llmcompressor supports quantizing fp8 KV Cache for memory savings and inference acceleration with vllm.

fp8 computation is supported on NVIDIA GPUs with compute capability > 8.9 (Ada Lovelace, Hopper).

To get started, install llmcompressor from source as this feature is new:

pip install git+https://github.com/vllm-project/llm-compressor.git@cb98f34d4ec9dd175e6995d12fb02dec39c6f27a

The example includes an end-to-end script for applying the quantization algorithm:

python3 llama3_fp8_kv_example.py

The resulting model Meta-Llama-3-8B-Instruct-FP8-KV is ready to be loaded into vLLM.

Let's walk through the main steps of the quantization process:

1. Load model
2. Prepare calibration data
3. Apply quantization
4. Evaluate and save the model

Load the model using AutoModelForCausalLM:

from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

Prepare the calibration data using the ultrachat dataset:

from datasets import load_dataset

DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048

ds = load_dataset(DATASET_ID, split=f"{DATASET_SPLIT}[:{NUM_CALIBRATION_SAMPLES}]")
ds = ds.shuffle(seed=42)

def process_and_tokenize(example):
    text = tokenizer.apply_chat_template(example["messages"], tokenize=False)
    return tokenizer(text, padding=False, max_length=MAX_SEQUENCE_LENGTH, truncation=True, add_special_tokens=False)

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

Configure and apply the FP8 quantization for weights, activations, and KV cache. Notice the new kv_cache_scheme section:

from llmcompressor import oneshot

recipe = """
quant_stage:
    quant_modifiers:
        QuantizationModifier:
            ignore: ["lm_head"]
            config_groups:
                group_0:
                    weights:
                        num_bits: 8
                        type: float
                        strategy: tensor
                        dynamic: false
                        symmetric: true
                    input_activations:
                        num_bits: 8
                        type: float
                        strategy: tensor
                        dynamic: false
                        symmetric: true
                    targets: ["Linear"]
            kv_cache_scheme:
                num_bits: 8
                type: float
                strategy: tensor
                dynamic: false
                symmetric: true
"""

oneshot(
    model=model,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
)

Test the quantized model with a sample generation:

input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=100)
print(tokenizer.decode(output[0]))

Save the quantized model:

SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-KV"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

For running the model in vLLM, make sure to specify the kv_cache_dtype="fp8" argument to enable quantization of the kv cache, and thus usage of your calibrated scales.

To evaluate the accuracy of your quantized model:

1. Install vllm and lm-evaluation-harness:

pip install "vllm>=0.5.5" lm_eval==0.4.3

2. Run an evaluation (e.g., on GSM-8K):

MODEL=$PWD/Meta-Llama-3-8B-Instruct-FP8-KV
lm_eval \
  --model vllm \
  --model_args pretrained=$MODEL,kv_cache_dtype=fp8,add_bos_token=True \
  --tasks gsm8k --num_fewshot 5 --batch_size auto

vllm (pretrained=Meta-Llama-3-8B-Instruct-FP8-KV,kv_cache_dtype=fp8,add_bos_token=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.7748|±  |0.0115|
|     |       |strict-match    |     5|exact_match|↑  |0.7763|±  |0.0115|

Note: Include add_bos_token=True as quantized models can be sensitive to the presence of the bos token.

Please open an issue on vllm-project/llm-compressor.