Arm/clip-vit-base-patch32-int8-xnnpack-executorch

openai/clip-vit-base-patch32 optimized for Arm-based Edge Linux systems

An INT8-quantized version of openai/clip-vit-base-patch32 for zero-shot image classification, exported to ExecuTorch (.pte) and optimized for Arm-based Edge Linux systems.

Summary

This repository contains an Arm-optimized version of openai/clip-vit-base-patch32 for zero-shot image classification. The vision encoder is quantized to dynamic INT8 — per-channel symmetric weights, dynamic INT8 activations computed at inference time — while the text encoder remains FP32, because PT2E could not capture its causal attention mask. The model is provided as a single multi-method ExecuTorch (.pte) file, targeting Edge Linux systems.

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 CIFAR-100 and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	ExecuTorch (.pte), multi-method
Target device class	Edge Linux
Reference device	Raspberry Pi 5 (Cortex-A76, Linux Raspberry Pi OS 64-bit, based on Debian 13 "Trixie")
Primary performance result	p50 latency 95.50 ms (3.12x faster than baseline), 10.47 frames per second
Accuracy result	Top-1 61.56%, Top-5 85.43%
Size / memory result	327.36 MB (1.77x smaller), peak memory 200.94 MB (2.41x less)

Original model

Field	Value
Original model	openai/clip-vit-base-patch32
Original source	Hugging Face
Original developer	OpenAI
Original model card	openai/clip-vit-base-patch32
Original license	MIT

Model files

File	Description
clip_raspberry_executorch_optimized.pte	Arm-optimized model for deployment
example.py	Minimal inference example
pyproject.toml	Pinned runtime dependencies for example.py, resolved with uv
uv.lock	Locked dependency resolution for pyproject.toml
config.yaml	Model I/O contract used by the example
benchmarks/	FP32 baseline and Arm-optimized benchmark records

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

Field	Value
Device / platform	Raspberry Pi 5
CPU / accelerator	Cortex-A76, 4 cores @ 2.4 GHz
OS	Linux, Raspberry Pi OS 64-bit, based on Debian 13 "Trixie"
Runtime	ExecuTorch 1.1.0
Backend / delegate	XNNPACK + KleidiAI
Batch size	1
Precision	vision encoder: dynamic INT8, per-channel symmetric weights, dynamic INT8 activations; text encoder: FP32
Runs	10 warmup + 100 measured

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
Mean Latency	298.20 ms	95.50 ms	3.12x faster
Latency p50	298.20 ms	95.50 ms	3.12x faster
Latency p90	298.36 ms	95.66 ms	3.12x faster
Latency p99	298.55 ms	96.36 ms	3.10x faster
Cold Start	214.35 ms	228.03 ms	+6.4% (larger model graph)
Frames per Second	3.35 FPS	10.47 FPS	3.12x throughput

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

Field	Value
Dataset	CIFAR-100
Split	N/A
Number of samples	10000
Metric(s)	Top-1 accuracy, Top-5 accuracy
Evaluation runtime	ExecuTorch 1.1.0

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
Top-1 accuracy (%)	60.73	61.56	+0.83 pp
Top-5 accuracy (%)	85.25	85.43	+0.18 pp

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:

Optimization area	Applied?	Notes
Model conversion	Yes	Exported to ExecuTorch (.pte) as a single multi-method artifact via PT2E capture, exposing encode_image, encode_text, and forward
Quantization	Yes	Vision encoder quantized to dynamic INT8 via PT2E PTQ-dynamic — per-channel symmetric weights, dynamic INT8 activations computed at inference time; the text encoder was excluded from quantization (kept at FP32) because PT2E could not capture its causal attention mask, preserving export correctness and accuracy
Runtime/backend selection	Yes	XNNPACK + KleidiAI CPU delegate
Graph/runtime compatibility updates	Yes	Performed as part of the ExecuTorch export pipeline
Accuracy validation	Yes	Compared against the original model or published baseline
Performance validation	Yes	Measured on the reference Arm platform

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

Property	Value
Input shape	[1, 3, 224, 224] (encode_image)
Input type	float32
Input range	N/A
Preprocessing	resize shortest side to 224 px (bicubic), center crop to 224x224, convert to float32, normalize with mean [0.48145466, 0.4578275, 0.40821073] and std [0.26862954, 0.26130258, 0.27577711], format NCHW

Expected output

Property	Value
encode_image output shape	[1, 512], float32, L2-normalised image embedding
encode_text output shape	[1, 512], float32, L2-normalised text embedding
forward output shape	[B_img, B_text], float32, scaled cosine similarity logits
Postprocessing	N/A

Intended use

This model is intended for developers evaluating zero-shot image classification 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 CIFAR-100 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 text encoder (CLIPTextModel) remains FP32 because PT2E could not capture it due to incompatibility with the causal attention mask, so it was excluded from quantization to preserve export correctness and accuracy — only the vision encoder is INT8.

The exported .pte is a multi-method artifact exposing three callable methods: encode_image, encode_text, and forward. The text methods support dynamic sequence length, while encode_image and the overall benchmark use a static batch size of 1.

Benchmarks were run single-threaded (n_threads = 1); multi-threaded performance may differ.

• Sample input: sample_input.jpg is derived from Samoyed on Beach by Appleinfl, via Wikimedia Commons (public domain).

About this version

Original Model: openai/clip-vit-base-patch32 by OpenAI - 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 MIT.

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.