Arm/mobile-sam-int8-xnnpack-executorch

MobileSAM optimized for Arm-based Edge Linux systems

An INT8-quantized version of MobileSAM for box-prompted instance segmentation, exported to ExecuTorch (.pte) and optimized for Arm-based Edge Linux systems.

Summary

This repository contains an Arm-optimized version of MobileSAM for box-prompted instance segmentation, quantized to INT8 via dynamic post-training quantization — per-channel symmetric weights, dynamic per-tensor activations. The model is provided in ExecuTorch (.pte) format, 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 COCO 2017 val and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	ExecuTorch (.pte)
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 1892.892 ms (1.08x faster than baseline)
Accuracy result	mIoU 73.11%
Size / memory result	21.744 MB (1.96x smaller), peak memory 320.83 MB

Original model

Field	Value
Original model	MobileSAM
Original source	GitHub
Original developer	MobileSAM (Chaoning Zhang et al., KAIST)
Original model card	MobileSAM repository
Original checkpoint	weights/mobile_sam.pt — the TinyViT-5M checkpoint this model was quantized from
Original license	Apache-2.0

Model files

File	Description
mobile_sam_raspberry_executorch_optimized.pte	Arm-optimized INT8 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 input/output 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 architecture	arm64
CPU	Cortex-A76, 4 cores @ 2.4 GHz
System memory	8 GB
OS	Linux, Raspberry Pi OS 64-bit, based on Debian 13 "Trixie"
Runtime	ExecuTorch 1.1.0
Execution backend	CPU (XNNPACK, KleidiAI)
Precision	INT8, dynamic PTQ — per-channel symmetric weights, dynamic INT8 activations
Batch size	1
Input resolution	1024x1024
Runs / warmup	100 / 10

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
Model size (MB)	42.685	21.744	1.96x smaller
End-to-end latency p50 (ms)	2053.196	1892.892	1.08x faster
End-to-end latency p90 (ms)	2064.113	1898.845	1.09x faster
Model load time (ms)	34.061	42.683	1.25x slower
Time to first inference (ms)	2110.23	1957.949	1.08x faster
Peak memory (MB)	334.23	320.83	1.04x less
Average memory (MB)	332.17	318.53	1.04x less
Frames per second	0.49	0.53	1.08x
XNNPACK delegation (%)	—	80.11	—

Model load time is the one metric that regresses. Dynamic quantization stores per-channel scales alongside the weights, so the optimized program has more tensors to map at load time even though it is half the size. It is paid once per process and is repaid by the first inference.

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	COCO 2017
Split	val2017
Sample count	5000
Metric(s)	mIoU (box-prompted instance segmentation)
Runtime	ExecuTorch 1.1.0 (CPU, XNNPACK, KleidiAI)

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
mIoU (%)	73.09	73.11	+0.02 pp

Each ground-truth box is used as a prompt and the predicted mask is compared with the corresponding ground-truth mask. 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	Converted to ExecuTorch .pte
Quantization	Yes	INT8 PTQ-dynamic — TorchAO 8da8w on Linear modules, PT2E dynamic INT8 on Conv2d / ConvTranspose2d; no calibration data required
Runtime/backend selection	Yes	XNNPACK + KleidiAI delegate
Graph/runtime compatibility updates	Yes	27 Conv2d_BN modules fused for PT2E pattern matching; LayerNorm2d patched to keep rank-4 tensors so XNNPACK can form large partitions
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

Download mobile_sam_raspberry_executorch_optimized.pte and place it beside example.py.

Run the example

uv run example.py

The script reads sample_input.jpg, prompts the model with a box around the vehicle in the center of the road, runs inference with mobile_sam_raspberry_executorch_optimized.pte, prints a finite summary, and writes:

• sample_output.png — the input image at its native resolution with the predicted mask overlaid in green and the box prompt drawn in orange
• segmentation.json — the box prompt, the per-proposal IoU scores, and the mask pixel count and coverage

Both files ship with this repository as expected results, and the run overwrites them in place. Copy the directory first if you want to keep them for comparison.

Expected input

Property	Value
Shape	Image [1, 3, 1024, 1024], box prompt [1, 1, 2, 2]
Dtype	float32
Range	[0.0, 1.0]; ImageNet normalization is applied inside the exported graph, so callers must not normalize externally
Preprocessing	to tensor and divide by 255, then resize to 1024x1024 with F.interpolate(mode="bilinear", align_corners=False) — no aspect-ratio preservation
Box prompt	Two corner points [[[x_tl, y_tl], [x_br, y_br]]] in 1024-pixel space, scaled per axis from native pixels and clamped to [0, 1024]

Resizing runs on the [0, 1] tensor rather than the PIL image. torchvision.transforms.Resize antialiases on downscale and produces different pixels, which would shift the mask away from the results above.

Expected output

Property	Value
Shape	low_res_masks [1, 3, 256, 256], iou_predictions [1, 3]
Format	Three candidate mask logits plus a per-proposal IoU quality score
Postprocessing	argmax over iou_predictions, bilinear upsample to 1024x1024, threshold at 0, then resize to the image's native resolution with mode="nearest"

Intended use

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

Limitations

• Evaluated on 5000 images from COCO 2017 val using box-prompted segmentation
• Fixed input size of 1024x1024 — images are resized without preserving the aspect ratio, which may affect accuracy on extreme aspect ratios
• Requires a box prompt around the object of interest; a loose or off-target box segments whatever dominates that box rather than the intended object
• Performance measurements are from Raspberry Pi 5 hardware; results may differ on other Arm devices

Additional notes

Quantization was performed with TorchAO Int8DynamicActivationIntxWeightConfig (8da8w, HQQ scale-only) on the Linear modules and a PT2E dynamic INT8 pass on the Conv2d / ConvTranspose2d operators. Dynamic quantization needs no calibration data: activation scales are computed at inference time from the actual input tensor.

Selecting the best proposal, upsampling, and thresholding run in Python rather than inside the exported graph, so PT2E canonicalization of argmax, gather, and bilinear cannot affect the result.

• Sample input: sample_input.jpg is derived from Richmond Skyline by JuxtaposedJacob, via Wikimedia Commons (CC0 1.0).

About this version

Original Model: MobileSAM by MobileSAM (Chaoning Zhang et al., KAIST) - 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 Apache-2.0.

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.