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

Install specific version:

uv pip install -U git+https://github.com/vllm-project/vllm.git \
  --extra-index-url https://wheels.vllm.ai/nightly \
  --no-deps \
  --no-cache

uv pip install compressed-tensors==0.12.3a20251114 --no-cache
uv pip install --upgrade torchvision --break-system-packages --no-cache
uv pip install cloudpickle msgspec zmq blake3 cachetools prometheus_client fastapi openai openai_harmony pybase64 llguidance diskcache xgrammar lm-format-enforcer partial-json-parser cbor2 einops gguf numba --no-cache

Initialize vLLM server:

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

Send requests to the server:

python

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)

Creation

This model was quantized using the llm-compressor library as shown below.

Creation details

Install specific llm-compression version:

uv pip install git+https://github.com/vllm-project/llm-compressor.git@refs/pull/2001/head --no-cache
uv pip install --upgrade torchvision --break-system-packages --no-cache

python

from transformers import AutoModelForCausalLM, AutoTokenizer

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.utils import dispatch_for_generation
from llmcompressor.modeling import replace_modules_for_calibration
from llmcompressor.modeling.granite4 import pack_3d_experts

MODEL_ID = "ibm-granite/granite-4.0-h-small"

model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

model = replace_modules_for_calibration(model)

ignore_lay = ["lm_head", "re:.*block_sparse_moe.router", "re:.*mamba.in_proj", "re:.*shared_mlp.input_linear"]

recipe = QuantizationModifier(
    targets=["Linear"],
    scheme="FP8_BLOCK",
    ignore=ignore_lay,
)

oneshot(model=model, recipe=recipe)

print("========== SAMPLE GENERATION ==============")
dispatch_for_generation(model)
input_ids = tokenizer(
    "Describe Large Language Model", return_tensors="pt"
).input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=35)
print(tokenizer.decode(output[0]))
print("==========================================")

SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-block"
print(f"Saving to {SAVE_DIR}")

model.save_pretrained(SAVE_DIR)
tokenizer.save_pretrained(SAVE_DIR)
pack_3d_experts(SAVE_DIR)

Evaluation

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

Evaluation details

Install specific version:

uv pip install -U git+https://github.com/vllm-project/vllm.git \
  --extra-index-url https://wheels.vllm.ai/nightly \
  --no-deps \
  --no-cache


uv pip install compressed-tensors==0.12.3a20251114 --no-cache
uv pip install --upgrade torchvision --break-system-packages --no-cache
uv pip install cloudpickle msgspec zmq blake3 cachetools prometheus_client fastapi openai openai_harmony pybase64 llguidance diskcache xgrammar lm-format-enforcer partial-json-parser cbor2 einops gguf numba --no-cache

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

Accuracy Comparison

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	RedHatAI/granite-4.0-h-small-FP8-dynamic
OpenLLM V1	ARC-Challenge (Acc-Norm, 25-shot)	72.27	72.10 (99.76%)	72.27 (100.00%)	72.10 (99.76%)
GSM8K (Strict-Match, 5-shot)	85.22	85.29 (100.09%)	85.52 (100.36%)	84.84 (99.56%)
HellaSwag (Acc-Norm, 10-shot)	86.08	85.88 (99.77%)	85.96 (99.86%)	85.88 (99.77%)
MMLU (Acc, 5-shot)	77.15	77.18 (100.03%)	77.23 (100.09%)	77.18 (100.03%)
TruthfulQA (MC2, 0-shot)	57.64	57.63 (99.99%)	57.94 (100.52%)	57.63 (100.00%)
Winogrande (Acc, 5-shot)	81.37	81.45 (100.10%)	80.82 (99.32%)	81.45 (100.10%)
Average Score	76.62	76.59 (99.96%)	76.62 (100.00%)	76.51 (99.86%)
OpenLLM V2	IFEval (Inst Level Strict Acc, 0-shot)	87.53	87.17 (99.59%)	86.69 (99.04%)	87.41 (99.86%)
BBH (Acc-Norm, 3-shot)	61.52	61.31 (99.66%)	61.40 (99.80%)	61.19 (99.46%)
Math-Hard (Exact-Match, 4-shot)	46.22	43.73 (94.61%)	43.88 (94.93%)	41.77 (90.36%)
GPQA (Acc-Norm, 0-shot)	35.23	34.98 (99.29%)	34.23 (97.14%)	34.23 (97.14%)
MUSR (Acc-Norm, 0-shot)	46.69	46.56 (99.72%)	45.77 (98.02%)	45.77 (98.02%)
MMLU-Pro (Acc, 5-shot)	47.99	47.63 (99.26%)	47.93 (99.88%)	47.58 (99.15%)
Average Score	54.20	53.56 (98.82%)	53.32 (98.38%)	52.99 (97.77%)

ekurtic/granite-4.0-h-small-FP8-block

Granite-4.0-h-small

Model Overview