--- tags: - int4 - vllm language: - en - de - fr - it - pt - hi - es - th pipeline_tag: text-generation license: llama3.1 base_model: meta-llama/Meta-Llama-3.1-405B-Instruct --- # Meta-Llama-3.1-405B-Instruct-quantized.w4a16 ## Model Overview - **Model Architecture:** Meta-Llama-3 - **Input:** Text - **Output:** Text - **Model Optimizations:** - **Weight quantization:** INT4 - **Intended Use Cases:** Intended for commercial and research use in English. Similarly to [Meta-Llama-3.1-405B-Instruct](https://huggingface.co./meta-llama/Meta-Llama-3.1-405B-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). Use in languages other than English. - **Release Date:** 8/9/2024 - **Version:** 1.0 - **License(s):** Llama3.1 - **Model Developers:** Neural Magic This model is a quantized version of [Meta-Llama-3.1-405B-Instruct](https://huggingface.co./meta-llama/Meta-Llama-3.1-405B-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-3.1-405B-Instruct-quantized.w4a16 achieves 98.7% recovery for the Arena-Hard evaluation, 100.0% for OpenLLM v1 (using Meta's prompting when available), 99.0% for OpenLLM v2, 98.0% for HumanEval pass@1, and 98.5% for HumanEval+ pass@1. ### Model Optimizations This model was obtained by quantizing the weights of [Meta-Llama-3.1-405B-Instruct](https://huggingface.co./meta-llama/Meta-Llama-3.1-405B-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. Symmetric per-channel quantization is applied, in which a linear scaling per output dimension maps the INT4 and floating point representations of the quantized weights. 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 1% damping factor and 512 sequences of 4,096 random tokens. ## 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-3.1-405B-Instruct-quantized.w4a16" number_gpus = 8 max_model_len = 4096 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 from datasets import Dataset from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot from llmcompressor.modifiers.quantization import GPTQModifier import random model_id = "meta-llama/Meta-Llama-3.1-405B-Instruct" num_samples = 512 max_seq_len = 8192 tokenizer = AutoTokenizer.from_pretrained(model_id) preprocess_fn = lambda example: {"text": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n{text}".format_map(example)} dataset_name = "neuralmagic/LLM_compression_calibration" dataset = load_dataset(dataset_name, split="train") ds = dataset.shuffle().select(range(num_samples)) ds = ds.map(preprocess_fn) recipe = GPTQModifier( targets="Linear", scheme="W4A16", ignore=["lm_head"], dampening_frac=0.01, ) model = SparseAutoModelForCausalLM.from_pretrained( model_id, device_map="auto", trust_remote_code=True, ) oneshot( model=model, dataset=ds, recipe=recipe, max_seq_length=max_seq_len, num_calibration_samples=num_samples, ) model.save_pretrained("Meta-Llama-3.1-405B-Instruct-quantized.w4a16") ``` ## 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-3.1-405B-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 #### Open LLM Leaderboard evaluation scores
Benchmark | Meta-Llama-3.1-405B-Instruct | Meta-Llama-3.1-405B-Instruct-quantized.w4a16 (this model) | Recovery |
Arena Hard | 67.4 (67.3 / 67.5) | 66.5 (66.5 / 66.4) | 98.7% |
OpenLLM v1 | |||
MMLU (5-shot) | 87.4 | 87.2 | 99.8% |
ARC Challenge (0-shot) | 95.0 | 95.3 | 100.4% |
GSM-8K (CoT, 8-shot, strict-match) | 96.4 | 96.3 | 99.8% |
Hellaswag (10-shot) | 88.3 | 88.3 | 99.9% |
Winogrande (5-shot) | 87.2 | 87.4 | 100.2% |
TruthfulQA (0-shot) | 64.6 | 65.3 | 101.0% |
Average | 86.8 | 86.8 | 100.0% |
OpenLLM v2 | |||
MMLU-Pro (5-shot) | 59.7 | 59.4 | 99.3% |
IFEval (0-shot) | 87.7 | 88.0 | 100.4% |
BBH (3-shot) | 67.0 | 67.5 | 100.7% |
Math-|v|-5 (4-shot) | 39.0 | 37.6 | 96.5% |
GPQA (0-shot) | 19.5 | 17.5 | 89.8% |
MuSR (0-shot) | 19.5 | 19.4 | 99.5% |
Average | 48.7 | 48.2 | 99.0% |
Coding | |||
HumanEval pass@1 | 86.8 | 85.1 | 98.0% |
HumanEval+ pass@1 | 80.1 | 78.9 | 98.5% |