Qwen3-32B model quantized with torchao float8 dynamic activation and float8 weight quantization (per row granularity), by PyTorch team. Use it directly, or serve using vLLM with 47% VRAM reduction (34.54 GB needed), around 1.7x speedup and little to no accuracy impact on H100.

Inference with vLLM

Shell

# Server
VLLM_DISABLE_COMPILE_CACHE=1 vllm serve pytorch/Qwen3-32B-FP8 --tokenizer Qwen/Qwen3-32B -O3

Shell

# Client
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "pytorch/Qwen3-32B-FP8",
  "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
}'

Inference with transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "pytorch/Qwen3-32B-FP8"

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

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

Quantization Recipe

Install the required packages:

Shell

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

Use the following code to get the float8 model using torchao library:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

model_id = "Qwen/Qwen3-32B"
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, PerRow
quant_config = Float8DynamicActivationFloat8WeightConfig(granularity=PerRow())
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)

Optionally, upload to your HF hub

USER_ID = "YOUR_USER_ID"
MODEL_NAME = model_id.split("/")[-1]
save_to = f"{USER_ID}/{MODEL_NAME}-FP8"
quantized_model.push_to_hub(save_to, safe_serialization=False)
tokenizer.push_to_hub(save_to)

Model Quality

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

Benchmark
	Qwen3-32B	Qwen3-32B-FP8
General
mmlu	80.71	80.67
bbh	37.49	38.01
Multilingual
mgsm_en_cot_es	58.4	52.0
Math
gpqa_main_zeroshot	41.96	42.63
Overall	54.64	53.33

Reproduce Model Quality Results

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

baseline

lm_eval --model hf --model_args pretrained=Qwen/Qwen3-32B --tasks mmlu --device cuda:0 --batch_size 8

float8 dynamic quantization (FP8)

Shell

export MODEL=pytorch/Qwen3-32B-FP8
# or
# export MODEL=Qwen/Qwen3-32B
lm_eval --model hf --model_args pretrained=$MODEL --tasks mmlu --device cuda:0 --batch_size 8

Memory Usage

Memory (tested on H100)
	Qwen3-32B	Qwen3-32B-FP8
Peak Memory	65.63 GB	34.71 GB (47.1% reduction)

Reproduce Peak Memory Usage Results

Code

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-32B" # pytorch/Qwen3-32B-FP8

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="cuda:0"
)

torch.cuda.reset_peak_memory_stats()

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

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


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

Model Performance

Benchmark (Tested on H100)
	Qwen3-32B	Qwen3-32B-FP8
latency (batch_size=1)	8.93s	5.16s (1.73x speedup)
latency (batch_size=256)	33.85s	16.15s (2.10x speedup)

Note: tested with fbgemm-gpu-genai installed.

Reproduce latency benchmarks

1. Setup

Shell

git clone git@github.com:vllm-project/vllm.git
cd vllm
VLLM_USE_PRECOMPILED=1 pip install --editable .

To use fbgemm kernels:

pip install fbgemm-gpu-genai

2. Latency benchmarking

Shell

export MODEL=Qwen/Qwen3-32B # or pytorch/Qwen3-32B-FP8
VLLM_DISABLE_COMPILE_CACHE=1 vllm bench latency --input-len 256 --output-len 256 --model $MODEL --batch-size 1

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.