Llama-3.2-3B-Instruct-quantized.w8a8

Llama-3.2-3B-Instruct-quantized.w8a8

Model Overview

• Model Architecture: Llama-3
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Activation quantization: INT8
  • Weight quantization: INT8
• Intended Use Cases: Intended for commercial and research use multiple languages. Similarly to Llama-3.2-3B-Instruct, this models is intended for assistant-like chat.
• Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws).
• Release Date: 9/25/2024
• Version: 1.0
• License(s): Llama3.2
• Model Developers: Neural Magic

Quantized version of Llama-3.2-3B-Instruct. It achieves scores within 1% of the scores of the unquantized model for MMLU, ARC-Challenge, GSM-8k, Hellaswag, Winogrande and TruthfulQA.

Model Optimizations

This model was obtained by quantizing the weights of Llama-3.2-3B-Instruct to INT8 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, where a fixed linear scaling factor is applied between INT8 and floating point representations for each output channel dimension. Activations are quantized with a symmetric dynamic per-token scheme, computing a linear scaling factor at runtime for each token between INT8 and floating point representations. The SmoothQuant algorithm is used to alleviate outliers in the activations, whereas rhe GPTQ algorithm is applied for quantization. Both algorithms are implemented in the llm-compressor library. GPTQ used a 1% damping factor and 512 sequences sequences taken from Neural Magic's LLM compression calibration dataset.

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 = "neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8"
number_gpus = 1
max_model_len = 8192

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, add_generation_prompt=True, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus, max_model_len=max_model_len)

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

This model was created by using the llm-compressor library as presented in the code snipet below.

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

model_id = "meta-llama/Llama-3.2-3B-Instruct"

num_samples = 512
max_seq_len = 8192

tokenizer = AutoTokenizer.from_pretrained(model_id)

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.shuffle().select(range(num_samples))
ds = ds.map(preprocess_fn)

recipe = [
  SmoothQuantModifier(
    smoothing_strength=0.7,
    mappings=[
      [["re:.*q_proj", "re:.*k_proj", "re:.*v_proj"], "re:.*input_layernorm"],
      [["re:.*gate_proj", "re:.*up_proj"], "re:.*post_attention_layernorm"],
      [["re:.*down_proj"], "re:.*up_proj"],
    ],
  ),
  GPTQModifier(
    sequential=True,
    targets="Linear",
    scheme="W8A8",
    ignore=["lm_head"],
    dampening_frac=0.01,
  )
]

model = SparseAutoModelForCausalLM.from_pretrained(
  model_id,
  device_map="auto",
)

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

model.save_pretrained("Llama-3.2-3B-Instruct-quantized.w8a8")

Evaluation

The model was evaluated on MMLU, ARC-Challenge, GSM-8K, Hellaswag, Winogrande and TruthfulQA. Evaluation was conducted using the Neural Magic fork of lm-evaluation-harness (branch llama_3.1_instruct) and the vLLM engine. This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challenge and GSM-8K that match the prompting style of Meta-Llama-3.1-Instruct-evals.

Accuracy

Open LLM Leaderboard evaluation scores

<table> <tr> <td><strong>Benchmark</strong> </td> <td><strong>Llama-3.2-3B-Instruct </strong> </td> <td><strong>Llama-3.2-3B-Instruct-quantized.w8a8 (this model)</strong> </td> <td><strong>Recovery</strong> </td> </tr> <tr> <td>MMLU (5-shot) </td> <td>62.98 </td> <td>62.75 </td> <td>99.6% </td> </tr> <tr> <td>MMLU (CoT, 0-shot) </td> <td>65.40 </td> <td>65.05 </td> <td>99.5% </td> </tr> <tr> <td>ARC Challenge (0-shot) </td> <td>77.13 </td> <td>76.45 </td> <td>99.1% </td> </tr> <tr> <td>GSM-8K (CoT, 8-shot, strict-match) </td> <td>77.94 </td> <td>77.56 </td> <td>99.5% </td> </tr> <tr> <td>Hellaswag (10-shot) </td> <td>73.62 </td> <td>73.63 </td> <td>100.0% </td> </tr> <tr> <td>Winogrande (5-shot) </td> <td>71.11 </td> <td>71.90 </td> <td>101.1% </td> </tr> <tr> <td>TruthfulQA (0-shot, mc2) </td> <td>51.47 </td> <td>51.38 </td> <td>98.4% </td> </tr> <tr> <td><strong>Average</strong> </td> <td><strong>68.52</strong> </td> <td><strong>68.39</strong> </td> <td><strong>99.81%</strong> </td> </tr> </table>

Reproduction

The results were obtained using the following commands:

MMLU

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=3850,max_gen_toks=10,tensor_parallel_size=1 \
  --tasks mmlu_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 5 \
  --batch_size auto

MMLU-CoT

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=4064,max_gen_toks=1024,tensor_parallel_size=1 \
  --tasks mmlu_cot_0shot_llama_3.1_instruct \
  --apply_chat_template \
  --num_fewshot 0 \
  --batch_size auto

ARC-Challenge

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=3940,max_gen_toks=100,tensor_parallel_size=1 \
  --tasks arc_challenge_llama_3.1_instruct \
  --apply_chat_template \
  --num_fewshot 0 \
  --batch_size auto

GSM-8K

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=4096,max_gen_toks=1024,tensor_parallel_size=1 \
  --tasks gsm8k_cot_llama_3.1_instruct \
  --fewshot_as_multiturn \
  --apply_chat_template \
  --num_fewshot 8 \
  --batch_size auto

Hellaswag

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",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="neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",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="neuralmagic/Llama-3.2-3B-Instruct-quantized.w8a8",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1 \
  --tasks truthfulqa \
  --num_fewshot 0 \
  --batch_size auto