate2
from transformers import AutoTokenizer

def generate_response(prompt, system_message, model_path):
    generator = ctranslate2.Generator(
        model_path,
        device="cuda",
        compute_type="int8"
    )
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    formatted_prompt = f"""<|im_start|>system<|im_sep|>{system_message}<|im_end|>
<|im_start|>user<|im_sep|>{prompt}<|im_end|>
<|im_start|>assistant<|im_sep|>"""
    tokens = tokenizer.convert_ids_to_tokens(tokenizer.encode(formatted_prompt))
    results = generator.generate_batch(
        [tokens],
        max_length=1024,
        sampling_temperature=0.7
    )
    response = tokenizer.decode(results[0].sequences_ids[0], skip_special_tokens=True)
    return response

if __name__ == "__main__":
    model_path = "path/to/your/phi-4-ct2-model"
    system_message = "You are a helpful AI assistant."
    user_prompt = "Write a short poem about a cat."
    response = generate_response(user_prompt, system_message, model_path)
    print("\nGenerated response:")
    print(response)

```
</details>

<details><summary>Streaming Example:</summary>

```python
import ctranslate2
from transformers import AutoTokenizer
import sys

def generate_response(prompt, system_message, model_path):
    """
    Generates and streams a response from an AI assistant.

    Initializes the CTranslate2 generator and tokenizer, formats the input prompt,
    tokenizes it, and streams the generated tokens by printing them as they are produced.

    Parameters:
        prompt (str): The user's input prompt.
        system_message (str): The system-level instruction.
        model_path (str): Path to the CTranslate2 model directory.
    """
    generator = ctranslate2.Generator(model_path, device="cuda", compute_type="int8")
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    formatted_prompt = f"""<|im_start|>system<|im_sep|>{system_message}<|im_end|>
<|im_start|>user<|im_sep|>{prompt}<|im_end|>
<|im_start|>assistant<|im_sep|>"""
    tokens = tokenizer.tokenize(formatted_prompt)
    for step in generator.generate_tokens([tokens], max_length=1024, sampling_temperature=0.7):
        token = step.tokens[0]
        decoded_token = tokenizer.decode([step.token_ids[0]])
        print(decoded_token, end="", flush=True)
        if token in tokenizer.eos_token or token in tokenizer.all_special_tokens:
            break

if __name__ == "__main__":
    model_path = "path/to/your/phi-4-ct2-model"
    system_message = "You are a helpful AI assistant."
    user_prompt = "Write a short poem about a cat."
    print("\nGenerating response:")
    generate_response(user_prompt, system_message, model_path)

```
</details>