--- tags: - fp8 - vllm license: llama3.1 license_link: https://huggingface.co/meta-llama/Meta-Llama-3.1-8B/blob/main/LICENSE language: - en --- # Meta-Llama-3.1-405B-Instruct-FP8-dynamic ## 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 English. Similarly to [Meta-Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-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:** 7/24/2024 - **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-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-405B-Instruct). It achieves an average recovery of 100.1% on the [OpenLLM](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) benchmark (version 1), compared to the unquantized model. ### Model Optimizations This model was obtained by quantizing the weights and activations of [Meta-Llama-3.1-405B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-405B-Instruct) 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%. In particular, this model can now be loaded and evaluated with a single node of 8xH100 GPUs, as opposed to multiple nodes. Only the weights and activations 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 FP8 representations of the quantized weights and activations. Activations are also quantized on a per-token dynamic basis. [LLM Compressor](https://github.com/vllm-project/llm-compressor) is used for quantization with 512 sequences of UltraChat. ## Deployment ### Use with vLLM 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-FP8-dynamic" number_gpus = 8 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=4096) 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 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 transformers import AutoTokenizer from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot from llmcompressor.transformers.compression.helpers import ( # noqa 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: channel dynamic: false symmetric: true input_activations: num_bits: 8 type: float strategy: token dynamic: true symmetric: true targets: ["Linear"] """ model_stub = "meta-llama/Meta-Llama-3.1-405B-Instruct" 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 ) output_dir = f"./{model_name}-FP8-dynamic" oneshot( model=model, recipe=recipe, output_dir=output_dir, save_compressed=True, tokenizer=AutoTokenizer.from_pretrained(model_stub), ) ``` ## Evaluation The model was evaluated on the [OpenLLM](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) leaderboard tasks (version 1) with the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) and the [vLLM](https://docs.vllm.ai/en/stable/) engine, using the following command: ``` lm_eval \ --model vllm \ --model_args pretrained="neuralmagic/Meta-Llama-3.1-405B-Instruct-FP8-dynamic",dtype=auto,tensor_parallel_size=8,gpu_memory_utilization=0.755,add_bos_token=True,max_model_len=4096 \ --tasks openllm \ --batch_size auto ``` Certain benchmarks for the full precision model are still being acquired. Average recovery is calculated only with metrics that both models have been evaluated on. ### Accuracy #### Open LLM Leaderboard evaluation scores
Benchmark Meta-Llama-3.1-405B-Instruct Meta-Llama-3.1-405B-Instruct-FP8-dynamic(this model) Recovery
MMLU (5-shot) * 86.17 *
ARC Challenge (25-shot) * * *
GSM-8K (5-shot, strict-match) 95.07 95.00 99.93%
Hellaswag (10-shot) * 88.34 *
Winogrande (5-shot) 87.21 87.45 100.2%
TruthfulQA (0-shot) 64.64 64.71 100.1%
Average * * 100.1%