Dataset split fallbacks

[kylesayrs]

Purpose

• Allow fallback dataset splits to be used for compression tasks requiring specific dataset splits
• Allow datasets without a calibration split (which is most datasets) to be used in oneshot without explicit preprocessing

Previously

oneshot(
    dataset="some_dataset",
    split="test[:128]"  # intended {"calibration": "test[:128]"}
)

ValueError: --do_oneshot requires a calibration dataset

Now

oneshot(
    dataset="some_dataset",
    split="test[:128]"
)

UserWarning: oneshot expects one of ['calibration', 'train'] dataset split, falling back to test.
Use splits={"calibration": "test"} to silence this warning

Changes

• Added typing definitions in src/llmcompressor/typing.py
• Implemented _get_split_with_fallbacks helper function which allows strict and non-strict split retrieval with fallbacks

Postrequisites

• Vision Datasets #943
• Simplify Examples #954

Testing

• Added tests in tests/llmcompressor/transformers/finetune/data/test_dataset_helpers.py

test.py

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 = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
    torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# Select calibration dataset.
DATASET_ID = "ultrachat-200k"
DATASET_SPLIT = "train_sft[:512]"

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

# Configure the quantization algorithm to run.
#   * quantize the weights to 4 bit with GPTQ with a group size 128
recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"])

# Apply algorithms.
oneshot(
    model=model,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    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)

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

[kylesayrs]

This change isn't strictly necessary for #943. I think the original intention was that users would manually assign dataset splits, ie oneshot(splits={"calibration": "train[:128]"}. How to do this isn't immediately obvious from the error we currently give, so automatic split inference is a nice to have.

[dsikka]

If this is a nice to have, let's not prioritize for now.
We can follow-up.