inference-optimization/granite-4.0-h-tiny-FP8-dynamic

Granite-4.0-h-tiny

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-tiny.

Model Optimizations

This model was obtained by quantizing the weights and activations of ibm-granite/granite-4.0-h-tiny 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

1. Install vLLM from main:

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

2. Initialize vLLM server:

vllm serve RedHatAI/granite-4.0-h-tiny-FP8-dynamic --tensor_parallel_size 1

3. 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-tiny-FP8-dynamic"

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

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-tiny"

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"]

recipe = QuantizationModifier(
    targets=["Linear"],
    scheme="FP8_DYNAMIC",
    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-dynamic"
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 vLLM from main:

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-tiny-FP8-dynamic",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-tiny-FP8-dynamic",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-tiny-FP8-dynamic" \
                  --dataset "humaneval" \
                  --backend vllm \
                  --tp 1 \
                  --greedy

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

Accuracy Comparison

Category	Benchmark	ibm-granite/granite-4.0-h-tiny	RedHatAI/granite-4.0-h-tiny-FP8-dynamic	Recovery (%)
OpenLLM V1	ARC-Challenge (Acc, 25-shot)	62.97	62.37	99.05
	GSM8K (Strict-Match, 5-shot)	80.44	79.83	99.24
	HellaSwag (Acc-Norm, 10-shot)	61.75	61.56	99.69
	MMLU (Acc, 5-shot)	66.46	66.33	99.80
	TruthfulQA (MC2, 0-shot)	58.48	58.11	99.37
	Winogrande (Acc, 5-shot)	71.43	72.30	101.22
	Average	66.92	66.75	99.73
OpenLLM V2	IFEval (Inst Level Strict Acc, 0-shot)	70.62	71.10	100.68
	MMLU-Pro (Acc, 5-shot)	46.24	46.05	99.59
	Average	58.43	58.58	100.13