Quark (original) (raw)
Quark is a deep learning quantization toolkit designed to be agnostic to specific data types, algorithms, and hardware. Different pre-processing strategies, algorithms and data-types can be combined in Quark.
The PyTorch support integrated through 🤗 Transformers primarily targets AMD CPUs and GPUs, and is primarily meant to be used for evaluation purposes. For example, it is possible to use lm-evaluation-harness with 🤗 Transformers backend and evaluate a wide range of models quantized through Quark seamlessly.
Users interested in Quark can refer to its documentation to get started quantizing models and using them in supported open-source libraries!
Although Quark has its own checkpoint / configuration format, the library also supports producing models with a serialization layout compliant with other quantization/runtime implementations (AutoAWQ, native fp8 in 🤗 Transformers).
To be able to load Quark quantized models in Transformers, the library first needs to be installed:
Support matrix
Models quantized through Quark support a large range of features, that can be combined together. All quantized models independently of their configuration can seamlessly be reloaded through PretrainedModel.from_pretrained.
The table below shows a few features supported by Quark:
| Feature | Supported subset in Quark |
|---|---|
| Data types | int8, int4, int2, bfloat16, float16, fp8_e5m2, fp8_e4m3, fp6_e3m2, fp6_e2m3, fp4, OCP MX, MX6, MX9, bfp16 |
| Pre-quantization transformation | SmoothQuant, QuaRot, SpinQuant, AWQ |
| Quantization algorithm | GPTQ |
| Supported operators | nn.Linear, nn.Conv2d, nn.ConvTranspose2d, nn.Embedding, nn.EmbeddingBag |
| Granularity | per-tensor, per-channel, per-block, per-layer, per-layer type |
| KV cache | fp8 |
| Activation calibration | MinMax / Percentile / MSE |
| Quantization strategy | weight-only, static, dynamic, with or without output quantization |
Models on Hugging Face Hub
Public models using Quark native serialization can be found at https://huggingface.co/models?other=quark.
Although Quark also supports models using quant_method="fp8" and models using quant_method="awq", Transformers loads these models rather through AutoAWQ or uses the native fp8 support in 🤗 Transformers.
Using Quark models in Transformers
Here is an example of how one can load a Quark model in Transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "EmbeddedLLM/Llama-3.1-8B-Instruct-w_fp8_per_channel_sym" model = AutoModelForCausalLM.from_pretrained(model_id) model = model.to("cuda")
print(model.model.layers[0].self_attn.q_proj)
tokenizer = AutoTokenizer.from_pretrained(model_id) inp = tokenizer("Where is a good place to cycle around Tokyo?", return_tensors="pt") inp = inp.to("cuda")
res = model.generate(**inp, min_new_tokens=50, max_new_tokens=100)
print(tokenizer.batch_decode(res)[0])