Embedl-optimized
- timm/vit_huge_patch16_sapiens2.fb
- INT8
- L4
- Labels
- Image
- Build
- Embedl-optimized
- Precision
- INT8
- Runtime
- TensorRT 11.3.0.99 with Embedl optimizations
- Input size
- 1024×768
vit_huge_patch16_sapiens2.fb is an image encoder from timm: it turns an image into a 1,280-dimensional embedding, for search, retrieval and as features for other models.
Methodology
- Latency: trtexec on the NVIDIA L4, batch 1, 201 iterations after warm-up, during the build: GPU compute time per inference, mean and p99, and throughput.
- Quantization: Embedl Deploy, activations INT8, weights INT8 per-channel; calibrated on Imagenette (1,000 images). Exported as ONNX opset 20, compiled with TensorRT 11.3.0.99, built on NVIDIA L4 (sm_89).
- Build: config vit_huge_patch16_sapiens2/fb-timm_int8_embedl_tensorrt_l4, on an NVIDIA L4.