Instructions to use Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171 with Transformers:

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

pipe = pipeline("text-generation", model="Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171")
model = AutoModelForCausalLM.from_pretrained("Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171", 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171

SGLang

How to use Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171 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 "Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171" \
    --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": "Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171",
		"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 "Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171" \
        --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": "Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171 with Docker Model Runner:
```
docker model run hf.co/Crusadersk/Llama-3.2-3B-Instruct-FP8-Dynamic-TR171
```

Built with Llama

llama3.2-3b — FP8_DYNAMIC (W8A8-e4m3)

Weight-FP8 checkpoint of unsloth/Llama-3.2-3B-Instruct, produced for the TR171 deployment-time safety-tax benchmark.

Provenance

Field	Value
Base model	`unsloth/Llama-3.2-3B-Instruct`
Base revision	`006f5dcd1393c3add266de40994ba96225e9689d` (INFERRED — recovered from local HF cache snapshot, not a recorded fact)
Recipe	`FP8_DYNAMIC` (W8A8-e4m3), `llmcompressor`
Quantization method	`compressed-tensors`
Calibration data	none — FP8_DYNAMIC is data-free
Build date	2026-07-02
Shard size	3.63 GB
Quantize wall time	37.7 s
Integrity record	per-file `sha256` from Hub LFS metadata; `shard_bytes` verified

Reproducing

Producer: research/tr171/expansion/fp8_support_probe.py; environment: research/tr171/expansion/Dockerfile.fp8. The recipe takes no calibration corpus, so there is no dataset or seed to reproduce — only the base checkpoint and the toolchain version.

Known reproducibility gap: llmcompressor was unpinned at build time, so the exact version used on 2026-07-02 is unrecorded. The Dockerfile now pins it. A rebuild may therefore not be bit-identical to this artifact.

Integrity, stated honestly: the 2026-07-02 build recorded no sha256 of its own, and the local build directory is now empty, so no aggregate directory digest exists for it. What is verifiable instead: the per-file sha256 below is read from this repo's Git-LFS metadata, and the mirror was checked against the build record — summing the file sizes in this repo, excluding the generated README.md, NOTICE and .gitattributes, reproduces the matrix's shard_bytes of 3,625,745,288 exactly. So these hashes describe the same bytes the probe measured, and you can verify a download against them directly:

File	sha256
`model.safetensors`	`9045dfd85fc49efd353fb1060d9bf52b5b745dc2c05e7c1595d7f7b54b91fcf4`
`tokenizer.json`	`6b9e4e7fb171f92fd137b777cc2714bf87d11576700a1dcd7a399e7bbe39537b`

LFS-tracked files only; the small JSON/text files are git blobs and carry no sha256. fp8_support_probe.py now records a real digest at build time, so future shards will not need this reconstruction.