--- tags: - int8 - vllm language: - en - de - fr - it - pt - hi - es - th pipeline_tag: text-generation license: llama3.1 base_model: meta-llama/Meta-Llama-3.1-70B-Instruct --- # Meta-Llama-3.1-70B-Instruct-quantized.w8a8 ## Model Overview - **Model Architecture:** LlamaForCausalLM - **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 [Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-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:** 7/29/2024 - **Version:** 1.0 - **License(s):** Llama3.1 - **Model Developers:** Neural Magic This model is a quantized version of [Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-Instruct). It was evaluated on a several tasks to assess the its quality in comparison to the unquatized model, including multiple-choice, math reasoning, and open-ended text generation. Meta-Llama-3.1-70B-Instruct-quantized.w8a8 achieves 98.8% recovery for the Arena-Hard evaluation, 99.9% for OpenLLM v1 (using Meta's prompting when available), 100.0% for OpenLLM v2, 98.7% for HumanEval pass@1, and 98.9% for HumanEval+ pass@1. ### Model Optimizations This model was obtained by quantizing the weights of [Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-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 [GPTQ](https://arxiv.org/abs/2210.17323) algorithm is applied for quantization, as implemented in the [llm-compressor](https://github.com/vllm-project/llm-compressor) library. GPTQ used a 10% damping factor and 256 sequences sequences taken from Neural Magic's [LLM compression calibration dataset](https://huggingface.co/datasets/neuralmagic/LLM_compression_calibration). ## Deployment This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below. ```python from vllm import LLM, SamplingParams from transformers import AutoTokenizer model_id = "neuralmagic/Meta-Llama-3.1-70B-Instruct-quantized.w8a8" number_gpus = 2 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](https://docs.vllm.ai/en/latest/) for more details. ## Creation This model was created by using the [llm-compressor](https://github.com/vllm-project/llm-compressor) library as presented in the code snipet below. ```python from transformers import AutoTokenizer, AutoModelForCausalLM from datasets import load_dataset from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot from llmcompressor.modifiers.quantization import GPTQModifier from llmcompressor.transformers.compression.helpers import calculate_offload_device_map model_stub = "meta-llama/Meta-Llama-3.1-70B-Instruct" model_name = model_stub.split("/")[-1] num_samples = 256 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.map(preprocess_fn) recipe = GPTQModifier( targets="Linear", scheme="W8A8", ignore=["lm_head"], dampening_frac=0.1, ) device_map = calculate_offload_device_map( model_stub, reserve_for_hessians=True, num_gpus=2, torch_dtype="auto", ) model = AutoModelForCausalLM.from_pretrained( model_stub, device_map="auto", dtype="auto", ) oneshot( model=model, dataset=ds, recipe=recipe, max_seq_length=max_seq_len, num_calibration_samples=num_samples, ) save_path = model_name + "-quantized.w8a8 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 Arena-Hard, OpenLLM v1, OpenLLM v2, HumanEval, and HumanEval+ benchmarks. In all cases, model outputs were generated with the [vLLM](https://docs.vllm.ai/en/stable/) engine. Arena-Hard evaluations were conducted using the [Arena-Hard-Auto](https://github.com/lmarena/arena-hard-auto) repository. The model generated a single answer for each prompt form Arena-Hard, and each answer was judged twice by GPT-4. We report below the scores obtained in each judgement and the average. OpenLLM v1 and v2 evaluations were conducted using Neural Magic's fork of [lm-evaluation-harness](https://github.com/neuralmagic/lm-evaluation-harness/tree/llama_3.1_instruct) (branch llama_3.1_instruct). 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](https://huggingface.co/datasets/meta-llama/Meta-Llama-3.1-70B-Instruct-evals) and a few fixes to OpenLLM v2 tasks. HumanEval and HumanEval+ evaluations were conducted using Neural Magic's fork of the [EvalPlus](https://github.com/neuralmagic/evalplus) repository. Detailed model outputs are available as HuggingFace datasets for [Arena-Hard](https://huggingface.co/datasets/neuralmagic/quantized-llama-3.1-arena-hard-evals), [OpenLLM v2](https://huggingface.co/datasets/neuralmagic/quantized-llama-3.1-leaderboard-v2-evals), and [HumanEval](https://huggingface.co/datasets/neuralmagic/quantized-llama-3.1-humaneval-evals). **Note:** Results have been updated after Meta modified the chat template. ### Accuracy
Benchmark | Meta-Llama-3.1-70B-Instruct | Meta-Llama-3.1-70B-Instruct-quantized.w8a8 (this model) | Recovery |
Arena Hard | 57.0 (55.8 / 58.2) | 56.3 (56.0 / 56.6) | 98.8% |
OpenLLM v1 | |||
MMLU (5-shot) | 83.9 | 83.7 | 99.7% |
MMLU (CoT, 0-shot) | 86.2 | 85.8 | 99.5% |
ARC Challenge (0-shot) | 93.3 | 93.1 | 99.7% |
GSM-8K (CoT, 8-shot, strict-match) | 95.4 | 94.2 | 98.8% |
Hellaswag (10-shot) | 86.7 | 86.7 | 100.0% |
Winogrande (5-shot) | 85.3 | 85.1 | 100.1% |
TruthfulQA (0-shot, mc2) | 60.7 | 61.4 | 101.3% |
Average | 84.5 | 84.3 | 99.9% |
OpenLLM v2 | |||
MMLU-Pro (5-shot) | 48.1 | 47.1 | 97.9% |
IFEval (0-shot) | 86.4 | 86.6 | 100.2% |
BBH (3-shot) | 55.8 | 55.2 | 98.9% |
Math-|v|-5 (4-shot) | 26.1 | 23.9 | 91.8% |
GPQA (0-shot) | 15.4 | 13.6 | 88.4% |
MuSR (0-shot) | 18.2 | 16.8 | 92.6% |
Average | 41.7 | 40.5 | 97.3% |
Coding | |||
HumanEval pass@1 | 79.7 | 78.7 | 98.7% |
HumanEval+ pass@1 | 74.8 | 74.0 | 98.9% |