llama-3-2-3b-instruct-onnx-genai-int4-kquantlast-emb-int8

Llama-3.2-3B-Instruct optimized for Arm-based Cloud CPU

A quantized version of meta-llama/Llama-3.2-3B-Instruct for text generation, provided in ONNX for the ONNX Runtime GenAI runtime with INT4 weights, dynamic INT8 activations and INT8 embeddings, and optimized for Arm-based Cloud CPU systems.

Summary

This repository contains an Arm-optimized version of meta-llama/Llama-3.2-3B-Instruct for text generation. The model is provided in ONNX for the ONNX Runtime GenAI runtime, targeting Cloud CPU systems.

The weights are quantized with GPTQ to 4-bit asymmetric per-group values (group size 32), activations are quantized to INT8 per-token dynamic, and the token-embedding table and LM head weights are kept in INT8. The composite scheme is labelled W4A8_dyn_emb_int8. The graph is packed for the ONNX Runtime GenAI runtime, which fuses attention into a single GroupQueryAttention operator and wires up the KV cache.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on MMLU and measured performance on a representative evaluation target.

This is the instruction-tuned checkpoint. The bundled chat template is applied to the prompt before tokenization, so the model answers chat-style turns.

Key results

Original model

Model files

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Benchmark workload: 128 prompt tokens, 60 generated tokens.

Performance results

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Accuracy results

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

Install dependencies

Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:

uv python install
uv sync --frozen

Run the example

uv run example.py

Expected input

Expected output

Intended use

This model is intended for developers evaluating text-generation workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• Accuracy was evaluated on MMLU, 5-shot, and may not generalize to all domains.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.

Additional notes

• The ONNX Runtime GenAI artifacts use the fixed names model.onnx (graph) and model.onnx.data (external weights); the runtime requires the .data sidecar to sit alongside model.onnx.
• The exported graph is packed for the ONNX Runtime GenAI runtime: attention is fused into a single GroupQueryAttention operator, the KV cache is wired up by the runtime, and the INT4 weights are packed into MatMulNBits.
• A QuaRot R1/R2 rotation is applied before quantization to make the weight and activation distributions easier to quantize.
• The k_quant_last recipe leaves the token-embedding table and the final LM head projection at INT8 instead of INT4. Those layers were kept at higher precision to keep accuracy within acceptable ranges.
• MMLU was evaluated with lm-evaluation-harness in the cloze (continuation log-likelihood) format, scoring the four answer letters as continuations of the 5-shot prompt, on a fixed subset of the MMLU test split.

About this version

Original Model: meta-llama/Llama-3.2-3B-Instruct by Meta - Repository

Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.

Converted/optimized by: Arm

License: The Original Model and the Optimized Model are subject to Llama 3.2 Community License.

This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.