granite-3.1-2b-base-quantized.w8a8

Model Overview

• Model Architecture: granite-3.1-2b-base
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT8
  • Activation quantization: INT8
• Release Date: 1/8/2025
• Version: 1.0
• Model Developers: Neural Magic

Quantized version of ibm-granite/granite-3.1-2b-base. It achieves an average score of 57.22 on the OpenLLM benchmark (version 1), whereas the unquantized model achieves 57.65.

Model Optimizations

This model was obtained by quantizing the weights and activations of ibm-granite/granite-3.1-2b-base to INT8 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.

Deployment

Use with vLLM

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

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

max_model_len, tp_size = 4096, 1
model_name = "neuralmagic/granite-3.1-2b-base-quantized.w8a8"
tokenizer = AutoTokenizer.from_pretrained(model_name)
llm = LLM(model=model_name, tensor_parallel_size=tp_size, max_model_len=max_model_len, trust_remote_code=True)
sampling_params = SamplingParams(temperature=0.3, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id])

messages_list = [
 [{"role": "user", "content": "Who are you? Please respond in pirate speak!"}],
]

prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list]

outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)

generated_text = [output.outputs[0].text for output in outputs]
print(generated_text)

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

Creation

This model was created with llm-compressor by running the code snippet below.

Evaluation

The model was evaluated on OpenLLM Leaderboard V1, OpenLLM Leaderboard V2 and on HumanEval, using the following commands:

Accuracy

Category	Metric	ibm-granite/granite-3.1-2b-base	neuralmagic/granite-3.1-2b-base-quantized.w8a8	Recovery (%)
OpenLLM V1	ARC-Challenge (Acc-Norm, 25-shot)	53.75	54.01	100.48
	GSM8K (Strict-Match, 5-shot)	47.84	46.55	97.30
	HellaSwag (Acc-Norm, 10-shot)	77.94	77.94	100.00
	MMLU (Acc, 5-shot)	52.88	52.34	98.98
	TruthfulQA (MC2, 0-shot)	39.04	38.12	97.64
	Winogrande (Acc, 5-shot)	74.43	74.35	99.89
	Average Score	57.65	57.22	99.26
Coding	HumanEval Pass@1	30.00	29.60	98.67

Inference Performance

This model achieves up to 1.4x speedup in single-stream deployment and up to 1.1x speedup in multi-stream asynchronous deployment, depending on hardware and use-case scenario. The following performance benchmarks were conducted with vLLM version 0.6.6.post1, and GuideLLM.

Single-stream performance (measured with vLLM version 0.6.6.post1)

Latency (s)
GPU class	Model	Speedup	Code Completion prefill: 256 tokens decode: 1024 tokens	Docstring Generation prefill: 768 tokens decode: 128 tokens	Code Fixing prefill: 1024 tokens decode: 1024 tokens	RAG prefill: 1024 tokens decode: 128 tokens	Instruction Following prefill: 256 tokens decode: 128 tokens	Multi-turn Chat prefill: 512 tokens decode: 256 tokens	Large Summarization prefill: 4096 tokens decode: 512 tokens
A5000	granite-3.1-2b-base	10.9	1.4	11.0	1.5	1.4	2.8	6.1
	granite-3.1-2b-base-quantized.w8a8 (this model)	1.37	7.9	1.0	8.0	1.1	1.0	2.0	4.7
	granite-3.1-2b-base-quantized.w4a16	1.94	5.4	0.7	5.5	0.8	0.7	1.4	3.4
A6000	granite-3.1-2b-base	9.8	1.3	10.0	1.3	1.3	2.6	5.4
	granite-3.1-2b-base-quantized.w8a8 (this model)	1.31	7.8	1.0	7.6	1.0	0.9	1.9	4.5
	granite-3.1-2b-base-quantized.w4a16	1.87	5.1	0.7	5.2	0.7	0.7	1.3	3.1

Multi-stream asynchronous performance (measured with vLLM version 0.6.6.post1)

Maximum Throughput (Queries per Second)
GPU class	Model	Speedup	Code Completion prefill: 256 tokens decode: 1024 tokens	Docstring Generation prefill: 768 tokens decode: 128 tokens	Code Fixing prefill: 1024 tokens decode: 1024 tokens	RAG prefill: 1024 tokens decode: 128 tokens	Instruction Following prefill: 256 tokens decode: 128 tokens	Multi-turn Chat prefill: 512 tokens decode: 256 tokens	Large Summarization prefill: 4096 tokens decode: 512 tokens
A5000	granite-3.1-2b-base	2.9	10.2	1.8	8.2	19.3	9.1	1.3
	granite-3.1-2b-base-quantized.w8a8 (this model)	1.13	3.1	12.1	2.0	9.6	22.2	10.2	1.4
	granite-3.1-2b-base-quantized.w4a16	0.98	2.8	10.0	1.8	8.1	18.6	9.0	1.2
A6000	granite-3.1-2b-base	3.7	12.4	2.4	10.3	23.6	11.0	1.6
	granite-3.1-2b-base-quantized.w8a8 (this model)	1.12	3.6	14.4	2.7	12.0	28.3	12.9	1.7
	granite-3.1-2b-base-quantized.w4a16	0.95	3.7	11.4	2.5	9.8	22.1	10.4	1.4

Base model

ibm-granite/granite-3.1-2b-base

(9)

Quantized Granite models from IBM Research. • 18 items • Updated Apr 30 • 2