đ Qwen3-235B-A22B-FP8-dynamic
A quantized large language model based on Qwen3-235B-A22B, optimized with FP8 quantization for efficient deployment.
đ Quick Start
This section provides a quick guide on how to use the Qwen3-235B-A22B-FP8-dynamic
model.
Deployment Example
The model can be efficiently deployed using the vLLM backend. Here is an example code snippet:
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
model_id = "RedHatAI/Qwen3-235B-A22B-FP8-dynamic"
number_gpus = 4
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)
vLLM also supports OpenAI-compatible serving. For more details, refer to the documentation.
⨠Features
Model Overview
- Model Architecture: Qwen3MoeForCausalLM
- 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: 05/05/2025
- Version: 1.0
- Model Developers: RedHat (Neural Magic)
Model Optimizations
This model was obtained by quantizing activations and weights of Qwen3-235B-A22B 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 library is used for quantization.
đĻ Installation
This section does not contain specific installation steps, so it is skipped.
đģ Usage Examples
Basic Usage
The deployment example above shows the basic usage of the model. Here is the code again for reference:
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
model_id = "RedHatAI/Qwen3-235B-A22B-FP8-dynamic"
number_gpus = 4
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 example provided in the original document, so this section is skipped.
đ Documentation
Creation
Creation details
This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot
from transformers import AutoModelForCausalLM, AutoTokenizer
model_stub = "Qwen/Qwen3-235B-A22B"
model_name = model_stub.split("/")[-1]
model = AutoModelForCausalLM.from_pretrained(model_stub)
tokenizer = AutoTokenizer.from_pretrained(model_stub)
recipe = QuantizationModifier(
ignore=["lm_head"],
targets="Linear",
scheme="FP8_dynamic",
)
oneshot(
model=model,
recipe=recipe,
)
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 (version 1), using lm-evaluation-harness and vLLM.
Evaluation details
lm_eval \
--model vllm \
--model_args pretrained="RedHatAI/Qwen3-235B-A22B-FP8-dynamic",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=4 \
--tasks openllm \
--apply_chat_template\
--fewshot_as_multiturn \
--batch_size auto
Accuracy
Category |
Benchmark |
Qwen3-235B-A22B |
Qwen3-235B-A22B-FP8-dynamic (this model) |
Recovery |
OpenLLM v1 |
MMLU (5-shot) |
84.77 |
84.61 |
99.8% |
OpenLLM v1 |
ARC Challenge (25-shot) |
71.84 |
70.90 |
98.7% |
OpenLLM v1 |
GSM-8K (5-shot, strict-match) |
74.22 |
74.98 |
101.0% |
OpenLLM v1 |
Hellaswag (10-shot) |
76.56 |
76.10 |
99.4% |
OpenLLM v1 |
Winogrande (5-shot) |
73.95 |
75.06 |
101.5% |
OpenLLM v1 |
TruthfulQA (0-shot, mc2) |
61.18 |
60.93 |
99.6% |
OpenLLM v1 |
Average |
73.75 |
73.76 |
100.0% |
đ§ Technical Details
There is no specific technical details section in the original document, so this section is skipped.
đ License
The model is released under the apache-2.0
license.