Llama-3.3-70B-Instruct-quantized.w4a16

Model Overview

• Model Architecture: Meta-Llama-3.1
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT4
• Intended Use Cases: Intended for commercial and research use in multiple languages. Instruction tuned text only models are intended for assistant-like chat, whereas pretrained models can be adapted for a variety of natural language generation tasks. The Llama 3.3 model also supports the ability to leverage the outputs of its models to improve other models including synthetic data generation and distillation. The Llama 3.3 Community License allows for these use cases.
• Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in any other way that is prohibited by the Acceptable Use Policy and Llama 3.3 Community License. Use in languages beyond English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
• Release Date: 12/11/2024
• Version: 1.0
• License(s): llama3.3
• Model Developers: Red Hat (Neural Magic)

Model Optimizations

This model was obtained by quantizing the weights of Llama-3.3-70B-Instruct to INT4 data type. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%.

Only the weights of the linear operators within transformers blocks are quantized. Weights are quantized using a symmetric per-group scheme, with group size 128. The GPTQ algorithm is applied for quantization, as implemented in the llm-compressor library.

Deployment

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/Llama-3.3-70B-Instruct-quantized.w4a16"
number_gpus = 1

sampling_params = SamplingParams(temperature=0.6, top_p=0.9, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

prompts = tokenizer.apply_chat_template(messages, 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 aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

Creation details

This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot
from datasets import load_dataset

# Load model
model_stub = "meta-llama/Llama-3.3-70B-Instruct"
model_name = model_stub.split("/")[-1]

num_samples = 1024
max_seq_len = 8192

tokenizer = AutoTokenizer.from_pretrained(model_stub)

model = AutoModelForCausalLM.from_pretrained(
    model_stub,
    device_map="auto",
    torch_dtype="auto",
)

def preprocess_fn(example):
  return {"text": tokenizer.apply_chat_template(example["messages"], add_generation_prompt=False, tokenize=False)}

ds = load_dataset("neuralmagic/LLM_compression_calibration", split="train")
ds = ds.map(preprocess_fn)

# Configure the quantization algorithm and scheme
recipe = GPTQModifier(
    targets="Linear",
    scheme="W4A16",
    ignore=["lm_head"],
    sequential_targets=["LlamaDecoderLayer"],
    dampening_frac=0.01,
)

# Apply quantization
oneshot(
    model=model,
    dataset=ds, 
    recipe=recipe,
    max_seq_length=max_seq_len,
    num_calibration_samples=num_samples,
)

# Save to disk in compressed-tensors format
save_path = model_name + "-quantized.w4a16"
model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)
print(f"Model and tokenizer saved to: {save_path}")

Evaluation

This model was evaluated on the well-known OpenLLM v1, HumanEval, and HumanEval+ benchmarks. In all cases, model outputs were generated with the vLLM engine.

OpenLLM v1 evaluations were conducted using lm-evaluation-harness and the prompting style of Meta-Llama-3.1-Instruct-evals when available.

HumanEval and HumanEval+ evaluations were conducted using Neural Magic's fork of the EvalPlus repository.

Evaluation details

MMLU

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16",dtype=auto,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_llama \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto

MMLU-CoT

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16",dtype=auto,max_model_len=4064,max_gen_toks=1024,tensor_parallel_size=1 \
  --tasks mmlu_cot_llama \
  --apply_chat_template \
  --num_fewshot 0 \
  --batch_size auto

ARC-Challenge

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16",dtype=auto,max_model_len=3940,max_gen_toks=100,tensor_parallel_size=1 \
  --tasks arc_challenge_llama \
  --apply_chat_template \
  --num_fewshot 0 \
  --batch_size auto

GSM-8K

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16",dtype=auto,max_model_len=4096,max_gen_toks=1024,tensor_parallel_size=1 \
  --tasks gsm8k_llama \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 8 \
  --batch_size auto

Hellaswag

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1 \
  --tasks hellaswag \
  --num_fewshot 10 \
  --batch_size auto

Winogrande

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1 \
  --tasks winogrande \
  --num_fewshot 5 \
  --batch_size auto

TruthfulQA

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1 \
  --tasks truthfulqa \
  --num_fewshot 0 \
  --batch_size auto

HumanEval and HumanEval+ Generation

python3 codegen/generate.py \
  --model RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16 \
  --bs 16 \
  --temperature 0.2 \
  --n_samples 50 \
  --root "." \
  --dataset humaneval

Sanitization

python3 evalplus/sanitize.py \
  humaneval/RedHatAI--Llama-3.3-70B-Instruct-quantized.w4a16_vllm_temp_0.2

Evaluation

evalplus.evaluate \
  --dataset humaneval \
  --samples humaneval/RedHatAI--Llama-3.3-70B-Instruct-quantized.w4a16_vllm_temp_0.2-sanitized

Accuracy

Category	Benchmark	Llama-3.3-70B-Instruct	Llama-3.3-70B-Instruct-quantized.w4a16 (this model)	Recovery
OpenLLM v1	MMLU (5-shot)	81.60	80.62	98.8%
	MMLU (CoT, 0-shot)	86.58	85.81	99.1%
	ARC Challenge (0-shot)	49.23	49.49	100.5%
	GSM-8K (CoT, 8-shot, strict-match)	94.16	94.47	100.3%
	Hellaswag (10-shot)	86.49	85.97	99.4%
	Winogrande (5-shot)	84.77		%
	TruthfulQA (0-shot, mc2)	62.75	61.66	98.3%
	Average	77.94	77.49	98.3%
Coding	HumanEval pass@1	83.20	83.40	100.2%
	HumanEval+ pass@1	78.40	78.60	100.3%