Implement aliasable mixin and alias activation ordering

[kylesayrs]

Purpose

• Add aliases for activation ordering options to improve usability for researchers familiar with autogptq

Changes

• Implement AliasableEnum mixin which allows enums to be aliased
• Add AliasableEnum to ActivationOrdering with the following map:

{
    "dynamic": "group",
    "static": "weight",
}

Testing

• Added passing tests in tests/test_quantization/test_quant_args.py
• Successfully quantized a model using dynamic actorder and tested e2e with vllm

llama3.py

from accelerate import cpu_offload
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot

# Select model and load it.
# MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
MODEL_ID = "meta-llama/Llama-3.2-1B-Instruct"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="cuda:0",
    torch_dtype="auto",
)
# cpu_offload(model)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# Select calibration dataset.
DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"

# Select number of samples. 512 samples is a good place to start.
# Increasing the number of samples can improve accuracy.
NUM_CALIBRATION_SAMPLES = 285  # 2048
MAX_SEQUENCE_LENGTH = 2048

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))


def preprocess(example):
    return {
        "text": tokenizer.apply_chat_template(
            example["messages"],
            tokenize=False,
        )
    }


ds = ds.map(preprocess)


# Tokenize inputs.
def tokenize(sample):
    return tokenizer(
        sample["text"],
        padding=False,
        max_length=MAX_SEQUENCE_LENGTH,
        truncation=True,
        add_special_tokens=False,
    )


ds = ds.map(tokenize, remove_columns=ds.column_names)

# Configure the quantization algorithm to run.
from compressed_tensors.quantization import QuantizationArgs, QuantizationType, QuantizationStrategy, ActivationOrdering, QuantizationScheme
recipe = GPTQModifier(
    targets="Linear",
    config_groups={
        "config_group": QuantizationScheme(
            targets=["Linear"],
            weights=QuantizationArgs(
                num_bits=4,
                type=QuantizationType.INT,
                strategy=QuantizationStrategy.GROUP,
                group_size=128,
                symmetric=True,
                dynamic=False,
                actorder="dynamic",
            ),
        ),
    },
    ignore=["lm_head"],
    dampening_frac=0.5
)

# Apply algorithms.
oneshot(
    model=model,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
)

# Confirm generations of the quantized model look sane.
print("\n\n")
print("========== SAMPLE GENERATION ==============")
input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
output = model.generate(input_ids, max_new_tokens=100)
print(tokenizer.decode(output[0]))
print("==========================================\n\n")

# Save to disk compressed.
SAVE_DIR = MODEL_ID.split("/")[1] + "-W4A16-G128"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

[horheynm]

lgtm,
for aliases on enum, is this the only way? Looks pretty complicated with setting hash for dicts

[kylesayrs]

@horheynm In order to support comparisons natively, at least these two methods must be overloaded

[rahul-tuli]

LGTM, pending response to comments!

[dsikka]

This LGTM. Seems like our changes to quant args would be backwards compatible but just wanted to confirm that this would be the case?

[kylesayrs]

@dsikka Yes, I am confident that any quantization config produced prior to these changes will be compatible with this new definition.

[dsikka]

This is causing llm-compressor tests to fail: https://github.com/vllm-project/llm-compressor/actions/runs/12090087246/job/33716500170?pr=945
It has been reverted: #217

[kylesayrs]

The failing tests pass locally with this branch. This points to it being related to different environments, may be a bug present in specific pydantic versions

[kylesayrs]

I was able to replicate on python3.9

python package environment

$ python3 -m pip list
Package                   Version     Editable project location
------------------------- ----------- --------------------------------
annotated-types           0.7.0
attrs                     24.2.0
beautifulsoup4            4.12.3
black                     22.12.0
bleach                    6.2.0
certifi                   2024.8.30
charset-normalizer        3.4.0
click                     8.1.7
compressed-tensors        0.8.0       /home/ksayers/compressed-tensors
defusedxml                0.7.1
exceptiongroup            1.2.2
fastjsonschema            2.21.0
filelock                  3.16.1
flake8                    7.1.1
fsspec                    2024.10.0
huggingface-hub           0.26.3
idna                      3.10
importlib_metadata        8.5.0
iniconfig                 2.0.0
isort                     5.8.0
Jinja2                    3.1.4
jsonschema                4.23.0
jsonschema-specifications 2024.10.1
jupyter_client            8.6.3
jupyter_core              5.7.2
jupyterlab_pygments       0.3.0
MarkupSafe                3.0.2
mccabe                    0.7.0
mistune                   3.0.2
mpmath                    1.3.0
mypy-extensions           1.0.0
nbclient                  0.10.1
nbconvert                 7.16.4
nbformat                  5.10.4
networkx                  3.2.1
numpy                     2.0.2
nvidia-cublas-cu12        12.4.5.8
nvidia-cuda-cupti-cu12    12.4.127
nvidia-cuda-nvrtc-cu12    12.4.127
nvidia-cuda-runtime-cu12  12.4.127
nvidia-cudnn-cu12         9.1.0.70
nvidia-cufft-cu12         11.2.1.3
nvidia-curand-cu12        10.3.5.147
nvidia-cusolver-cu12      11.6.1.9
nvidia-cusparse-cu12      12.3.1.170
nvidia-nccl-cu12          2.21.5
nvidia-nvjitlink-cu12     12.4.127
nvidia-nvtx-cu12          12.4.127
packaging                 24.2
pandocfilters             1.5.1
pathspec                  0.12.1
pip                       23.0.1
platformdirs              4.3.6
pluggy                    1.5.0
pycodestyle               2.12.1
pydantic                  2.10.2
pydantic_core             2.27.1
pyflakes                  3.2.0
Pygments                  2.18.0
pytest                    8.3.3
python-dateutil           2.9.0.post0
PyYAML                    6.0.2
pyzmq                     26.2.0
referencing               0.35.1
regex                     2024.11.6
requests                  2.32.3
rpds-py                   0.21.0
safetensors               0.4.5
setuptools                58.1.0
six                       1.16.0
soupsieve                 2.6
sympy                     1.13.1
tinycss2                  1.4.0
tokenizers                0.20.3
tomli                     2.2.1
torch                     2.5.1
tornado                   6.4.2
tqdm                      4.67.1
traitlets                 5.14.3
transformers              4.46.3
triton                    3.1.0
typing_extensions         4.12.2
urllib3                   2.2.3
webencodings              0.5.1
wheel                     0.45.1
zipp                      3.21.0

Python 3.9.20 (main, Sep  7 2024, 18:35:25) 
[GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from compressed_tensors.quantization import ActivationOrdering
>>> ActivationOrdering.DYNAMIC == ActivationOrdering.GROUP
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/ksayers/compressed-tensors/src/compressed_tensors/utils/helpers.py", line 144, in __eq__
    self.aliases.get(self.value, self.value)
TypeError: 'staticmethod' object is not callable

Potentially related to https://bugs.python.org/issue43682