compressed-tensors

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.

Why `compressed-tensors`?

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.

Installation

From PyPI

Stable release:

pip install compressed-tensors

Nightly release:

pip install compressed-tensors-nightly

From Source

git clone https://github.com/neuralmagic/compressed-tensors
cd compressed-tensors
pip install -e .

Getting started

Saving/Loading Compressed Tensors (Bitmask Compression)

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

Saving/Loading Compressed Models (Bitmask Compression)

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.

Saving a Compressed Model with PTQ

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.

Metadata

Release files for compressed-tensors-nightly 0.8.1.20241214

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for compressed-tensors-nightly 0.8.1.20241214
File	Size	Uploaded
compressed-tensors-nightly-0.8.1.20241214.tar.gz	56.8 kB	Dec 14, 2024	Details

Built distribution (wheel)

Table of built distributions (wheels) for compressed-tensors-nightly 0.8.1.20241214
File	Interpreter	ABI	Platform	Reset
compressed_tensors_nightly-0.8.1.20241214-py3-none-any.whl 87.7 kB Dec 14, 2024	Python 3	none	any	Details

Total release size: 144.5 kB

Release files / compressed-tensors-nightly-0.8.1.20241214.tar.gz

Download URL	compressed-tensors-nightly-0.8.1.20241214.tar.gz
Size	56.8 kB
Tags	Source
SHA-256 checksum How to use checksums	`dbf3c0006467fe562f8bc5272ddfc09834400cbe1e89a8623d0dc5c35d7354b3`
BLAKE2b-256 checksum How to use checksums	`9d659dc90b3994434baa21f85ee3e63e2e92781321ba72cedcd7fed70fbb17bb`
Upload date	Dec 14, 2024
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.0.1 CPython/3.10.12`

Release files / compressed_tensors_nightly-0.8.1.20241214-py3-none-any.whl

Download URL	compressed_tensors_nightly-0.8.1.20241214-py3-none-any.whl
Size	87.7 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`747b1e2a2680986b5d2efedc36d514c9e4761bb291c24bc912d3ecbdda5031d4`
BLAKE2b-256 checksum How to use checksums	`e3695c3876fb871553e10948707c3fdd1ecb67a2e6b8a1debcc502ca7020dc38`
Upload date	Dec 14, 2024
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.0.1 CPython/3.10.12`