pytorch/Phi-4-mini-instruct-int4wo-hqq

Phi4-mini quantized with torchao int4 weight only quantization, using hqq algorithm for improved accuracy, by PyTorch team. Use it directly or serve using vLLM for 67% VRAM reduction and 12-20% speedup on A100 GPUs.

Inference with vLLM

Install vllm nightly and torchao nightly to get some recent changes:

pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
pip install git+https://github.com/pytorch/ao.git

Code Example

from vllm import LLM, SamplingParams

# Sample prompts.
prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)


if __name__ == '__main__':
    # Create an LLM.
    llm = LLM(model="pytorch/Phi-4-mini-instruct-int4wo-hqq")
    # Generate texts from the prompts.
    # The output is a list of RequestOutput objects
    # that contain the prompt, generated text, and other information.
    outputs = llm.generate(prompts, sampling_params)
    # Print the outputs.
    print("\nGenerated Outputs:\n" + "-" * 60)
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt:    {prompt!r}")
        print(f"Output:    {generated_text!r}")
        print("-" * 60)

Note: please use VLLM_DISABLE_COMPILE_CACHE=1 to disable compile cache when running this code, e.g. VLLM_DISABLE_COMPILE_CACHE=1 python example.py, since there are some issues with the composability of compile in vLLM and torchao, this is expected be resolved in pytorch 2.8.

Serving

Then we can serve with the following command:

vllm serve pytorch/Phi-4-mini-instruct-int4wo-hqq --tokenizer microsoft/Phi-4-mini-instruct -O3

Inference with Transformers

Install the required packages:

pip install git+https://github.com/huggingface/transformers@main
pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
pip install torch
pip install accelerate

Example:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
 
torch.random.manual_seed(0)

model_path = "pytorch/Phi-4-mini-instruct-int4wo-hqq"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_path)
 
messages = [
    {"role": "system", "content": "You are a helpful AI assistant."},
    {"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"},
    {"role": "assistant", "content": "Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."},
    {"role": "user", "content": "What about solving an 2x + 3 = 7 equation?"},
]
 
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
)
 
generation_args = {
    "max_new_tokens": 500,
    "return_full_text": False,
    "temperature": 0.0,
    "do_sample": False,
}
 
output = pipe(messages, **generation_args)
print(output[0]['generated_text'])

Quantization Recipe

Install the required packages:

pip install git+https://github.com/huggingface/transformers@main
pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
pip install torch
pip install accelerate

Use the following code to get the quantized model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

model_id = "microsoft/Phi-4-mini-instruct"

from torchao.quantization import Int4WeightOnlyConfig
quant_config = Int4WeightOnlyConfig(group_size=128, use_hqq=True)
quantization_config = TorchAoConfig(quant_type=quant_config)
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16, quantization_config=quantization_config)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Push to hub
USER_ID = "YOUR_USER_ID"
MODEL_NAME = model_id.split("/")[-1]
save_to = f"{USER_ID}/{MODEL_NAME}-int4wo-hqq"
quantized_model.push_to_hub(save_to, safe_serialization=False)
tokenizer.push_to_hub(save_to)

# Manual Testing
prompt = "Hey, are you conscious? Can you talk to me?"
messages = [
    {
        "role": "system",
        "content": "",
    },
    {"role": "user", "content": prompt},
]
templated_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
print("Prompt:", prompt)
print("Templated prompt:", templated_prompt)
inputs = tokenizer(
    templated_prompt,
    return_tensors="pt",
).to("cuda")
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt):])

Note: to push_to_hub you need to run

pip install -U "huggingface_hub[cli]"
huggingface-cli login

and use a token with write access, from https://huggingface.co/settings/tokens

Model Quality

We rely on lm-evaluation-harness to evaluate the quality of the quantized model.

Need to install lm-eval from source: https://github.com/EleutherAI/lm-evaluation-harness#install

baseline

lm_eval --model hf --model_args pretrained=microsoft/Phi-4-mini-instruct --tasks hellaswag --device cuda:0 --batch_size 8

int4 weight only quantization with hqq (int4wo-hqq)

lm_eval --model hf --model_args pretrained=pytorch/Phi-4-mini-instruct-int4wo-hqq --tasks hellaswag --device cuda:0 --batch_size 8

Benchmark
	Phi-4-mini-ins	Phi-4-mini-ins-int4wo-hqq
Popular aggregated benchmark
mmlu (0-shot)	66.73	63.56
mmlu_pro (5-shot)	46.43	36.74
Reasoning
arc_challenge (0-shot)	56.91	54.86
gpqa_main_zeroshot	30.13	30.58
HellaSwag	54.57	53.54
openbookqa	33.00	34.40
piqa (0-shot)	77.64	76.33
social_iqa	49.59	47.90
truthfulqa_mc2 (0-shot)	48.39	46.44
winogrande (0-shot)	71.11	71.51
Multilingual
mgsm_en_cot_en	60.8	59.6
Math
gsm8k (5-shot)	81.88	74.37
mathqa (0-shot)	42.31	42.75
Overall	55.35	53.28

Peak Memory Usage

Results

Benchmark
	Phi-4 mini-Ins	Phi-4-mini-instruct-int4wo-hqq
Peak Memory (GB)	8.91	2.98 (67% reduction)

Code Example

We can use the following code to get a sense of peak memory usage during inference:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

# use "microsoft/Phi-4-mini-instruct" or "pytorch/Phi-4-mini-instruct-int4wo-hqq"
model_id = "pytorch/Phi-4-mini-instruct-int4wo-hqq"
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_id)

torch.cuda.reset_peak_memory_stats()

prompt = "Hey, are you conscious? Can you talk to me?"
messages = [
    {
        "role": "system",
        "content": "",
    },
    {"role": "user", "content": prompt},
]
templated_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
print("Prompt:", prompt)
print("Templated prompt:", templated_prompt)
inputs = tokenizer(
    templated_prompt,
    return_tensors="pt",
).to("cuda")
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt):])

mem = torch.cuda.max_memory_reserved() / 1e9
print(f"Peak Memory Usage: {mem:.02f} GB")

Model Performance

Our int4wo is only optimized for batch size 1, so expect some slowdown with larger batch sizes, we expect this to be used in local server deployment for single or a few users where the decode tokens per second will matters more than the time to first token.

Results (A100 machine)

Benchmark (Latency)
	Phi-4 mini-Ins	phi4-mini-int4wo-hqq
latency (batch_size=1)	2.46s	2.2s (12% speedup)
serving (num_prompts=1)	0.87 req/s	1.05 req/s (20% speedup)

Note the result of latency (benchmark_latency) is in seconds, and serving (benchmark_serving) is in number of requests per second. Int4 weight only is optimized for batch size 1 and short input and output token length, please stay tuned for models optimized for larger batch sizes or longer token length.

Setup

Get vllm source code:

git clone [email protected]:vllm-project/vllm.git

Install vllm

VLLM_USE_PRECOMPILED=1 pip install --editable .

Run the benchmarks under vllm root folder:

benchmark_latency

baseline

python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model microsoft/Phi-4-mini-instruct --batch-size 1

int4wo-hqq

VLLM_DISABLE_COMPILE_CACHE=1 python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model pytorch/Phi-4-mini-instruct-int4wo-hqq --batch-size 1

benchmark_serving

We benchmarked the throughput in a serving environment.

Download sharegpt dataset:

wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json

Other datasets can be found in: https://github.com/vllm-project/vllm/tree/main/benchmarks

Note: you can change the number of prompts to be benchmarked with --num-prompts argument for benchmark_serving script.

baseline

Server:

vllm serve microsoft/Phi-4-mini-instruct --tokenizer microsoft/Phi-4-mini-instruct -O3

Client:

python benchmarks/benchmark_serving.py --backend vllm --dataset-name sharegpt --tokenizer microsoft/Phi-4-mini-instruct --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json --model microsoft/Phi-4-mini-instruct --num-prompts 1

int4wo-hqq

Server:

VLLM_DISABLE_COMPILE_CACHE=1 vllm serve pytorch/Phi-4-mini-instruct-int4wo-hqq --tokenizer microsoft/Phi-4-mini-instruct -O3 --pt-load-map-location cuda:0

Client:

python benchmarks/benchmark_serving.py --backend vllm --dataset-name sharegpt --tokenizer microsoft/Phi-4-mini-instruct --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json --model pytorch/Phi-4-mini-instruct-int4wo-hqq --num-prompts 1

Disclaimer

PyTorch has not performed safety evaluations or red teamed the quantized models. Performance characteristics, outputs, and behaviors may differ from the original models. Users are solely responsible for selecting appropriate use cases, evaluating and mitigating for accuracy, safety, and fairness, ensuring security, and complying with all applicable laws and regulations.

Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the licenses the models are released under, including any limitations of liability or disclaimers of warranties provided therein.