Qwen3 32B FP8 Dynamic

🚀 Qwen3-32B-FP8-dynamic

This project is based on the Qwen3-32B model, which uses FP8 quantization technology to optimize the model, reducing GPU memory requirements and increasing compute throughput.

✨ Features

• Model Architecture: Based on Qwen3ForCausalLM, it takes text as input and outputs text.
• Model Optimizations: Both weights and activations are quantized to FP8, reducing GPU memory requirements by approximately 50% and increasing matrix - multiply compute throughput by approximately 2x.
• Intended Use Cases: Suitable for reasoning, function calling, fine - tuning for subject matter experts, multilingual instruction following, and translation.

📦 Installation

There is no specific installation content provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "RedHatAI/Qwen3-32B-FP8-dynamic"
number_gpus = 1
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, top_k=20, min_p=0, max_tokens=256)

messages = [
    {"role": "user", "content": prompt}
]

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [{"role": "user", "content": "Give me a short introduction to large language model."}]

prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

Advanced Usage

There is no advanced usage content provided in the original document, so this part is not added.

📚 Documentation

Model Overview

• Model Architecture: Qwen3ForCausalLM
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Activation quantization: FP8
  • Weight quantization: FP8
• Intended Use Cases:
  • Reasoning.
  • Function calling.
  • Subject matter experts via fine - tuning.
  • Multilingual instruction following.
  • Translation.
• Out - of - scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws).
• Release Date: 05/02/2025
• Version: 1.0
• Model Developers: RedHat (Neural Magic)

Model Optimizations

This model was obtained by quantizing activations and weights of [Qwen3 - 32B](https://huggingface.co/Qwen/Qwen3 - 32B) to FP8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix - multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size requirements by approximately 50%.

Only weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized with a symmetric static per - channel scheme, whereas activations are quantized with a symmetric dynamic per - token scheme. The [llm - compressor](https://github.com/vllm - project/llm - compressor) library is used for quantization.

Deployment

This model can be deployed efficiently using the vLLM backend. vLLM also supports OpenAI - compatible serving. See the documentation for more details.

Creation

from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model
model_stub = "Qwen/Qwen3-32B"
model_name = model_stub.split("/")[-1]

model = AutoModelForCausalLM.from_pretrained(model_stub)

tokenizer = AutoTokenizer.from_pretrained(model_stub)

# Configure the quantization algorithm and scheme
recipe = QuantizationModifier(
    ignore=["lm_head"],
    targets="Linear",
    scheme="FP8_dynamic",
)

# Apply quantization
oneshot(
    model=model,
    recipe=recipe,
)

# Save to disk in compressed-tensors format
save_path = model_name + "-FP8-dynamic"
model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)
print(f"Model and tokenizer saved to: {save_path}")

Evaluation

The model was evaluated on the OpenLLM leaderboard tasks (versions 1 and 2), using [lm - evaluation - harness](https://github.com/EleutherAI/lm - evaluation - harness), and on reasoning tasks using lighteval. vLLM was used for all evaluations.

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Qwen3-32B-FP8-dynamic",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks openllm \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Qwen3-32B-FP8-dynamic",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks mgsm \
  --apply_chat_template\
  --batch_size auto

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Qwen3-32B-FP8-dynamic",dtype=auto,gpu_memory_utilization=0.5,max_model_len=16384,enable_chunk_prefill=True,tensor_parallel_size=2 \
  --tasks leaderboard \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

model_parameters:
  model_name: RedHatAI/Qwen3-32B-FP8-dynamic
  dtype: auto
  gpu_memory_utilization: 0.9
  tensor_parallel_size: 2
  max_model_length: 40960
  generation_parameters:
    temperature: 0.6
    top_k: 20
    min_p: 0.0
    top_p: 0.95
    max_new_tokens: 32768

lighteval vllm \
  --model_args lighteval_model_arguments.yaml \
  --tasks lighteval|aime24|0|0 \
  --use_chat_template = true

lighteval vllm \
  --model_args lighteval_model_arguments.yaml \
  --tasks lighteval|aime25|0|0 \
  --use_chat_template = true

lighteval vllm \
  --model_args lighteval_model_arguments.yaml \
  --tasks lighteval|math_500|0|0 \
  --use_chat_template = true

lighteval vllm \
  --model_args lighteval_model_arguments.yaml \
  --tasks lighteval|gpqa:diamond|0|0 \
  --use_chat_template = true

lighteval vllm \
  --model_args lighteval_model_arguments.yaml \
  --tasks extended|lcb:codegeneration \
  --use_chat_template = true

Accuracy

🔧 Technical Details

Only weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized with a symmetric static per - channel scheme, whereas activations are quantized with a symmetric dynamic per - token scheme. The [llm - compressor](https://github.com/vllm - project/llm - compressor) library is used for quantization.

📄 License

The model is under the apache - 2.0 license.