Phi-4-mini-instruct-AWQ-INT4

This repository hosts the Phi4-mini-instruct model quantized with torchao using int4 weight-only quantization and the awq algorithm. This work is brought to you by the PyTorch team. This model can be used directly or served using vLLM for 56% VRAM reduction (3.95 GB needed) and 1.17x speedup on H100 GPUs. The model is calibrated with 2 samples from mmlu_pro task to recover the accuracy for mmlu_pro specifically. It recovered accuracy from mmlu_pro from INT4 checkpoint from 36.98 to 43.13, while bfloat16 baseline accuracy is 46.43.

Inference with vLLM

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

pip install --pre torchao torch vllm --index-url https://download.pytorch.org/whl/nightly/cu128

Serving

Then we can serve with the following command:

# Server
export MODEL=pytorch/Phi-4-mini-instruct-AWQ-INT4
VLLM_DISABLE_COMPILE_CACHE=1 vllm serve $MODEL --tokenizer $MODEL -O3

# Client
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "pytorch/Phi-4-mini-instruct-AWQ-INT4",
  "messages": [
    {"role": "user", "content": "Give me a short introduction to large language models."}
  ],
  "temperature": 0.6,
  "top_p": 0.95,
  "top_k": 20,
  "max_tokens": 32768
}'

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.

Inference with Transformers

Install the required packages:

# for compatibility with modeling file in checkpoint
pip install transformers==4.53.0
pip install --pre torchao torch --index-url https://download.pytorch.org/whl/nightly/cu126
pip install accelerate

Example:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "pytorch/Phi-4-mini-instruct-AWQ-INT4"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("")

print("thinking content:", thinking_content)
print("content:", content)

Quantization Recipe

Install the required packages:

# for compatibility with modeling file in checkpoint
pip install transformers==4.53.0
pip install accelerate
pip install --pre torchao torch --index-url https://download.pytorch.org/whl/nightly/cu128

Use the following code to get the quantized model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

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


from torchao.quantization import Int4WeightOnlyConfig, quantize_
from torchao.prototype.awq import (
    AWQConfig,
)
from torchao._models._eval import TransformerEvalWrapper
model = AutoModelForCausalLM.from_pretrained(
    model_to_quantize,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Note: this is only compatible with H100
base_config = Int4WeightOnlyConfig(group_size=128)
# for A100, please use the following for base_config:
# base_config = Int4WeightOnlyConfig(group_size=128, int4_packing_format="tile_packed_to_4d", int4_choose_qparams_algorithm="hqq")
quant_config = AWQConfig(base_config, step="prepare")
quantize_(
    model,
    quant_config,
)
tasks = ["mmlu_pro"]
calibration_limit = 2
TransformerEvalWrapper(
    model=model,
    tokenizer=tokenizer,
    max_seq_length=max_seq_length,
).run_eval(
    tasks=tasks,
    limit=calibration_limit,
)
quant_config = AWQConfig(base_config, step="convert")
quantize_(model, quant_config)

quantized_model = model
quant_config = AWQConfig(base_config, step="prepare_for_loading")
quantized_model.config.quantization_config = TorchAoConfig(quant_config)


# Push to hub
USER_ID = "YOUR_USER_ID"
MODEL_NAME = model_id.split("/")[-1]
save_to = f"{USER_ID}/{MODEL_NAME}-AWQ-INT4"
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. Here we only run on mmlu for sanity check.

Since the checkpoint is tuned on mmlu_pro, we check against the accuracy for mmlu_pro:

<details> <summary> Reproduce Model Quality Results </summary>

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 mmlu --device cuda:0 --batch_size 8

AWQ-INT4

export MODEL=pytorch/Phi-4-mini-instruct-AWQ-INT4
lm_eval --model hf --model_args pretrained=$MODEL --tasks mmlu --device cuda:0 --batch_size 8

</details>

Peak Memory Usage

Results

<details> <summary> Reproduce Peak Memory Usage Results </summary>

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-AWQ-INT4"
model_id = "pytorch/Phi-4-mini-instruct-AWQ-INT4"
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cuda:0", 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")

</details>

Model Performance

Results (H100 machine)

Note: it's expected that the awq-int4 checkpoint is slower when batch size is 256 since the problem is not memory bound but becomes compute bound when batch size is larger, while int4 weight only checkpoint is only expected to have speedup for memory bound situations. Note: we are comparing to jerryzh168/Phi-4-mini-instruct-INT4 which is a checkpoint for H100, since the AWQ-INT4 is using the new INT4 config that's optimized for H100 that doesn't regress the performance for batch size 256. It's possible to generate AWQ-INT4 for A100 as well using Int4WeightOnlyConfig(group_size=128, int4_packing_foramt="tile_packed_to_4d", int4_choose_qparams_algorithm="hqq")

<details> <summary> Reproduce Model Performance Results </summary>

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

export MODEL=microsoft/Phi-4-mini-instruct
vllm bench latency --input-len 256 --output-len 256 --model $MODEL --batch-size 1

AWQ-INT4

export MODEL=pytorch/Phi-4-mini-instruct-AWQ-INT4
VLLM_DISABLE_COMPILE_CACHE=1 vllm bench latency --input-len 256 --output-len 256 --model $MODEL --batch-size 1

</details>

Paper: TorchAO: PyTorch-Native Training-to-Serving Model Optimization

The model's quantization is powered by TorchAO, a framework presented in the paper TorchAO: PyTorch-Native Training-to-Serving Model Optimization.

Abstract: We present TorchAO, a PyTorch-native model optimization framework leveraging quantization and sparsity to provide an end-to-end, training-to-serving workflow for AI models. TorchAO supports a variety of popular model optimization techniques, including FP8 quantized training, quantization-aware training (QAT), post-training quantization (PTQ), and 2:4 sparsity, and leverages a novel tensor subclass abstraction to represent a variety of widely-used, backend agnostic low precision data types, including INT4, INT8, FP8, MXFP4, MXFP6, and MXFP8. TorchAO integrates closely with the broader ecosystem at each step of the model optimization pipeline, from pre-training (TorchTitan) to fine-tuning (TorchTune, Axolotl) to serving (HuggingFace, vLLM, SGLang, ExecuTorch), connecting an otherwise fragmented space in a single, unified workflow. TorchAO has enabled recent launches of the quantized Llama 3.2 1B/3B and LlamaGuard3-8B models and is open-source at this https URL .

Resources

• Official TorchAO GitHub Repository: https://github.com/pytorch/ao
• TorchAO Documentation: https://docs.pytorch.org/ao/stable/index.html

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.