metadata

pipeline_tag: text-generation
datasets:
  - bigcode/the-stack-v2-train
license: bigcode-openrail-m
library_name: transformers
tags:
  - code
model-index:
  - name: starcoder2-15b-quantized.w8a16
    results:
      - task:
          type: text-generation
        dataset:
          name: HumanEval+
          type: humanevalplus
        metrics:
          - type: pass@1
            value: 37.6
      - task:
          type: text-generation
        dataset:
          name: HumanEval
          type: humaneval
        metrics:
          - type: pass@1
            value: 44.3

starcoder2-15b-quantized.w8a16

Model Overview

Model Architecture: StarCoder2
- Input: Text
- Output: Text
Model Optimizations:
- Weight quantization: INT8
Intended Use Cases: Intended for commercial and research use. Similarly to starcoder2-15b, this model is intended for code generation and is not an instruction model. Commands like "Write a function that computes the square root." do not work well.
Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws).
Release Date: 8/1/2024
Version: 1.0
License(s): bigcode-openrail-m
Model Developers: Neural Magic

Quantized version of starcoder2-15b. It achieves a HumanEval pass@1 of 44.3, whereas the unquantized model achieves 44.8 when evaluated under the same conditions.

Model Optimizations

This model was obtained by quantizing the weights of starcoder2-15b to INT8 data type. 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 of the linear operators within transformers blocks are quantized. Symmetric per-channel quantization is applied, in which a linear scaling per output dimension maps the INT8 and floating point representations of the quantized weights. The GPTQ algorithm is applied for quantization, as implemented in the llm-compressor library. GPTQ used a 1% damping factor and 256 sequences of 8,192 random tokens.

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "neuralmagic/starcoder2-15b-quantized.w8a16"
number_gpus = 1

sampling_params = SamplingParams(temperature=0.2, top_p=0.95, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)

prompts = ["def print_hello_world():"]

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

vLLM aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

This model was created by using the llm-compressor library as presented in the code snipet below.

from transformers import AutoTokenizer
from datasets import Dataset
from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
import random

model_id = "bigcode/starcoder2-15b"

num_samples = 256
max_seq_len = 8192

tokenizer = AutoTokenizer.from_pretrained(model_id)

max_token_id = len(tokenizer.get_vocab()) - 1
input_ids = [[random.randint(0, max_token_id) for _ in range(max_seq_len)] for _ in range(num_samples)]
attention_mask = num_samples * [max_seq_len * [1]]
ds = Dataset.from_dict({"input_ids": input_ids, "attention_mask": attention_mask})

recipe = GPTQModifier(
  targets="Linear",
  scheme="W8A16",
  ignore=["lm_head"],
  dampening_frac=0.01,
)

model = SparseAutoModelForCausalLM.from_pretrained(
  model_id,
  device_map="auto",
  trust_remote_code=True,
)

oneshot(
  model=model,
  dataset=ds,
  recipe=recipe,
  max_seq_length=max_seq_len,
  num_calibration_samples=num_samples,
)
model.save_pretrained("starcoder2-15b-quantized.w8a16")

Evaluation

The model was evaluated on the HumanEval and HumanEval+ benchmarks, using the generation configuration from Big Code Models Leaderboard. We used Neural Magic's fork of evalplus and the vLLM engine, using the following commands:

python codegen/generate.py \
  --model neuralmagic/starcoder2-15b-quantized.w8a16 \
  --bs 8 \
  --temperature 0.2 \
  --n_samples 50 \
  --dataset humaneval \
  -- root "."

python3 evalplus/sanitize.py humaneval/neuralmagic--starcoder2-15b-quantized.w8a16_vllm_temp_0.2

evalplus.evaluate --dataset humaneval --samples humaneval/neuralmagic--starcoder2-15b-quantized.w8a16_vllm_temp_0.2-sanitized

Accuracy

Benchmark	starcoder2-15b	starcoder2-15b-quantized.w8a16 (this model)	Recovery
HumanEval pass@1	44.8	44.3	98.9%
HumanEval pass@10	62.7	62.6	99.8%
HumanEval+ pass@1	38.6	37.6	97.4%
HumanEval+ pass@10	54.9	54.5	99.3%