HarmenWessels's picture
Card: remove benchmark info; GitHub links -> repo root
d84b1d1 verified
|
Raw
History Blame Contribute Delete
1.94 kB
metadata
license: apache-2.0
base_model: ibm-granite/granite-4.1-8b
base_model_relation: quantized
tags:
  - openvino
  - int4
  - awq
  - intel
  - arc
  - igpu
  - core-ultra
language:
  - en
pipeline_tag: text-generation

granite-4.1-8b — OpenVINO INT4 (channel-wise, symmetric, AWQ-calibrated)

OpenVINO IR conversion of ibm-granite/granite-4.1-8b (dense 40-layer transformer, 128k context, Apache-2.0). To our knowledge the first OpenVINO IR of this model. Quantized INT4 symmetric channel-wise with AWQ + scale-estimation calibration (data-aware calibration, applied from the start).

Converted with optimum-cli (optimum-intel git-master, transformers 4.57.6):

optimum-cli export openvino -m ibm-granite/granite-4.1-8b \
  --task text-generation-with-past --weight-format int4 --sym --group-size -1 \
  --awq --scale-estimation --dataset wikitext2 \
  granite-4.1-8b-int4-cw-ov

Related

Related artifacts: granite-4.1-3b-int4-cw-ov (same recipe, smaller/faster) and granite-4.1-3b-int8-ov.

Usage (OpenVINO GenAI)

import openvino_genai as ov_genai

pipe = ov_genai.LLMPipeline("granite-4.1-8b-int4-cw-ov", "GPU", CACHE_DIR="./.ovcache")
print(pipe.generate("Write a Python function that merges overlapping intervals.",
                    max_new_tokens=256))

Tested end-to-end as an OpenAI-compatible Continue.dev backend via core-ultra-llm-server.

Provenance

  • Base model: ibm-granite/granite-4.1-8b (IBM, Apache-2.0, released 2026-04)
  • Calibration: AWQ + scale estimation on wikitext2 (NNCF defaults)
  • Conversion date: 2026-06-05
  • No finetuning — weights are a direct quantization of the original