The compressed-tensors
library extends the safetensors format, providing a versatile and efficient way to store and manage compressed tensor data. This library supports various quantization and sparsity schemes, making it a unified format for handling different model optimizations like GPTQ, AWQ, SmoothQuant, INT8, FP8, SparseGPT, and more.
As model compression becomes increasingly important for efficient deployment of LLMs, the landscape of quantization and compression techniques has become increasingly fragmented.
Each method often comes with its own storage format and loading procedures, making it challenging to work with multiple techniques or switch between them.
compressed-tensors
addresses this by providing a single, extensible format that can represent a wide variety of compression schemes.
- Unified Checkpoint Format: Supports various compression schemes in a single, consistent format.
- Wide Compatibility: Works with popular quantization methods like GPTQ, SmoothQuant, and FP8. See llm-compressor
- Flexible Quantization Support:
- Weight-only quantization (e.g., W4A16, W8A16, WnA16)
- Activation quantization (e.g., W8A8)
- KV cache quantization
- Non-uniform schemes (different layers can be quantized in different ways!)
- Sparsity Support: Handles both unstructured and semi-structured (e.g., 2:4) sparsity patterns.
- Open-Source Integration: Designed to work seamlessly with Hugging Face models and PyTorch.
This allows developers and researchers to easily experiment with composing different quantization methods, simplify model deployment pipelines, and reduce the overhead of supporting multiple compression formats in inference engines.
From PyPI
Stable release:
pip install compressed-tensors
Nightly release:
pip install compressed-tensors-nightly
git clone https://github.com/neuralmagic/compressed-tensors
cd compressed-tensors
pip install -e .
The function save_compressed
uses the compression_format
argument to apply compression to tensors.
The function load_compressed
reverses the process: converts the compressed weights on disk to decompressed weights in device memory.
from compressed_tensors import save_compressed, load_compressed, BitmaskConfig
from torch import Tensor
from typing import Dict
# the example BitmaskConfig method efficiently compresses
# tensors with large number of zero entries
compression_config = BitmaskConfig()
tensors: Dict[str, Tensor] = {"tensor_1": Tensor(
[[0.0, 0.0, 0.0],
[1.0, 1.0, 1.0]]
)}
# compress tensors using BitmaskConfig compression format (save them efficiently on disk)
save_compressed(tensors, "model.safetensors", compression_format=compression_config.format)
# decompress tensors (load_compressed returns a generator for memory efficiency)
decompressed_tensors = {}
for tensor_name, tensor in load_compressed("model.safetensors", compression_config = compression_config):
decompressed_tensors[tensor_name] = tensor
We can apply bitmask compression to a whole model. For more detailed example see example
directory.
from compressed_tensors import save_compressed_model, load_compressed, BitmaskConfig
from transformers import AutoModelForCausalLM
model_name = "neuralmagic/llama2.c-stories110M-pruned50"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto")
original_state_dict = model.state_dict()
compression_config = BitmaskConfig()
# save compressed model weights
save_compressed_model(model, "compressed_model.safetensors", compression_format=compression_config.format)
# load compressed model weights (`dict` turns generator into a dictionary)
state_dict = dict(load_compressed("compressed_model.safetensors", compression_config))
For more in-depth tutorial on bitmask compression, refer to the notebook.
We can use compressed-tensors to run basic post training quantization (PTQ) and save the quantized model compressed on disk
model_name = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cuda:0", torch_dtype="auto")
config = QuantizationConfig.parse_file("./examples/bit_packing/int4_config.json")
config.quantization_status = QuantizationStatus.CALIBRATION
apply_quantization_config(model, config)
dataset = load_dataset("ptb_text_only")["train"]
tokenizer = AutoTokenizer.from_pretrained(model_name)
def tokenize_function(examples):
return tokenizer(examples["sentence"], padding=False, truncation=True, max_length=1024)
tokenized_dataset = dataset.map(tokenize_function, batched=True)
data_loader = DataLoader(tokenized_dataset, batch_size=1, collate_fn=DefaultDataCollator())
with torch.no_grad():
for idx, sample in tqdm(enumerate(data_loader), desc="Running calibration"):
sample = {key: value.to(device) for key,value in sample.items()}
_ = model(**sample)
if idx >= 512:
break
model.apply(freeze_module_quantization)
model.apply(compress_quantized_weights)
output_dir = "./ex_llama1.1b_w4a16_packed_quantize"
compressor = ModelCompressor(quantization_config=config)
compressed_state_dict = compressor.compress(model)
model.save_pretrained(output_dir, state_dict=compressed_state_dict)
For more in-depth tutorial on quantization compression, refer to the notebook.