--- tags: - fp8 - 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 --- # Meta-Llama-3.1-405B-FP8 ## Model Overview - **Model Architecture:** Meta-Llama-3.1 - **Input:** Text - **Output:** Text - **Model Optimizations:** - **Weight quantization:** FP8 - **Activation quantization:** FP8 - **Intended Use Cases:** Intended for commercial and research use in multiple languages. Similarly to [Meta-Llama-3.1-8B](https://huggingface.co./meta-llama/Meta-Llama-3.1-8B), this model serves as a base version. - **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:** 8/6/2024 - **Version:** 1.0 - **License(s):** [llama3.1](https://huggingface.co./meta-llama/Meta-Llama-3.1-8B/blob/main/LICENSE) - **Model Developers:** Neural Magic Quantized version of [Meta-Llama-3.1-405B](https://huggingface.co./meta-llama/Meta-Llama-3.1-405B). It achieves an average score of 82.00 on the [OpenLLM](https://huggingface.co./spaces/open-llm-leaderboard/open_llm_leaderboard) benchmark (version 1), recovering 98.7% of dense performance. ### Model Optimizations This model was obtained by quantizing the weights and activations of [Meta-Llama-3.1-405B](https://huggingface.co./meta-llama/Meta-Llama-3.1-405B) to FP8 data type, ready for inference with vLLM built from source. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%. Only the weights and activations of the linear operators within transformers blocks are quantized. Symmetric per-tensor quantization is applied, in which a single linear scaling maps the FP8 representations of the quantized weights and activations. [LLM Compressor](https://github.com/vllm-project/llm-compressor) is used for quantization with 512 sequences of UltraChat. ## Creation This model was created by applying [LLM Compressor with calibration samples from UltraChat](https://github.com/vllm-project/llm-compressor/blob/sa/big_model_support/examples/big_model_offloading/big_model_w8a8_calibrate.py), as presented in the code snipet below. ```python import torch from datasets import load_dataset from transformers import AutoTokenizer from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot from llmcompressor.transformers.compression.helpers import ( calculate_offload_device_map, custom_offload_device_map, ) recipe = """ quant_stage: quant_modifiers: QuantizationModifier: ignore: ["lm_head"] config_groups: group_0: weights: num_bits: 8 type: float strategy: tensor dynamic: false symmetric: true input_activations: num_bits: 8 type: float strategy: tensor dynamic: false symmetric: true targets: ["Linear"] """ model_stub = "meta-llama/Meta-Llama-3.1-405B" model_name = model_stub.split("/")[-1] device_map = calculate_offload_device_map( model_stub, reserve_for_hessians=False, num_gpus=8, torch_dtype=torch.float16 ) model = SparseAutoModelForCausalLM.from_pretrained( model_stub, torch_dtype=torch.float16, device_map=device_map ) tokenizer = AutoTokenizer.from_pretrained(model_stub) output_dir = f"./{model_name}-FP8" DATASET_ID = "HuggingFaceH4/ultrachat_200k" DATASET_SPLIT = "train_sft" NUM_CALIBRATION_SAMPLES = 512 MAX_SEQUENCE_LENGTH = 4096 ds = load_dataset(DATASET_ID, split=DATASET_SPLIT) ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES)) def preprocess(example): return { "text": tokenizer.apply_chat_template( example["messages"], tokenize=False, ) } ds = ds.map(preprocess) def tokenize(sample): return tokenizer( sample["text"], padding=False, max_length=MAX_SEQUENCE_LENGTH, truncation=True, add_special_tokens=False, ) ds = ds.map(tokenize, remove_columns=ds.column_names) oneshot( model=model, output_dir=output_dir, dataset=ds, recipe=recipe, max_seq_length=MAX_SEQUENCE_LENGTH, num_calibration_samples=NUM_CALIBRATION_SAMPLES, save_compressed=True, ) ``` ## 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](https://github.com/neuralmagic/lm-evaluation-harness/tree/llama_3.1_instruct) (branch llama_3.1_instruct) and the [vLLM](https://docs.vllm.ai/en/stable/) engine. This version of the lm-evaluation-harness includes versions of ARC-Challenge that matches the prompting style of [Meta-Llama-3.1-evals](https://huggingface.co./datasets/meta-llama/Meta-Llama-3.1-8B-evals). An asterisk indicates that some evaluations are still being collected. ### Accuracy #### Open LLM Leaderboard evaluation scores
Benchmark | Meta-Llama-3.1-405B | Meta-Llama-3.1-405B-FP8(this model) | Recovery |
MMLU (5-shot) | * | 84.72 | * |
ARC Challenge (0-shot) | 95.99 | 95.82 | 99.82% |
GSM-8K (5-shot, strict-match) | 88.10 | 87.94 | 99.82% |
Hellaswag (10-shot) | 90.02 | 89.14 | 99.02% |
Winogrande (5-shot) | 87.61 | 86.42 | 98.64% |
TruthfulQA (0-shot) | 49.83 | 47.93 | 96.19% |
Average | * | 82.00 | 98.70% |