Instructions to use AINovice2005/quantized-Phi-4-reasoning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AINovice2005/quantized-Phi-4-reasoning with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="AINovice2005/quantized-Phi-4-reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AINovice2005/quantized-Phi-4-reasoning")
model = AutoModelForCausalLM.from_pretrained("AINovice2005/quantized-Phi-4-reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use AINovice2005/quantized-Phi-4-reasoning with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AINovice2005/quantized-Phi-4-reasoning"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AINovice2005/quantized-Phi-4-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AINovice2005/quantized-Phi-4-reasoning

SGLang

How to use AINovice2005/quantized-Phi-4-reasoning with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "AINovice2005/quantized-Phi-4-reasoning" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AINovice2005/quantized-Phi-4-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "AINovice2005/quantized-Phi-4-reasoning" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AINovice2005/quantized-Phi-4-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AINovice2005/quantized-Phi-4-reasoning with Docker Model Runner:
```
docker model run hf.co/AINovice2005/quantized-Phi-4-reasoning
```

Phi-4 Reasoning Quantized

🚀 Model Description

This is an int8 quantized version of Phi-4 Reasoning, optimized using torchao for reduced memory footprint and accelerated inference. The quantization applies int8 weights with dynamic int8 activations, maintaining high task performance while enabling efficient deployment on consumer and edge hardware.

Quantization Details

Method: torchao quantization
Weight Precision: int8
Activation Precision: int8 dynamic
Technique: Symmetric mapping
Impact: Significant reduction in model size with minimal loss in reasoning, coding, and general instruction-following capabilities.

🎯 Intended Use

Fast inference in production environments with limited VRAM
Research on int8 quantization deployment performance
Tasks: general reasoning, chain-of-thought, code generation, and long-context tasks.

⚠️ Limitations

Slight degradation in performance compared to full-precision (bfloat16) models
English-centric training data; may underperform in other languages or nuanced tasks
Further finetuning or quantization-aware calibration can enhance task-specific performance.

Downloads last month: 1

Model tree for AINovice2005/quantized-Phi-4-reasoning

Base model

microsoft/phi-4

Finetuned

microsoft/Phi-4-reasoning

Quantized

(32)

this model

Collection including AINovice2005/quantized-Phi-4-reasoning

🚀 Optimized Models: torchao & Pruna Quantization

Collection

Quantized Models using torchao & Pruna for efficient inference and deployment. • 8 items • Updated Mar 12 • 1