Mixtral-8x7B-Instruct-v0.1-fp8

<!-- header start --> <p align="center"> <img src="https://i.imgur.com/mNM6Cai.png" width="100%" alt="Friendli Logo"> </p> <!-- header end -->

Mixtral-8x7B-Instruct-v0.1 - FP8

• Model creator: Mistral AI
• Original model: Mixtral-8x7B-Instruct-v0.1

Description

This repo contains the Mixtral-8x7B-Instruct-v0.1 model quantized to FP8 by FriendliAI, significantly enhancing its inference efficiency while maintaining high accuracy. Note that FP8 is only supported by NVIDIA Ada, Hopper, and Blackwell GPU architectures. Check out FriendliAI documentation for more details.

Compatibility

This model is compatible with Friendli Container.

Prerequisites

• Before you begin, make sure you have signed up for Friendli Suite. You can use Friendli Containers free of charge for four weeks.
• Prepare a Personal Access Token following this guide.
• Prepare a Friendli Container Secret following this guide.

Preparing Personal Access Token

PAT (Personal Access Token) is the user credential for for logging into our container registry.

1. Sign in Friendli Suite.
2. Go to User Settings > Tokens and click 'Create new token'.
3. Save your created token value.

Pulling Friendli Container Image

1. Log in to the Docker client using the personal access token created as outlined in this guide.

export FRIENDLI_PAT="YOUR PAT"
docker login registry.friendli.ai -u $YOUR_EMAIL -p $FRIENDLI_PAT

1. Pull image

docker pull registry.friendli.ai/trial

Running Friendli Container

Once you've prepared the image of Friendli Container, you can launch it to create a serving endpoint.

docker run \
  --gpus '"device=0"' \
  -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -e FRIENDLI_CONTAINER_SECRET="YOUR CONTAINER SECRET" \
  registry.friendli.ai/trial \
    --web-server-port 8000 \
    --hf-model-name FriendliAI/Mixtral-8x7B-Instruct-v0.1-fp8

Optimizing Inference Performance with Policy Search

To serve MoE models efficiently, it is required to run a policy search to explore the optimal execution policy:

export POLICY_DIR=$PWD/policy

mkdir -p $POLICY_DIR

docker run \
  --gpus '"device=0"' \
  -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -v $POLICY_DIR:/policy \
  -e FRIENDLI_CONTAINER_SECRET="YOUR CONTAINER SECRET" \
  registry.friendli.ai/trial \
    --web-server-port 8000 \
    --hf-model-name FriendliAI/Mixtral-8x7B-Instruct-v0.1-fp8 \
    --algo-policy-dir /policy \
    --search-policy true

When the optimal policy is successfully searched, the policy is compiled into a policy file and saved at $POLICY_DIR. Now you can create an inference endpoint with this optimal policy as follows:

docker run \
  --gpus '"device=0"' \
  -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -v $POLICY_DIR:/policy \
  -e FRIENDLI_CONTAINER_SECRET="YOUR CONTAINER SECRET" \
  registry.friendli.ai/trial \
    --web-server-port 8000 \
    --hf-model-name FriendliAI/Mixtral-8x7B-Instruct-v0.1-fp8 \
    --algo-policy-dir /policy

---

Original model card: MistralAI's Mixtr-8x7B Instruct v0.1

Mixtral-8x7B

The Mixtral-8x7B Large Language Model (LLM) is a pretrained generative Sparse Mixture of Experts. The Mixtral-8x7B outperforms Llama 2 70B on most benchmarks we tested.

For full details of this model please read our release blog post.

Warning

This repo contains weights that are compatible with vLLM serving of the model as well as Hugging Face transformers library. It is based on the original Mixtral torrent release, but the file format and parameter names are different. Please note that model cannot (yet) be instantiated with HF.

Instruction format

This format must be strictly respected, otherwise the model will generate sub-optimal outputs.

The template used to build a prompt for the Instruct model is defined as follows:

<s> [INST] Instruction [/INST] Model answer</s> [INST] Follow-up instruction [/INST]

Note that <s> and </s> are special tokens for beginning of string (BOS) and end of string (EOS) while [INST] and [/INST] are regular strings.

As reference, here is the pseudo-code used to tokenize instructions during fine-tuning:

def tokenize(text):
    return tok.encode(text, add_special_tokens=False)

[BOS_ID] + 
tokenize("[INST]") + tokenize(USER_MESSAGE_1) + tokenize("[/INST]") +
tokenize(BOT_MESSAGE_1) + [EOS_ID] +
…
tokenize("[INST]") + tokenize(USER_MESSAGE_N) + tokenize("[/INST]") +
tokenize(BOT_MESSAGE_N) + [EOS_ID]

In the pseudo-code above, note that the tokenize method should not add a BOS or EOS token automatically, but should add a prefix space.

In the Transformers library, one can use chat templates which make sure the right format is applied.

Run the model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mistralai/Mixtral-8x7B-Instruct-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "user", "content": "What is your favourite condiment?"},
    {"role": "assistant", "content": "Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!"},
    {"role": "user", "content": "Do you have mayonnaise recipes?"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")

outputs = model.generate(inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

By default, transformers will load the model in full precision. Therefore you might be interested to further reduce down the memory requirements to run the model through the optimizations we offer in HF ecosystem:

In half-precision

Note float16 precision only works on GPU devices

<details> <summary> Click to expand </summary>

+ import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mistralai/Mixtral-8x7B-Instruct-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)

+ model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

messages = [
    {"role": "user", "content": "What is your favourite condiment?"},
    {"role": "assistant", "content": "Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!"},
    {"role": "user", "content": "Do you have mayonnaise recipes?"}
]

input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")

outputs = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

</details>

Lower precision using (8-bit & 4-bit) using bitsandbytes

<details> <summary> Click to expand </summary>

+ import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mistralai/Mixtral-8x7B-Instruct-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)

+ model = AutoModelForCausalLM.from_pretrained(model_id, load_in_4bit=True, device_map="auto")

text = "Hello my name is"
messages = [
    {"role": "user", "content": "What is your favourite condiment?"},
    {"role": "assistant", "content": "Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!"},
    {"role": "user", "content": "Do you have mayonnaise recipes?"}
]

input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")

outputs = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

</details>

Load the model with Flash Attention 2

<details> <summary> Click to expand </summary>

+ import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mistralai/Mixtral-8x7B-Instruct-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)

+ model = AutoModelForCausalLM.from_pretrained(model_id, use_flash_attention_2=True, device_map="auto")

messages = [
    {"role": "user", "content": "What is your favourite condiment?"},
    {"role": "assistant", "content": "Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!"},
    {"role": "user", "content": "Do you have mayonnaise recipes?"}
]

input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")

outputs = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

</details>

Limitations

The Mixtral-8x7B Instruct model is a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance. It does not have any moderation mechanisms. We're looking forward to engaging with the community on ways to make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs.

The Mistral AI Team

Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Louis Ternon, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.