metadata

tags:
  - fp8
  - vllm
license: other
license_name: bigcode-openrail-m
license_link: https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement

starcoder2-7b-FP8

Model Overview

Model Architecture: starcoder2-7b
- Input: Text
- Output: Text
Model Optimizations:
- Weight quantization: FP8
- Activation quantization: FP8
Intended Use Cases: Intended for commercial and research use in English.
Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English.
Release Date: 8/1/2024
Version: 1.0
License(s): bigcode-openrail-m
Model Developers: Neural Magic

Quantized version of starcoder2-7b.

It achieves an average score of 39.30 on the HumanEval+ benchmark, whereas the unquantized model achieves 39.65.

Model Optimizations

This model was obtained by quantizing the weights and activations of starcoder2-7b to FP8 data type, ready for inference with vLLM >= 0.5.2. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%.

Only the weights and activations of the linear operators within transformers blocks are quantized. Symmetric per-tensor quantization is applied, in which a single linear scaling maps the FP8 representations of the quantized weights and activations. AutoFP8 is used for quantization with 512 sequences of UltraChat.

Creation

This model was created by applying LLM Compressor with calibration samples from UltraChat, as presented in the code snipet below. A slight modification to the code was made due to the parameters of the model. Running the below code will throw an index error, and simply replacing the erroneous line with max_quant_shape = param.shape[0] resolves the issue.

import torch
from datasets import load_dataset
from transformers import AutoTokenizer

from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
from llmcompressor.transformers.compression.helpers import (
    calculate_offload_device_map,
    custom_offload_device_map,
)

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"]
"""

model_stub = "bigcode/starcoder2-7b"
model_name = model_stub.split("/")[-1]

device_map = calculate_offload_device_map(
    model_stub, reserve_for_hessians=False, num_gpus=8, torch_dtype=torch.float16
)

model = SparseAutoModelForCausalLM.from_pretrained(
    model_stub, torch_dtype=torch.float16, device_map=device_map
)
tokenizer = AutoTokenizer.from_pretrained(model_stub)

output_dir = f"./{model_name}-FP8"

DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 4096

ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))

def preprocess(example):
    return {
        "text": " ".join([msg["content"] for msg in example["messages"]])
    }

ds = ds.map(preprocess)

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)

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

Evaluation

The model was evaluated on the HumanEval+ benchmark with the Neural Magic fork of the EvalPlus implementation of HumanEval+ and the vLLM engine, using the following command:

python codegen/generate.py --model neuralmagic/starcoder2-7b-FP8 --temperature 0.2 --n_samples 50 --resume --root ~ --dataset humaneval
python evalplus/sanitize.py ~/humaneval/neuralmagic--starcoder2-7b-FP8_vllm_temp_0.2
evalplus.evaluate --dataset humaneval --samples ~/humaneval/neuralmagic--starcoder2-7b-FP8_vllm_temp_0.2-sanitized

Accuracy

HumanEval+ evaluation scores

Benchmark	starcoder2-7b	starcoder2-7b-FP8(this model)	Recovery
base pass@1	34.9	34.6	99.14%
base pass@10	50.7	50.1	98.82%
base+extra pass@1	30.0	30.3	101.00%
base+extra pass@10	43.0	42.2	98.14%
Average	39.65	39.30	99.27%