English | 中文
? Hugging Face • ? ModelScope • ? wisemodel
? Ask questions or discuss ideas on GitHub
Join us on ? Discord or WeChat
Check out Yi Tech Report
Grow at Yi Learning Hub
? Learn at Yi Tech Blog
? The Yi series models are the next generation of open-source large language models trained from scratch by 01.AI.
? Targeted as a bilingual language model and trained on 3T multilingual corpus, the Yi series models become one of the strongest LLM worldwide, showing promise in language understanding, commonsense reasoning, reading comprehension, and more. For example,
Yi-34B-Chat model landed in second place (following GPT-4 Turbo), outperforming other LLMs (such as GPT-4, Mixtral, Claude) on the AlpacaEval Leaderboard (based on data available up to January 2024).
Yi-34B model ranked first among all existing open-source models (such as Falcon-180B, Llama-70B, Claude) in both English and Chinese on various benchmarks, including Hugging Face Open LLM Leaderboard (pre-trained) and C-Eval (based on data available up to November 2023).
(Credits to Llama) Thanks to the Transformer and Llama open-source communities, as they reduce the efforts required to build from scratch and enable the utilization of the same tools within the AI ecosystem.
TL;DR
The Yi series models adopt the same model architecture as Llama but are NOT derivatives of Llama.
Both Yi and Llama are based on the Transformer structure, which has been the standard architecture for large language models since 2018.
Grounded in the Transformer architecture, Llama has become a new cornerstone for the majority of state-of-the-art open-source models due to its excellent stability, reliable convergence, and robust compatibility. This positions Llama as the recognized foundational framework for models including Yi.
Thanks to the Transformer and Llama architectures, other models can leverage their power, reducing the effort required to build from scratch and enabling the utilization of the same tools within their ecosystems.
However, the Yi series models are NOT derivatives of Llama, as they do not use Llama's weights.
As Llama's structure is employed by the majority of open-source models, the key factors of determining model performance are training datasets, training pipelines, and training infrastructure.
Developing in a unique and proprietary way, Yi has independently created its own high-quality training datasets, efficient training pipelines, and robust training infrastructure entirely from the ground up. This effort has led to excellent performance with Yi series models ranking just behind GPT4 and surpassing Llama on the Alpaca Leaderboard in Dec 2023.
[ Back to top ⬆️ ]
Yi-9B-200K is open-sourced and available to the public.Yi-9B is open-sourced and available to the public.Yi-9B stands out as the top performer among a range of similar-sized open-source models (including Mistral-7B, SOLAR-10.7B, Gemma-7B, DeepSeek-Coder-7B-Base-v1.5 and more), particularly excelling in code, math, common-sense reasoning, and reading comprehension.
Yi-VL-34B and Yi-VL-6B, are open-sourced and available to the public.Yi-VL-34B has ranked first among all existing open-source models in the latest benchmarks, including MMMU and CMMMU (based on data available up to January 2024).
Yi-34B-ChatYi-34B-Chat-4bitsYi-34B-Chat-8bitsYi-6B-ChatYi-6B-Chat-4bitsYi-6B-Chat-8bitsYou can try some of them interactively at:
Yi-6B-200K and Yi-34B-200K, are open-sourced and available to the public.Yi-6B and Yi-34B, are open-sourced and available to the public.[ Back to top ⬆️ ]
Yi models come in multiple sizes and cater to different use cases. You can also fine-tune Yi models to meet your specific requirements.
If you want to deploy Yi models, make sure you meet the software and hardware requirements.
| Model | Download |
|---|---|
| Yi-34B-Chat | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-34B-Chat-4bits | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-34B-Chat-8bits | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-6B-Chat | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-6B-Chat-4bits | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-6B-Chat-8bits | • ? Hugging Face • ? ModelScope • ? wisemodel |
- 4-bit series models are quantized by AWQ.
- 8-bit series models are quantized by GPTQ
- All quantized models have a low barrier to use since they can be deployed on consumer-grade GPUs (e.g., 3090, 4090).
| Model | Download |
|---|---|
| Yi-34B | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-34B-200K | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-9B | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-9B-200K | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-6B | • ? Hugging Face • ? ModelScope • ? wisemodel |
| Yi-6B-200K | • ? Hugging Face • ? ModelScope • ? wisemodel |
- 200k is roughly equivalent to 400,000 Chinese characters.
- If you want to use the previous version of the Yi-34B-200K (released on Nov 5, 2023), run git checkout 069cd341d60f4ce4b07ec394e82b79e94f656cf to download the weight.
| Model | Intro | Default context window | Pretrained tokens | Training Data Date |
|---|---|---|---|---|
| 6B series models | They are suitable for personal and academic use. | 4K | 3T | Up to June 2023 |
| 9B series models | It is the best at coding and math in the Yi series models. | Yi-9B is continuously trained based on Yi-6B, using 0.8T tokens. | ||
| 34B series models | They are suitable for personal, academic, and commercial (particularly for small and medium-sized enterprises) purposes. It's a cost-effective solution that's affordable and equipped with emergent ability. | 3T |
For chat models
However, this higher diversity might amplify certain existing issues, including:
[ Back to top ⬆️ ]
Tip: If you want to get started with the Yi model and explore different methods for inference, check out the Yi Cookbook.
Select one of the following paths to begin your journey with Yi!
If you prefer to deploy Yi models locally,
?♀️ and you have sufficient resources (for example, NVIDIA A800 80GB), you can choose one of the following methods:
?♀️ and you have limited resources (for example, a MacBook Pro), you can use llama.cpp.
If you prefer not to deploy Yi models locally, you can explore Yi's capabilities using any of the following options.
If you want to explore more features of Yi, you can adopt one of these methods:
Yi APIs (Yi official)
Yi APIs (Replicate)
If you want to chat with Yi with more customizable options (e.g., system prompt, temperature, repetition penalty, etc.), you can try one of the following options:
Yi-34B-Chat-Playground (Yi official)
Yi-34B-Chat-Playground (Replicate)
If you want to chat with Yi, you can use one of these online services, which offer a similar user experience:
Yi-34B-Chat (Yi official on Hugging Face)
Yi-34B-Chat (Yi official beta)
[ Back to top ⬆️ ]
This tutorial guides you through every step of running Yi-34B-Chat locally on an A800 (80G) and then performing inference.
Make sure Python 3.10 or a later version is installed.
If you want to run other Yi models, see software and hardware requirements.
To set up the environment and install the required packages, execute the following command.
git clone https://github.com/01-ai/Yi.git
cd yi
pip install -r requirements.txtYou can download the weights and tokenizer of Yi models from the following sources:
You can perform inference with Yi chat or base models as below.
Create a file named quick_start.py and copy the following content to it.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_path = '<your-model-path>'
tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
# Since transformers 4.35.0, the GPT-Q/AWQ model can be loaded using AutoModelForCausalLM.
model = AutoModelForCausalLM.from_pretrained(
model_path,
device_map="auto",
torch_dtype='auto'
).eval()
# Prompt content: "hi"
messages = [
{"role": "user", "content": "hi"}
]
input_ids = tokenizer.apply_chat_template(conversation=messages, tokenize=True, add_generation_prompt=True, return_tensors='pt')
output_ids = model.generate(input_ids.to('cuda'))
response = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True)
# Model response: "Hello! How can I assist you today?"
print(response)Run quick_start.py.
python quick_start.pyThen you can see an output similar to the one below. ?
Hello! How can I assist you today?Yi-34B
The steps are similar to pip - Perform inference with Yi chat model.
You can use the existing file text_generation.py.
python demo/text_generation.py --model <your-model-path>Then you can see an output similar to the one below. ?
Prompt: Let me tell you an interesting story about cat Tom and mouse Jerry,
Generation: Let me tell you an interesting story about cat Tom and mouse Jerry, which happened in my childhood. My father had a big house with two cats living inside it to kill mice. One day when I was playing at home alone, I found one of the tomcats lying on his back near our kitchen door, looking very much like he wanted something from us but couldn’t get up because there were too many people around him! He kept trying for several minutes before finally giving up...
Yi-9B
Input
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_DIR = "01-ai/Yi-9B"
model = AutoModelForCausalLM.from_pretrained(MODEL_DIR, torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR, use_fast=False)
input_text = "# write the quick sort algorithm"
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_length=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))Output
# write the quick sort algorithm
def quick_sort(arr):
if len(arr) <= 1:
return arr
pivot = arr[len(arr) // 2]
left = [x for x in arr if x < pivot]
middle = [x for x in arr if x == pivot]
right = [x for x in arr if x > pivot]
return quick_sort(left) + middle + quick_sort(right)
# test the quick sort algorithm
print(quick_sort([3, 6, 8, 10, 1, 2, 1]))[ Back to top ⬆️ ]
Make sure you've installed Docker and nvidia-container-toolkit.
docker run -it --gpus all
-v <your-model-path>: /models
ghcr.io/01-ai/yi:latest
Alternatively, you can pull the Yi Docker image from registry.lingyiwanwu.com/ci/01-ai/yi:latest.
You can perform inference with Yi chat or base models as below.
The steps are similar to pip - Perform inference with Yi chat model.
Note that the only difference is to set model_path = '<your-model-mount-path>' instead of model_path = '<your-model-path>'.
The steps are similar to pip - Perform inference with Yi base model.
Note that the only difference is to set --model <your-model-mount-path>' instead of model <your-model-path>.
conda-lock to generate fully reproducible lock files for conda environments. micromamba for installing these dependencies.
Install micromamba by following the instructions available here.
Execute micromamba install -y -n yi -f conda-lock.yml to create a conda environment named yi and install the necessary dependencies.
The following tutorial will guide you through every step of running a quantized model (Yi-chat-6B-2bits) locally and then performing inference.
This tutorial assumes you use a MacBook Pro with 16GB of memory and an Apple M2 Pro chip.
Make sure git-lfs is installed on your machine.
llama.cpp
To clone the llama.cpp repository, run the following command.
git clone [email protected]:ggerganov/llama.cpp.git2.1 To clone XeIaso/yi-chat-6B-GGUF with just pointers, run the following command.
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/XeIaso/yi-chat-6B-GGUF2.2 To download a quantized Yi model (yi-chat-6b.Q2_K.gguf), run the following command.
git-lfs pull --include yi-chat-6b.Q2_K.ggufTo perform inference with the Yi model, you can use one of the following methods.
Method 1: Perform inference in terminal
Method 2: Perform inference in web
To compile llama.cpp using 4 threads and then conduct inference, navigate to the llama.cpp directory, and run the following command.
Tips
Replace
/Users/yu/yi-chat-6B-GGUF/yi-chat-6b.Q2_K.ggufwith the actual path of your model.By default, the model operates in completion mode.
For additional output customization options (for example, system prompt, temperature, repetition penalty, etc.), run
./main -hto check detailed descriptions and usage.
make -j4 && ./main -m /Users/yu/yi-chat-6B-GGUF/yi-chat-6b.Q2_K.gguf -p "How do you feed your pet fox? Please answer this question in 6 simple steps:nStep 1:" -n 384 -e
...
How do you feed your pet fox? Please answer this question in 6 simple steps:
Step 1: Select the appropriate food for your pet fox. You should choose high-quality, balanced prey items that are suitable for their unique dietary needs. These could include live or frozen mice, rats, pigeons, or other small mammals, as well as fresh fruits and vegetables.
Step 2: Feed your pet fox once or twice a day, depending on the species and its individual preferences. Always ensure that they have access to fresh water throughout the day.
Step 3: Provide an appropriate environment for your pet fox. Ensure it has a comfortable place to rest, plenty of space to move around, and opportunities to play and exercise.
Step 4: Socialize your pet with other animals if possible. Interactions with other creatures can help them develop social skills and prevent boredom or stress.
Step 5: Regularly check for signs of illness or discomfort in your fox. Be prepared to provide veterinary care as needed, especially for common issues such as parasites, dental health problems, or infections.
Step 6: Educate yourself about the needs of your pet fox and be aware of any potential risks or concerns that could affect their well-being. Regularly consult with a veterinarian to ensure you are providing the best care.
...
Now you have successfully asked a question to the Yi model and got an answer! ?
To initialize a lightweight and swift chatbot, run the following command.
cd llama.cpp
./server --ctx-size 2048 --host 0.0.0.0 --n-gpu-layers 64 --model /Users/yu/yi-chat-6B-GGUF/yi-chat-6b.Q2_K.ggufThen you can get an output like this:
...
llama_new_context_with_model: n_ctx = 2048
llama_new_context_with_model: freq_base = 5000000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M2 Pro
ggml_metal_init: picking default device: Apple M2 Pro
ggml_metal_init: ggml.metallib not found, loading from source
ggml_metal_init: GGML_METAL_PATH_RESOURCES = nil
ggml_metal_init: loading '/Users/yu/llama.cpp/ggml-metal.metal'
ggml_metal_init: GPU name: Apple M2 Pro
ggml_metal_init: GPU family: MTLGPUFamilyApple8 (1008)
ggml_metal_init: hasUnifiedMemory = true
ggml_metal_init: recommendedMaxWorkingSetSize = 11453.25 MB
ggml_metal_init: maxTransferRate = built-in GPU
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size = 128.00 MiB, ( 2629.44 / 10922.67)
llama_new_context_with_model: KV self size = 128.00 MiB, K (f16): 64.00 MiB, V (f16): 64.00 MiB
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size = 0.02 MiB, ( 2629.45 / 10922.67)
llama_build_graph: non-view tensors processed: 676/676
llama_new_context_with_model: compute buffer total size = 159.19 MiB
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size = 156.02 MiB, ( 2785.45 / 10922.67)
Available slots:
-> Slot 0 - max context: 2048
llama server listening at http://0.0.0.0:8080To access the chatbot interface, open your web browser and enter http://0.0.0.0:8080 into the address bar.
Enter a question, such as "How do you feed your pet fox? Please answer this question in 6 simple steps" into the prompt window, and you will receive a corresponding answer.
[ Back to top ⬆️ ]
You can build a web UI demo for Yi chat models (note that Yi base models are not supported in this senario).
Step 1: Prepare your environment.
Step 2: Download the Yi model.
Step 3. To start a web service locally, run the following command.
python demo/web_demo.py -c <your-model-path>You can access the web UI by entering the address provided in the console into your browser.
[ Back to top ⬆️ ]
bash finetune/scripts/run_sft_Yi_6b.shOnce finished, you can compare the finetuned model and the base model with the following command:
bash finetune/scripts/run_eval.shBy default, we use a small dataset from BAAI/COIG to finetune the base model.
You can also prepare your customized dataset in the following jsonl format:
{ "prompt": "Human: Who are you? Assistant:", "chosen": "I'm Yi." }And then mount them in the container to replace the default ones:
docker run -it
-v /path/to/save/finetuned/model/:/finetuned-model
-v /path/to/train.jsonl:/yi/finetune/data/train.json
-v /path/to/eval.jsonl:/yi/finetune/data/eval.json
ghcr.io/01-ai/yi:latest
bash finetune/scripts/run_sft_Yi_6b.shMake sure you have conda. If not, use
mkdir -p ~/miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
rm -rf ~/miniconda3/miniconda.sh
~/miniconda3/bin/conda init bash
source ~/.bashrcThen, create a conda env:
conda create -n dev_env python=3.10 -y
conda activate dev_env
pip install torch==2.0.1 deepspeed==0.10 tensorboard transformers datasets sentencepiece accelerate ray==2.7For the Yi-6B model, a node with 4 GPUs, each with GPU memory larger than 60GB, is recommended.
For the Yi-34B model, because the usage of the zero-offload technique consumes a lot of CPU memory, please be careful to limit the number of GPUs in the 34B finetune training. Please use CUDA_VISIBLE_DEVICES to limit the number of GPUs (as shown in scripts/run_sft_Yi_34b.sh).
A typical hardware setup for finetuning the 34B model is a node with 8 GPUs (limited to 4 in running by CUDA_VISIBLE_DEVICES=0,1,2,3), each with GPU memory larger than 80GB, and total CPU memory larger than 900GB.
Download a LLM-base model to MODEL_PATH (6B and 34B). A typical folder of models is like:
|-- $MODEL_PATH
| |-- config.json
| |-- pytorch_model-00001-of-00002.bin
| |-- pytorch_model-00002-of-00002.bin
| |-- pytorch_model.bin.index.json
| |-- tokenizer_config.json
| |-- tokenizer.model
| |-- ...Download a dataset from huggingface to local storage DATA_PATH, e.g. Dahoas/rm-static.
|-- $DATA_PATH
| |-- data
| | |-- train-00000-of-00001-2a1df75c6bce91ab.parquet
| | |-- test-00000-of-00001-8c7c51afc6d45980.parquet
| |-- dataset_infos.json
| |-- README.mdfinetune/yi_example_dataset has example datasets, which are modified from BAAI/COIG
|-- $DATA_PATH
|--data
|-- train.jsonl
|-- eval.jsonlcd into the scripts folder, copy and paste the script, and run. For example:
cd finetune/scripts
bash run_sft_Yi_6b.shFor the Yi-6B base model, setting training_debug_steps=20 and num_train_epochs=4 can output a chat model, which takes about 20 minutes.
For the Yi-34B base model, it takes a relatively long time for initialization. Please be patient.
cd finetune/scripts
bash run_eval.shThen you'll see the answer from both the base model and the finetuned model.
[ Back to top ⬆️ ]
python quantization/gptq/quant_autogptq.py
--model /base_model
--output_dir /quantized_model
--trust_remote_codeOnce finished, you can then evaluate the resulting model as follows:
python quantization/gptq/eval_quantized_model.py
--model /quantized_model
--trust_remote_codeGPT-Q is a PTQ (Post-Training Quantization) method. It saves memory and provides potential speedups while retaining the accuracy of the model.
Yi models can be GPT-Q quantized without a lot of efforts. We provide a step-by-step tutorial below.
To run GPT-Q, we will use AutoGPTQ and exllama. And the huggingface transformers has integrated optimum and auto-gptq to perform GPTQ quantization on language models.
The quant_autogptq.py script is provided for you to perform GPT-Q quantization:
python quant_autogptq.py --model /base_model
--output_dir /quantized_model --bits 4 --group_size 128 --trust_remote_codeYou can run a quantized model using the eval_quantized_model.py:
python eval_quantized_model.py --model /quantized_model --trust_remote_codepython quantization/awq/quant_autoawq.py
--model /base_model
--output_dir /quantized_model
--trust_remote_codeOnce finished, you can then evaluate the resulting model as follows:
python quantization/awq/eval_quantized_model.py
--model /quantized_model
--trust_remote_codeAWQ is a PTQ (Post-Training Quantization) method. It's an efficient and accurate low-bit weight quantization (INT3/4) for LLMs.
Yi models can be AWQ quantized without a lot of efforts. We provide a step-by-step tutorial below.
To run AWQ, we will use AutoAWQ.
The quant_autoawq.py script is provided for you to perform AWQ quantization:
python quant_autoawq.py --model /base_model
--output_dir /quantized_model --bits 4 --group_size 128 --trust_remote_codeYou can run a quantized model using the eval_quantized_model.py:
python eval_quantized_model.py --model /quantized_model --trust_remote_code[ Back to top ⬆️ ]
If you want to deploy Yi models, make sure you meet the software and hardware requirements.
Before using Yi quantized models, make sure you've installed the correct software listed below.
| Model | Software |
|---|---|
| Yi 4-bit quantized models | AWQ and CUDA |
| Yi 8-bit quantized models | GPTQ and CUDA |
Before deploying Yi in your environment, make sure your hardware meets the following requirements.
| Model | Minimum VRAM | Recommended GPU Example |
|---|---|---|
| Yi-6B-Chat | 15 GB | 1 x RTX 3090 (24 GB) 1 x RTX 4090 (24 GB) 1 x A10 (24 GB) 1 x A30 (24 GB) |
| Yi-6B-Chat-4bits | 4 GB | 1 x RTX 3060 (12 GB) 1 x RTX 4060 (8 GB) |
| Yi-6B-Chat-8bits | 8 GB | 1 x RTX 3070 (8 GB) 1 x RTX 4060 (8 GB) |
| Yi-34B-Chat | 72 GB | 4 x RTX 4090 (24 GB) 1 x A800 (80GB) |
| Yi-34B-Chat-4bits | 20 GB | 1 x RTX 3090 (24 GB) 1 x RTX 4090 (24 GB) 1 x A10 (24 GB) 1 x A30 (24 GB) 1 x A100 (40 GB) |
| Yi-34B-Chat-8bits | 38 GB | 2 x RTX 3090 (24 GB) 2 x RTX 4090 (24 GB) 1 x A800 (40 GB) |
Below are detailed minimum VRAM requirements under different batch use cases.
| Model | batch=1 | batch=4 | batch=16 | batch=32 |
|---|---|---|---|---|
| Yi-6B-Chat | 12 GB | 13 GB | 15 GB | 18 GB |
| Yi-6B-Chat-4bits | 4 GB | 5 GB | 7 GB | 10 GB |
| Yi-6B-Chat-8bits | 7 GB | 8 GB | 10 GB | 14 GB |
| Yi-34B-Chat | 65 GB | 68 GB | 76 GB | > 80 GB |
| Yi-34B-Chat-4bits | 19 GB | 20 GB | 30 GB | 40 GB |
| Yi-34B-Chat-8bits | 35 GB | 37 GB | 46 GB | 58 GB |
| Model | Minimum VRAM | Recommended GPU Example |
|---|---|---|
| Yi-6B | 15 GB | 1 x RTX 3090 (24 GB) 1 x RTX 4090 (24 GB) 1 x A10 (24 GB) 1 x A30 (24 GB) |
| Yi-6B-200K | 50 GB | 1 x A800 (80 GB) |
| Yi-9B | 20 GB | 1 x RTX 4090 (24 GB) |
| Yi-34B | 72 GB | 4 x RTX 4090 (24 GB) 1 x A800 (80 GB) |
| Yi-34B-200K | 200 GB | 4 x A800 (80 GB) |
[ Back to top ⬆️ ]
Yi-34Band Yi-34B-Chat comes down to the fine-tuning approach and outcomes.
Yi-34B could be your go-to.Yi-34B-Chat might be your best bet.
