--- 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-3.1-405B-Instruct --- # Meta-Llama-3.1-405B-Instruct-quantized.w8a8 ## Model Overview - **Model Architecture:** Meta-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 [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). - **Release Date:** 8/19/2024 - **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-FP8-dynamic achieves 95.8% recovery for the Arena-Hard evaluation, 99.3% for OpenLLM v1 (using Meta's prompting when available), 98.4% for OpenLLM v2, 100.1% for HumanEval pass@1, and 100.4% 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 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. Linear scaling factors are computed via by minimizing the mean squarred error (MSE). 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 1% damping factor and 512 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-3.1-405B-Instruct-quantized.w8a8" number_gpus = 8 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 (using 8 A100 80GB GPUs). ```python from transformers import AutoTokenizer from datasets import load_dataset from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot from llmcompressor.modifiers.quantization import GPTQModifier from llmcompressor.transformers.compression.helpers import custom_offload_device_map model_id = "meta-llama/Meta-Llama-3.1-405B-Instruct" num_samples = 512 max_seq_len = 4096 num_gpus = 8 max_memory_per_gpu = "20GB" 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 = GPTQModifier( sequential=True, targets="Linear", scheme="W8A8", ignore=["lm_head"], dampening_frac=0.01, observer="mse" ) device_map = custom_offload_device_map( model_id, max_memory_per_gpu=max_memory_per_gpu, num_gpus=num_gpus, torch_dtype="auto", ) 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("Meta-Llama-3.1-405B-Instruct-quantized.w8a8") ``` ## 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
Benchmark | Meta-Llama-3.1-405B-Instruct | Meta-Llama-3.1-405B-Instruct-quantized.w8a8 (this model) | Recovery |
Arena Hard | 67.4 (67.3 / 67.5) | 64.6 (64.3 / 64.8) | 95.8% |
OpenLLM v1 | |||
MMLU (5-shot) | 87.4 | 87.1 | 99.6% |
ARC Challenge (0-shot) | 95.0 | 94.4 | 99.4% |
GSM-8K (CoT, 8-shot, strict-match) | 96.4 | 95.5 | 99.0% |
Hellaswag (10-shot) | 88.3 | 88.2 | 99.8% |
Winogrande (5-shot) | 87.2 | 86.1 | 98.7% |
TruthfulQA (0-shot) | 64.6 | 64.4 | 99.6% |
Average | 86.8 | 86.2 | 99.3% |
OpenLLM v2 | |||
MMLU-Pro (5-shot) | 59.7 | 58.4 | 97.8% |
IFEval (0-shot) | 87.7 | 87.0 | 99.2% |
BBH (3-shot) | 67.0 | 66.7 | 99.6% |
Math-lvl-5 (4-shot) | 39.0 | 35.8 | 91.9% |
GPQA (0-shot) | 19.5 | 20.4 | 104.5% |
MuSR (0-shot) | 19.5 | 19.2 | 98.8% |
Average | 48.7 | 47.9 | 98.4% |
Coding | |||
HumanEval pass@1 | 86.8 | 86.9 | 100.1% |
HumanEval+ pass@1 | 80.1 | 80.4 | 100.4% |