Instructions to use amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0 with Transformers:

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

pipe = pipeline("text-generation", model="amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0")
model = AutoModelForMultimodalLM.from_pretrained("amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0")
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 amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0

SGLang

How to use amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0 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 "amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0" \
    --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": "amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0",
		"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 "amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0" \
        --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": "amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0 with Docker Model Runner:
```
docker model run hf.co/amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0
```

Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0

Model Overview

Model Architecture: LlamaForCausalLM
- Input: Text
- Output: Text
Source Model: Llama-3.3-70B-Instruct
Supported Hardware: AMD EPYC (CPU inference)
Preferred Operating System: Linux
Inference Engine: vLLM v0.22.0
Quantization Framework: LLM Compressor v0.11.0
Quantization Method: 4-bit Weight-Only Quantization (W4A16), Asymmetric
Compatible Stack:
- ZenDNN v6.0.0
- ZenTorch v2.11.0.1
- PyTorch v2.11.0
- LLM Compressor v0.11.0
- vLLM v0.22.0

ZenTorch v2.11.0.1 for PyTorch v2.11.0 has to be built from source.

This model was Built with Llama. This is a quantized version of Llama-3.3-70B-Instruct created by AMD using LLM Compressor (compressed-tensors) for ZenDNN-optimized CPU inference.

Quantization

The model was quantized from Llama-3.3-70B-Instruct using LLM Compressor v0.11.0.

Method: 4-bit Weight-Only Quantization (W4A16), Asymmetric
Config: compressed-tensors, num_bits=4, type=int, symmetric=false, group_size=128
Quantized Layers: All nn.Linear layers excluding lm_head

The model was produced using LLM Compressor as shown below:

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor import oneshot
from llmcompressor.modifiers.awq import AWQModifier

MODEL_ID   = "meta-llama/Llama-3.3-70B-Instruct"
OUTPUT_DIR = "./Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0"
NUM_CALIB, MAX_SEQ_LEN = 128, 2048

# Load the model and tokenizer (bf16)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, device_map="cpu", dtype="bfloat16", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)

# AWQ recipe: 4-bit asymmetric weight-only, group size 128, skip lm_head
recipe = [
    AWQModifier(ignore=["lm_head"], scheme="W4A16_ASYM", targets=["Linear"]),
]

# Load and tokenize calibration data (128 samples from ultrachat_200k)
calib = load_dataset("HuggingFaceH4/ultrachat_200k", split=f"train_sft[:{NUM_CALIB}]")
calib = calib.map(
    lambda ex: {"text": "\n".join(m["content"] for m in ex["messages"] if m.get("content"))},
    remove_columns=calib.column_names,
)
calib = calib.map(
    lambda ex: tokenizer(
        ex["text"], truncation=True, max_length=MAX_SEQ_LEN, add_special_tokens=False
    ),
    remove_columns=["text"],
)

# Quantize with the oneshot API
oneshot(
    model=model,
    dataset=calib,
    recipe=recipe,
    max_seq_length=MAX_SEQ_LEN,
    processor=tokenizer,
)

# Save in compressed-tensors format
model.save_pretrained(OUTPUT_DIR, save_compressed=True)
tokenizer.save_pretrained(OUTPUT_DIR)

Note: This quantization method (W4A16 asymmetric WOQ) is specific to the ZenDNN execution path.

Quick Start

Requirements

pip install \
    torch==2.11.0 \
    zentorch==2.11.0.1 \
    vllm==0.22.0 \
    huggingface_hub \
    "lm-eval[vllm]==0.4.12"

ZenTorch v2.11.0.1 for PyTorch v2.11.0 has to be built from source.

CPU runtime libraries (only needed if not already present):

conda install -c conda-forge gperftools=2.17.2=h65a8314_0 --no-deps -y
conda install -c conda-forge llvm-openmp=18.1.8=hf5423f3_1 --no-deps -y

Recommended environment variables

# vLLM CPU runtime tuning
export VLLM_USE_AOT_COMPILE=0
export VLLM_WORKER_MULTIPROC_METHOD=spawn

# TorchInductor
export TORCHINDUCTOR_FREEZING=1
export TORCHINDUCTOR_AUTOGRAD_CACHE=0

# ZenTorch / ZenDNN
export ZENDNNL_MATMUL_ALGO=1
export ZENTORCH_FUSED_MOE=1

For optimal performance, set LD_PRELOAD with libtcmalloc_minimal.so.4 (tcmalloc) and an OpenMP runtime (libiomp5.so for Intel OpenMP, or libomp.so for LLVM OpenMP):

export LD_PRELOAD=<path to lib>/libtcmalloc_minimal.so.4:<path to lib>/libiomp5.so${LD_PRELOAD:+:$LD_PRELOAD}

Locate the libraries with find / -name 'libtcmalloc_minimal.so.4' and find / -name 'libiomp5.so', then substitute the resulting directory for <path to lib>.

Note: Set LD_PRELOAD before launching vLLM or any inference script.

Evaluation

The model was evaluated against the BF16 (unquantized) baseline on GSM8K (5-shot) using lm-evaluation-harness with the vLLM engine.

Benchmark	BF16 Baseline	W4A16-Asym (this model)	Recovery
GSM8K (5-shot, flexible-extract)	-	-	-

Evaluation Command

lm_eval \
    --model vllm \
    --model_args pretrained="amd/Llama-3.3-70B-Instruct-w4a16-llmcompressor-v0.11.0",dtype=bfloat16 \
    --tasks gsm8k \
    --batch_size auto \
    --trust_remote_code \
    --num_fewshot 5 \
    --log_samples \
    --gen_kwargs "max_gen_toks=2048" \
    --apply_chat_template \
    --output_path .

Limitations

Version Lock: This model is compatible with ZenDNN v6.0.0 / PyTorch v2.11.0. It may not load correctly on other versions.
CPU Only: This model is optimized for AMD EPYC CPU inference via ZenDNN. It is not intended for GPU inference.