-
-
Notifications
You must be signed in to change notification settings - Fork 421
Quantization
Quantization is a process of:
- storage-optimization
reducing the memory footprint of the model by reducing the precision of parameters in a model - compute-optimization
speed up the inference process by providing optimized kernels for native execution in quantized precision
For storage-only quantization, the model is quantized to lower precision but the operations are still performed in the original precision which means that each operation needs to be upcasted to the original precision before execution resulting in a performance overhead.
Quantization can be done in multiple ways:
on-the-fly- by quantizing on-the-fly during model load
available by selecting settings -> quantization for some quantization types
- by quantizing immediately after model load
available by selecting settings -> quantization for all quantization types - by simply loading a pre-quantized model
quantization type will be auto-determined at the start of the load - during model training itself
out-of-scope for this document
Tip
If you're on Windows with a compatible GPU, you may try WSL2 for broader feature compatibiliy
See WSL Wiki for more details
Typical models pre-quantized with bitsandbytes
would have look like *nf4.safetensors
or *fp8.safetensors
Note
BnB is the only quantization method that allows for usage of balanced offload as well as quantization on-the-fly during load, thus it is considered most versatile choice, but it is not available on all platforms.
Limitations:
- default
bitsandbytes
package only supports nVidia GPUs
some quantization types require newer GPU with supported CUDA ops: e.g. nVidia Turing GPUs or newer -
bitsandbytes
relies ontriton
packages which are not available on windows unless manually compiled/installed
without them, performance is significantly reduced - for AMD/ROCm: link
- for Intel/IPEX: link
Typical models pre-quantized with optimum.quanto
would have look like *qint.safetensors
.
Note
OQ is highly efficient with its qint8/qint3 quantization types, but its usage is limited to specific platforms and cannot be used with broad offloading methods
Limitations:
- requires
torch==2.4.0
if you're running older torch, you can try upgrading it or running sdnext with--reinstall
flag - not compatible with balanced offload
- not supported on Intel Arc/IPEX since IPEX is still based on Torch 2.1
- not supported with Zluda since Zluda does not support torch 2.4
- not supported on AMD ROCm on Linux using official package due to explicit CUDA checks
for AMD/ROCm: link
GGUF is a binary file format used to package pre-quantized models.
GGUF is originally desiged by llama.cpp
project and intended to be used with its GGML execution runtime.
However, without GGML, GGUF provides storage-only quantization which means that every operation needs to be upcast to current device precision before execution (typically FP16 or BF16) which comes with a significant performance overhead.
Warning
Right now, all popular T2I inference UIs (SD.Next, Forge, ComfyUI, InvokeAI etc.) are using GGUF as storage-only and as such usage of GGUF is not recommended!
-
gguf
supports wide range of quantization types and is not platform or GPU dependent -
gguf
does not provide native GPU kernels which means thatgguf
is purely a storage optimization -
gguf
reduces model size and memory usage, but it does slow down model inference since all quantized weights are de-quantized on-the-fly
Limitations:
-
gguf
is not compatible with model offloading as it would trigger de-quantization -
note: only supported component in
gguf
binary format is UNET/Transformer
you cannot load all-in-one single-file GGUF model
NNCF provides full cross-platform storage-only quantization (referred to as model compression)
with optional platform-specific compute-optimization (available only on OpenVINO platform)
Note
Advantage of NNCF is that does works on any platform: if you're having issues with optimum-quanto
or bitsandbytes
, try it out!
- broad platform and GPU support
- enable in Settings -> Compute -> Compress model weights with NNCF
- see NNCF Wiki for more details
Caution
Using incompatible configurations will result in errors during model load:
- BitsAndBytes nf4 quantization is not compatible with sequential offload
Error: Blockwise quantization only supports 16/32-bit floats
- Quanto qint quantization is not compatible with balanced offload
Error: QBytesTensor.new() missing 5 required positional arguments
- Quanto qint quantization is not compatible with sequential offload
Error: Expected all tensors to be on the same device