openEuler/sam2.1_hiera_tiny_prompt_ascend

SAM 2.1 Hiera Tiny — box-prompt segmentation (Ascend 310P)

Compiled box-prompt SAM 2.1 Hiera Tiny deployment for IB-Robot's SegmentDetections perception service (/perception/*/segment_detections).

Deployments

deployment	backend	artifacts	notes
ascend_310p	Ascend ACL (Ascend310P1)	encoder OM + decoder OM (batch=4 box prompts)	board-side box-prompt masks

Files

• inference_manifest.json — deployment routing (schema v3)
• artifacts/ascend_310p/sam2_encoder.om — vision encoder (longest-side 1024)
• artifacts/ascend_310p/sam2_decoder_bs4.om — mask decoder, 4 box prompts per batch
• assets/sam2.1_hiera_tiny.pt — original PyTorch checkpoint the OMs were converted from (same weights as openEuler/sam2.1_hiera_tiny)
• assets/adapter.json — adapter identity (sam2-mask-logits-iou-v1)

Usage

Select deployment ascend_310p through the unified inference runtime (ibrobot_msgs/srv/SegmentDetections, adapter perception_service.model_service_plugins:SegmentDetectionsPlugin).

For host-side Torch execution (automatic-mask mode), use openEuler/sam2.1_hiera_tiny instead, which ships torch_cpu/torch_cuda deployments.

Selective download:

./scripts/download_models.py --models sam2.1_hiera_tiny_prompt_ascend --target 310p