Granite-4.0-h-small

Model Overview

Model Architecture: GraniteMoeHybridForCausalLM
- Input: Text
- Output: Text
Model Optimizations:
- Weight quantization: FP8
- Activation quantization: FP8
Release Date:
Version: 1.0
Model Developers:: Red Hat

Quantized version of ibm-granite/granite-4.0-h-small.

Model Optimizations

This model was obtained by quantizing the weights and activations of ibm-granite/granite-4.0-h-small to FP8 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 and activations of the linear operators within transformers blocks of the language model are quantized.

Deployment

Use with vLLM

Initialize vLLM server:

vllm serve RedHatAI/granite-4.0-h-small-FP8-block --tensor_parallel_size 1

Send requests to the server:

from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

model = "RedHatAI/granite-4.0-h-small-FP8-block"

messages = [
    {"role": "user", "content": "Explain quantum mechanics clearly and concisely."},
]


outputs = client.chat.completions.create(
    model=model,
    messages=messages,
)

generated_text = outputs.choices[0].message.content
print(generated_text)

Evaluation

The model was evaluated on the OpenLLM leaderboard task, using lm-evaluation-harness. vLLM was used for all evaluations.

Evaluation details

Openllm V1

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/granite-4.0-h-small-FP8-block",dtype=auto,add_bos_token=True,max_model_len=16384,tensor_parallel_size=1,gpu_memory_utilization=0.9,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks openllm \
  --write_out \
  --batch_size auto \
  --show_config

Openllm V2

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/granite-4.0-h-small-FP8-block",dtype=auto,add_bos_token=False,max_model_len=16384,tensor_parallel_size=1,gpu_memory_utilization=0.7,disable_log_stats=True,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks leaderboard \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --write_out \
  --batch_size auto \
  --show_config

Coding Benchmarks

evalplus.evaluate --model "RedHatAI/granite-4.0-h-small-FP8-block" \
                  --dataset "humaneval" \
                  --backend vllm \
                  --tp 1 \
                  --greedy

evalplus.evaluate --model "RedHatAI/granite-4.0-h-small-FP8-block" \
                --dataset "mbpp" \
                --backend vllm \
                --tp 1 \
                --greedy

* I/p Length = 2048, O/p Length = 2048, #Requests = 1024

Accuracy

Category	Metric	ibm-granite/granite-4.0-h-small	ibm-granite/granite-4.0-h-small-FP8	RedHatAI/granite-4.0-h-small-FP8-block
Model Size (GB)		64.41	33.48	36.43
Throughput (Requests/sec)*		2.031	2.144	2.066
OpenLLM V1	ARC-Challenge (Acc-Norm, 25-shot)	72.27	71.67 (99.17)	72.27 (100.00)
	GSM8K (Strict-Match, 5-shot)	85.06	85.60 (100.62)	85.60 (100.62)
	HellaSwag (Acc-Norm, 10-shot)	86.07	86.02 (99.94)	85.96 (99.87)
	MMLU (Acc, 5-shot)	77.15	76.94 (99.73)	77.23 (100.10)
	TruthfulQA (MC2, 0-shot)	57.97	57.62 (99.40)	57.85 (99.80)
	Winogrande (Acc, 5-shot)	81.45	81.14 (99.61)	80.82 (99.22)
	Average Score	76.66	76.50 (99.79)	76.62 (99.95)
OpenLLM V2	IFEval (Inst Level Strict Acc, 0-shot)	87.41	87.65 (100.27)	87.89 (100.55)
	BBH (Acc-Norm, 3-shot)	61.52	61.31 (99.66)	61.40 (99.80)
	Math-Hard (Exact-Match, 4-shot)	46.60	44.34 (95.14)	44.94 (96.43)
	GPQA (Acc-Norm, 0-shot)	32.55	32.05 (98.45)	34.23 (105.15)
	MUSR (Acc-Norm, 0-shot)	46.43	46.30 (99.72)	45.77 (98.58)
	MMLU-Pro (Acc, 5-shot)	47.96	47.91 (99.88)	47.93 (99.93)
	Average Score	53.75	53.26 (99.09)	53.69 (99.89)