Meta-Llama-3-70B-Instruct-FP8
Model Overview
- Model Architecture: Meta-Llama-3
- 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-3-70B-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). Use in languages other than English.
- Release Date: 6/8/2024
- Version: 1.0
- License(s): Llama3
- Model Developers: Neural Magic
Quantized version of Meta-Llama-3-70B-Instruct. It achieves an average score of 79.16 on the OpenLLM benchmark (version 1), whereas the unquantized model achieves 79.51.
Model Optimizations
This model was obtained by quantizing the weights and activations of Meta-Llama-3-70B-Instruct to FP8 data type, ready for inference with vLLM >= 0.5.0. 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. AutoFP8 is used for quantization with 512 sequences of UltraChat.
Deployment
Use with vLLM
This model can be deployed efficiently using the vLLM backend, as shown in the example below.
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
model_id = "neuralmagic/Meta-Llama-3-70B-Instruct-FP8"
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, tokenize=False)
llm = LLM(model=model_id)
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 for more details.
Creation
This model was created by applying AutoFP8 with calibration samples from ultrachat, as presented in the code snipet below. Although AutoFP8 was used for this particular model, Neural Magic is transitioning to using llm-compressor which supports several quantization schemes and models not supported by AutoFP8.
from datasets import load_dataset
from transformers import AutoTokenizer
from auto_fp8 import AutoFP8ForCausalLM, BaseQuantizeConfig
pretrained_model_dir = "meta-llama/Meta-Llama-3-70B-Instruct"
quantized_model_dir = "Meta-Llama-3-70B-Instruct-FP8"
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True, model_max_length=4096)
tokenizer.pad_token = tokenizer.eos_token
ds = load_dataset("mgoin/ultrachat_2k", split="train_sft").select(range(512))
examples = [tokenizer.apply_chat_template(batch["messages"], tokenize=False) for batch in ds]
examples = tokenizer(examples, padding=True, truncation=True, return_tensors="pt").to("cuda")
quantize_config = BaseQuantizeConfig(
quant_method="fp8",
activation_scheme="dynamic", # or "static"
ignore_patterns=["re:.*lm_head"],
)
model = AutoFP8ForCausalLM.from_pretrained(
pretrained_model_dir, quantize_config=quantize_config
)
model.quantize(examples)
model.save_quantized(quantized_model_dir)
Evaluation
The model was evaluated on the OpenLLM leaderboard tasks (version 1) with the lm-evaluation-harness (commit 383bbd54bc621086e05aa1b030d8d4d5635b25e6) and the vLLM engine, using the following command:
lm_eval \
--model vllm \
--model_args pretrained="neuralmagic/Meta-Llama-3-70B-Instruct-FP8",dtype=auto,gpu_memory_utilization=0.4,add_bos_token=True,max_model_len=4096 \
--tasks openllm \
--batch_size auto
Accuracy
Open LLM Leaderboard evaluation scores
Benchmark | Meta-Llama-3-70B-Instruct | Meta-Llama-3-70B-Instruct-FP8(this model) | Recovery |
MMLU (5-shot) | 80.18 | 80.06 | 99.85% |
ARC Challenge (25-shot) | 72.69 | 72.61 | 99.88% |
GSM-8K (5-shot, strict-match) | 92.49 | 91.12 | 98.51% |
Hellaswag (10-shot) | 85.50 | 85.41 | 99.89% |
Winogrande (5-shot) | 83.34 | 83.03 | 99.62% |
TruthfulQA (0-shot) | 62.90 | 62.73 | 99.72% |
Average | 79.51 | 79.16 | 99.55% |
- Downloads last month
- 106,834