carpedm20/sapiens2-seg-tensorrt

Sapiens2 TensorRT Engine

Results and visual examples

Hardware-specific Sapiens2 semantic human-part segmentation engines. An engine's runtime metadata and output dtype/geometry are part of its contract. No matching saved visual inference was recovered for the exact listed builds.

Available build records: h100-sm90-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp/README.md, h100-sm90-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp/metadata.json, h100-sm90-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp/sapiens2_0.4b_seg_fp8_mask_uint8_single-engine-native.metadata.json, h100-sm90-trt10.14.1.48-max-batch-1/README.md, h100-sm90-trt10.14.1.48-max-batch-1/metadata.json, h100-sm90-trt10.14.1.48-max-batch-1/sapiens2_0.8b_seg_fp16.metadata.json, rtx5090-sm120-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp/README.md, rtx5090-sm120-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp/metadata.json.

No checkpoint-matched photo gallery was located in the existing evidence reviewed for this update. The files and runtime records below are the available evidence; no visual quality claim is inferred from their presence.

Evidence provenance

This documentation update reuses saved results; it does not rerun inference or change weights. The starting repository revision is ef85bd8048fc. Captions distinguish model predictions, training diagnostics, and aggregate metrics. Qualitative examples are not a representative accuracy estimate.

Gallery sources and SHA-256 checksums.

Pre-built single-engine TensorRT artifact for facebook/sapiens2-seg-0.4b.

Specifications

Setting	Value
Variant	0.4b
Revision	h100-sm90-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp
Precision	fp8
Output kind	mask
Output dtype	uint8
Input shape	[1, 3, 1024, 768]
Max batch size	1
TensorRT	10.14.1.48
CUDA	13.1
Built on	h100-sm90
Engine size	780.51 MB
Model engine	sapiens2_0.4b_seg_fp8_mask_uint8.engine

Usage

Fury downloads this engine when runtime="tensorrt" and SAPIENS2_TENSORRT_ENGINE_REVISION=h100-sm90-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp.

from ai.models.segmentation.sapiens2 import Sapiens2SegmentationModel

model = Sapiens2SegmentationModel(
    model_name="0.4b",
    runtime="tensorrt",
    engine_revision="h100-sm90-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp",
)

Local Download

make cmd-ai-local -- python -m scripts.tensorrt.sapiens2 download \
  --model-name 0.4b \
  --engine-revision h100-sm90-trt10.14.1.48-max-batch-1-fp8-default-normqk-exp

Built at: 2026-07-04 10:08:28 UTC