Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Universal Quantized File Format: UQFF

The uniquely powerful quantized file format.

  1. Flexible 🌀: Multiple quantization formats in one file format with one framework to run them all.
  2. Reliable 🔒: Compatibility ensured with embedded and checked semantic versioning information from day 1.
  3. Easy 🤗: Download UQFF models easily and quickly from Hugging Face, or use a local file.
  4. Customizable 🛠️: Make and publish your own UQFF files in minutes.

ToC

Motivation

UQFF builds on our ISQ feature by allowing serialization and deserialization for models.

While ISQ is a powerful feature enabling easy quantization of models, the key limitation has been the time required for requantization. While the process is relatively fast with parallelization and other techniques, multiple runs can make the experience slow.

Comparting UQFF to GGUF:

In contrast to GGUF, which only supports the GGUF quantizations, UQFF is designed with flexibiliuty in mind. At its code, it extends the power and flexibility of ISQ. The ability to support multiple quantization types (more to come!) in one simple, easy-to-use file is a critical feature.

Additionally, users will no longer need to wait for GGUF support to begin using post-training quantized models. As we add new models and quantization schemes to mistral.rs, the feature set of UQFF will grow.

Support

The following quantization formats are supported in UQFF. One can, of course, be combined arbitrarily during UQFF generation or ISQ using a model topology. When loading a UQFF model, only the per-layer device mapping feature of the topology applies.

  • GGUF quantized:

    • Q4_0
    • Q4_1
    • Q5_0
    • Q5_1
    • Q8_0
    • Q8_1 (not available on CUDA)
    • Q2K
    • Q3K
    • Q4K
    • Q5K
    • Q6K
    • Q8K (not available on CUDA)

  • HQQ quantized:

    • HQQ4
    • HQQ8

  • FP8:

    • FP8 E4M3 (4-bit exponent, 3-bit mantissa)

  • AFQ quantized (🔥 AFQ is fast on Metal):

    • AFQ2
    • AFQ3
    • AFQ4
    • AFQ6
    • AFQ8

  • F8Q8:

    • F8Q8

Loading a UQFF model

To load a UQFF model, specify the filename of the first (or only) UQFF shard. This will be located based on the model ID, and can be loaded locally or from Hugging Face based on the model ID.

  • phi3.5-mini-instruct-q4k-0.uqff
  • ../UQFF/phi3.5-mini-instruct-q4k-0.uqff

You can find a collection of UQFF models here, which each include a simple command to get started.

Note: when loading an UQFF model, any ISQ setting will be ignored.

Shard auto-discovery

Large models produce multiple shard files (e.g., q4k-0.uqff, q4k-1.uqff, q4k-2.uqff). You only need to specify one shard file – the remaining shards are auto-discovered from the same directory or Hugging Face repository.

For example, if a model has shards q4k-0.uqff, q4k-1.uqff, and q4k-2.uqff:

# Just specify the first shard -- the rest are found automatically
mistralrs run -m EricB/MyModel-UQFF --from-uqff q4k-0.uqff

This also works when multiple quantizations exist in the same repo (e.g., q4k-* and q8_0-*). Only the shards matching the specified prefix are loaded.

Running with the CLI

mistralrs run -m EricB/Phi-3.5-mini-instruct-UQFF --from-uqff phi3.5-mini-instruct-f8e4m3-0.uqff

Using with the Rust SDK

Check out the following examples:

Using the Python SDK

Modify the Which instantiation as follows:

Which.Plain(
    model_id="EricB/Phi-3.5-mini-instruct-UQFF",
+   from_uqff="phi3.5-mini-instruct-q4k-0.uqff"
),

Using topology for device mapping with UQFF

When loading a UQFF model, the quantization is already baked in, so ISQ settings in the topology are ignored. However, device mapping from a topology file still applies. This is useful for splitting a pre-quantized model across multiple GPUs or offloading layers to CPU.

CLI example:

mistralrs run -m EricB/Phi-3.5-mini-instruct-UQFF --from-uqff phi3.5-mini-instruct-q4k.uqff --topology device_map.yml

Topology file for device mapping only (device_map.yml):

0-16:
  device: cuda[0]
16-32:
  device: cuda[1]

Rust SDK example:

#![allow(unused)]
fn main() {
use mistralrs::{UqffTextModelBuilder, Topology, LayerTopology, Device};

let model = UqffTextModelBuilder::new(
    "EricB/Phi-3.5-mini-instruct-UQFF",
    vec!["phi3.5-mini-instruct-q4k.uqff".into()],
)
.into_inner()
.with_topology(
    Topology::empty()
        .with_range(0..16, LayerTopology { isq: None, device: Some(Device::Cuda(0)) })
        .with_range(16..32, LayerTopology { isq: None, device: Some(Device::Cuda(1)) })
)
.build()
.await?;
}

Python SDK example:

runner = Runner(
    which=Which.Plain(
        model_id="EricB/Phi-3.5-mini-instruct-UQFF",
        from_uqff="phi3.5-mini-instruct-q4k.uqff",
        topology="device_map.yml",
    ),
)

Note: The isq field in topology entries is ignored when loading UQFF models since quantization is pre-applied.

Creating a UQFF model

Creating a UQFF model requires you to generate the UQFF file.

  • Specify an output path: either a .uqff file path or a directory where files will be auto-named.
  • The quantization of a UQFF model is determined from the ISQ or model topology (see the topology docs for more details on how ISQ and the topology mix).

Along with the UQFF file, the generation process will also output several .json configuration files and residual.safetensors. All of these files are considered the UQFF model, and should be kept together or uploaded.

Note: Only the .uqff files are unique to the quantization level(s). If you are generating multiple UQFF files, it is OK for the others to be overwritten.

Single quantization (file output):

mistralrs quantize -m microsoft/Phi-3.5-mini-instruct --isq q4k -o phi3.5-uqff/phi3.5-mini-instruct-q4k.uqff

Single quantization (directory output):

mistralrs quantize -m microsoft/Phi-3.5-mini-instruct --isq q4k -o phi3.5-uqff/

Multiple quantizations at once (directory output):

Generate multiple UQFF files by specifying multiple --isq types. All quantizations go to the same output directory.

# Comma-separated ISQ types
mistralrs quantize -m microsoft/Phi-3.5-mini-instruct --isq q4k,q8_0 -o phi3.5-uqff/

# Equivalent: repeated --isq flags
mistralrs quantize -m microsoft/Phi-3.5-mini-instruct --isq q4k --isq q8_0 -o phi3.5-uqff/

This produces the following in phi3.5-uqff/:

  • q4k-0.uqff (and additional shards q4k-1.uqff, … if the model is large)
  • q8_0-0.uqff (and additional shards if needed)
  • README.md (auto-generated model card for Hugging Face)
  • Shared files: config.json, tokenizer.json, residual.safetensors, etc.

Note: Multiple --isq values require a directory output path (not a .uqff file path).

Model card generation

When using directory output mode, the quantize command automatically generates a README.md model card in the output directory. This model card includes Hugging Face YAML frontmatter, a description, and an examples table with the appropriate --from-uqff commands for each quantization.

To skip model card generation, use --no-readme:

mistralrs quantize -m microsoft/Phi-3.5-mini-instruct --isq q4k -o phi3.5-uqff/ --no-readme

Uploading to Hugging Face

After quantization completes in directory mode, the quantize command prints the huggingface-cli upload command you can use. The general form is:

huggingface-cli upload <YOUR_USERNAME>/<MODEL_NAME>-UQFF <output_dir> --repo-type model --private

Alternatively, you can upload with Git LFS:

  1. Install git-lfs
  2. Run git lfs install
  3. (If the files are larger than 5GB) Run huggingface-cli lfs-enable-largefiles . (you will need to pip install huggingface_hub)

After this, you can use Git to track, commit, and push files.

List of models

You can find a list of models in the Hugging Face model collection.

Have you created a UQFF model on Hugging Face? If so, please create an issue.