timm/edgenext_small.usi_in1k

int8 (INT8) · embedl_tensorrt_l4

timm/edgenext_small.usi_in1k
INT8
L4

Labels

Image
Classification

Build

Embedl-optimized

Precision

INT8

Runtime

TensorRT 11.3.0.99 with Embedl optimizations

Input size

256×256

Run edgenext_small.usi_in1k on L4

edgenext_small.usi_in1k is an image classifier from timm. It labels an image with one of the 1,000 ImageNet classes.

License: mit (the upstream model's).

Methodology

Latency: trtexec on the NVIDIA L4 with TensorRT 11.3.0.99: batch 1, GPU compute time (no copies to and from the GPU).
Accuracy: the ImageNet validation set, 50,000 images, on this INT8 engine.

Files

Artifacts

What ships in the model bundle.

Bundle

docker.io/embedl/model-artifacts:edgenext_small.usi_in1k-timm_int8_embedl_tensorrt_l4.054163824bc1cf1e

File	Kind	Size	sha256
Benchmark results benchmark_results.json	Results	1 KiB	c373660fb24b
Engine layers (trtexec --exportLayerInfo) compile/layer_info.json	Layer info	8 KiB	cdf936f8b88b
TensorRT engine compile/model.engine INT8, NVIDIA L4 (sm_89), TensorRT 11.3.0.99	Engine	10.6 MiB	b449dc1313b4
Per-layer timings (trtexec --exportProfile) compile/profile.json	Profile	27 KiB	636714cb9328
trtexec benchmark log compile/trtexec_benchmark.log	Log	9 KiB	77859102dcaf
Source run: trtexec benchmark log compile/trtexec_benchmark_source.log	Log	9 KiB	4c3ba70c33e7
trtexec profiling log compile/trtexec_profile.log	Log	9 KiB	fb299b83b25b
Per-inference timings (trtexec --exportTimes) compile/trtexec_times.json	Results	53 KiB	9501761a42a3
Source run: per-inference timings compile/trtexec_times_source.json	Results	53 KiB	058969b32417
Build config config.json	Config	2 KiB	054163824bc1
Example input (0.jpg) examples/0.jpg	example	94 KiB	899c0c9a0fbf
Example input (1.jpg) examples/1.jpg	example	74 KiB	f275d4150279
Example input (2.jpg) examples/2.jpg	example	100 KiB	848fc2aaa45e
Model config (Hugging Face format) hf/config.json	Config	68 KiB	ffdf8ee9845b
Preprocessing (Hugging Face format) hf/preprocessor_config.json	Config	363 B	2acb4ab927f1
Source run: host phase log (export, quantization) host.log	Log	2 KiB	763b21549fa8
Runtime manifest (embedl.model/1) model.json	Manifest	1 KiB	af499efc8f76
Optimized PyTorch export optimization/model.pt2	PT2	21.6 MiB	995b9f21bc36
Quantized ONNX (Q/DQ) quantization/model.onnx	ONNX	21.8 MiB	d855e1c45ff5
Quantized PyTorch export quantization/model.pt2	PT2	21.8 MiB	5b6f128295ba
Self-test: embedl benchmark selftest/benchmark.json	Results	547 B	bf6b6dd014b6
Self-test: embedl infer selftest/infer.json	Results	726 B	b560fe511e62
Self-test: embedl info selftest/info.json	Results	902 B	688db868574a
Self-test: embedl serve selftest/serve.json	Results	726 B	b560fe511e62
Source build (where the engine and its accuracy come from) source-build.json	Manifest	3 KiB	b7bb4b0f1c1e
Source run: target phase log (compile, benchmark, validation) target.log	Log	4 KiB	131391dab42a
Validation results validation_results.json	Results	109 B	409dd85cd439