jartine/Mixtral-8x22B-Instruct-v0.1-llamafile

Mixtral 8x22B Instruct v0.1 - llamafile

This repository contains executable weights (which we call llamafiles) that run on Linux, MacOS, Windows, FreeBSD, OpenBSD, and NetBSD for AMD64 and ARM64.

• Model creator: Mistral AI
• Original model: mistralai/Mixtral-8x22B-Instruct-v0.1

Quickstart

Assuming your system has at least 128GB of RAM, you can try running the following command which download, concatenate, and execute the model.

( curl -L https://huggingface.co/jartine/Mixtral-8x22B-Instruct-v0.1-llamafile/resolve/main/Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile.cat0
  curl -L https://huggingface.co/jartine/Mixtral-8x22B-Instruct-v0.1-llamafile/resolve/main/Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile.cat1
) > Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile
chmod +x Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile
./Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile --help   # view manual
./Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile          # launch web gui + oai api
./Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile -p ...   # cli interface (scriptable)

Alternatively, you may download an official llamafile executable from Mozilla Ocho on GitHub, in which case you can use the Mixtral llamafiles as a simple weights data file.

llamafile -m Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile ...

For further information, please see the llamafile README.

Having trouble? See the "Gotchas" section of the README.

Prompting

Prompt template:

[INST] {{prompt}} [/INST]

Command template:

./Mixtral-8x22B-Instruct-v0.1.Q4_0.llamafile -p "[INST]{{prompt}}[/INST]"

About llamafile

llamafile is a new format introduced by Mozilla Ocho on Nov 20th 2023. It uses Cosmopolitan Libc to turn LLM weights into runnable llama.cpp binaries that run on the stock installs of six OSes for both ARM64 and AMD64.

In addition to being executables, llamafiles are also zip archives. Each llamafile contains a GGUF file, which you can extract using the unzip command. If you want to change or add files to your llamafiles, then the zipalign command (distributed on the llamafile github) should be used instead of the traditional zip command.

About Upload Limits

Files which exceed the Hugging Face 50GB upload limit have a .cat𝑋 extension. You need to use the cat command locally to turn them back into a single file, using the same order.

About Quantization Formats (General Advice)

Your choice of quantization format depends on three things:

1. Will it fit in RAM or VRAM?
2. Is your use case reading (e.g. summarization) or writing (e.g. chatbot)?
3. llamafiles bigger than 4.30 GB are hard to run on Windows (see gotchas)

Good quants for writing (prediction speed) are Q5_K_M, and Q4_0. Text generation is bounded by memory speed, so smaller quants help, but they cause the LLM to hallucinate more. However that doesn't mean they can't think correctly. A highly degraded quant like Q2_K may not make a great encyclopedia, but it's still capable of logical reasoning and the emergent capabilities LLMs exhibit.

Good quants for reading (evaluation speed) are BF16, F16, Q8_0, and Q4_0 (ordered from fastest to slowest). Prompt evaluation is bounded by flop count, which means perf can be improved through software engineering alone, e.g. BLAS algorithms, in which case quantization starts hurting more than it helps, since it competes for CPU resources and makes it harder for the compiler to parallelize instructions. You want to ideally use the simplest smallest floating point format that's natively implemented by your hardware. In most cases, that's BF16 or FP16. However, llamafile is able to still offer respectable tinyBLAS speedups for llama.cpp's simplest quants: Q8_0 and Q4_0.

Hardware Choices (Mixtral 8x22B Specific)

This model is very large. Even at Q2 quantization, it's still well-over twice as large the highest tier NVIDIA gaming GPUs. llamafile supports splitting models over multiple GPUs (for NVIDIA only currently) if you have such a system. The easiest way to have one, if you don't, is to pay a few bucks an hour to rent a 4x RTX 4090 rig off vast.ai.

Mac Studio is a good option for running this model locally. An M2 Ultra desktop from Apple is affordable and has 128GB of unified RAM+VRAM. If you have one, then llamafile will use your Metal GPU. Try starting out with the Q4_0 quantization level.

Another good option for running large, large language models locally and fully under your control is to just use CPU inference. We developed new tensor multiplication kernels on the llamafile project specifically to speed up "mixture of experts" LLMs like Mixtral. On a AMD Threadripper Pro 7995WX with 256GB of 5200 MT/s RAM, llamafile v0.8 runs Mixtral 8x22B Q4_0 on Linux at 98 tokens per second for evaluation, and it predicts 9.44 tokens per second.

---

Model Card for Mixtral-8x22B-Instruct-v0.1

The Mixtral-8x22B-Instruct-v0.1 Large Language Model (LLM) is an instruct fine-tuned version of the Mixtral-8x22B-v0.1.

Run the model

from transformers import AutoModelForCausalLM
from mistral_common.protocol.instruct.messages import (
    AssistantMessage,
    UserMessage,
)
from mistral_common.protocol.instruct.tool_calls import (
    Tool,
    Function,
)
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.tokens.instruct.normalize import ChatCompletionRequest

device = "cuda" # the device to load the model onto

tokenizer_v3 = MistralTokenizer.v3()

mistral_query = ChatCompletionRequest(
    tools=[
        Tool(
            function=Function(
                name="get_current_weather",
                description="Get the current weather",
                parameters={
                    "type": "object",
                    "properties": {
                        "location": {
                            "type": "string",
                            "description": "The city and state, e.g. San Francisco, CA",
                        },
                        "format": {
                            "type": "string",
                            "enum": ["celsius", "fahrenheit"],
                            "description": "The temperature unit to use. Infer this from the users location.",
                        },
                    },
                    "required": ["location", "format"],
                },
            )
        )
    ],
    messages=[
        UserMessage(content="What's the weather like today in Paris"),
    ],
    model="test",
)

encodeds = tokenizer_v3.encode_chat_completion(mistral_query).tokens
model = AutoModelForCausalLM.from_pretrained("mistralai/Mixtral-8x22B-Instruct-v0.1")
model_inputs = encodeds.to(device)
model.to(device)

generated_ids = model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
sp_tokenizer = tokenizer_v3.instruct_tokenizer.tokenizer
decoded = sp_tokenizer.decode(generated_ids[0])
print(decoded)

Instruct tokenizer

The HuggingFace tokenizer included in this release should match our own. To compare: pip install mistral-common

from mistral_common.protocol.instruct.messages import (
    AssistantMessage,
    UserMessage,
)
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.tokens.instruct.normalize import ChatCompletionRequest

from transformers import AutoTokenizer

tokenizer_v3 = MistralTokenizer.v3()

mistral_query = ChatCompletionRequest(
    messages=[
        UserMessage(content="How many experts ?"),
        AssistantMessage(content="8"),
        UserMessage(content="How big ?"),
        AssistantMessage(content="22B"),
        UserMessage(content="Noice 🎉 !"),
    ],
    model="test",
)
hf_messages = mistral_query.model_dump()['messages']

tokenized_mistral = tokenizer_v3.encode_chat_completion(mistral_query).tokens

tokenizer_hf = AutoTokenizer.from_pretrained('mistralai/Mixtral-8x22B-Instruct-v0.1')
tokenized_hf = tokenizer_hf.apply_chat_template(hf_messages, tokenize=True)

assert tokenized_hf == tokenized_mistral

Function calling and special tokens

This tokenizer includes more special tokens, related to function calling :

• [TOOL_CALLS]
• [AVAILABLE_TOOLS]
• [/AVAILABLE_TOOLS]
• [TOOL_RESULTS]
• [/TOOL_RESULTS]

If you want to use this model with function calling, please be sure to apply it similarly to what is done in our SentencePieceTokenizerV3.

The Mistral AI Team

Albert Jiang, Alexandre Sablayrolles, Alexis Tacnet, Antoine Roux, Arthur Mensch, Audrey Herblin-Stoop, Baptiste Bout, Baudouin de Monicault, Blanche Savary, Bam4d, Caroline Feldman, Devendra Singh Chaplot, Diego de las Casas, Eleonore Arcelin, Emma Bou Hanna, Etienne Metzger, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Harizo Rajaona, Jean-Malo Delignon, Jia Li, Justus Murke, Louis Martin, Louis Ternon, Lucile Saulnier, Lélio Renard Lavaud, Margaret Jennings, Marie Pellat, Marie Torelli, Marie-Anne Lachaux, Nicolas Schuhl, Patrick von Platen, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Thibaut Lavril, Timothée Lacroix, Théophile Gervet, Thomas Wang, Valera Nemychnikova, William El Sayed, William Marshall