Instructions to use wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic")
model = AutoModelForCausalLM.from_pretrained("wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic

SGLang

How to use wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic with Docker Model Runner:
```
docker model run hf.co/wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic
```

Qwen2.5-7B-Instruct-FP8-Dynamic

An FP8 quantization of Qwen/Qwen2.5-7B-Instruct, produced with llm-compressor for the benchmark project vllm-serve-bench. Not an official Qwen release.

Scheme

Weights: every Linear layer is FP8 (E4M3) with one static scale per output channel.
Activations: FP8, with a scale computed per token at run time (dynamic), so no calibration data was used.
Kept in bf16: lm_head. It maps to the full vocabulary, and its errors land directly on the output distribution.
KV cache: not quantized.
Format: compressed-tensors. vLLM reads the scheme from config.json, so no --quantization flag is needed.

recipe.yaml is the exact llm-compressor recipe. provenance.json records the source snapshot, tool versions, device and time taken.

Reproduce

pip install -r scripts/quantize/requirements.txt   # pinned toolchain
python scripts/quantize/quantize.py --model Qwen/Qwen2.5-7B-Instruct --out <dir>

Both files are in the repository.

Serve

vllm serve wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic

Tested with vLLM 0.29.0 on an RTX 5090 (Blackwell, sm_120), where vLLM selects a CUTLASS FP8 kernel.

Measured

Serving performance against the bf16 original, with the same engine flags, sweep and SLOs, is in docs/03-results.md (Experiment B2). Every figure there is generated from committed raw results.

A quality smoke check is in results/quality/report.md: greedy decoding on a small fixed set of exact-answer and open prompts, compared with bf16.

Limitations

The quality check is a smoke test on a small hand-written prompt set, scored by containment. It is not an evaluation. Use a proper eval suite before relying on this checkpoint.
FP8 kernels need hardware support (Ada, Hopper, Blackwell). Elsewhere vLLM may fall back or refuse.
Performance results come from one consumer GPU. They do not transfer to other hardware.

License

Apache 2.0, as for the base model.

Downloads last month: 34

Safetensors

Model size

8B params

Tensor type

BF16

F8_E4M3

Model tree for wrbooth/Qwen2.5-7B-Instruct-FP8-Dynamic

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Quantized

(433)

this model