[Bugfix] Support model offloading SparseGPTQ

[kylesayrs]

Purpose

• Fix bug related to running calibration with offloaded models
• Add offloading support to SparseGPTQ
• Add additional logging to SparseGPTQ to validate sparsity

Changes

• Update logic which determines which device to put calibration inputs on
  • New logic incorporates offloading assumptions. Previously, if the model was offloaded, the logic would attempt to put the inputs on the meta device

NotImplementedError: Cannot copy out of meta tensor; no data!

• Add weight onloading and offloading to SparseGPTQ

if is_module_offloaded(self.layer):
    self.layer._hf_hook.pre_forward(self.layer)

• Added sparsity logging to SparseGPTQ for algorithm validation

===== Compressing layer 110/113 to sparsity 0.5 =====                                                                  
2024-11-15T23:56:01.649029+0000 | compress_module | INFO - Compressing model.layers.15.mlp.gate_proj.model.layers.15.mlp.gate_proj...
2024-11-15T23:56:02.016416+0000 | compress | INFO - time 0.36
2024-11-15T23:56:02.016682+0000 | compress | INFO - error 39350.72
2024-11-15T23:56:02.072700+0000 | compress | INFO - sparsity 0.50
2024-11-15T23:56:02.094924+0000 | apply_compression | INFO -

Testing

llama_example.py

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

from llmcompressor.modifiers.obcq import SparseGPTModifier
from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot

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

model = SparseAutoModelForCausalLM.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 = 2  # 512
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.
#   * quantize the weights to 4 bit with GPTQ with a group size 128
recipe = SparseGPTModifier(targets="Linear", sparsity=0.5, ignore=["lm_head"])

# 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)

[github-actions]

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