---

name: embedding-engine description: Embedding backends (InsightFace/PyTorch+ONNXRuntime vs TensorRT). Use when optimizing embedding throughput or debugging drift/fallbacks.

Embedding Engine Skill

Use this skill to optimize embedding performance and debug embedding drift/fallback behavior.

When to Use

• Embedding pipeline running slowly
• Need to switch between PyTorch and TensorRT
• Debugging embedding drift between backends
• Building/caching TensorRT engines
• Verifying ONNXRuntime/CoreML provider selection (macOS)

Sub-agents

Current Backends

• pytorch (default): ArcFace via the insightface Python package (used by tools/episode_run.py)
• tensorrt (optional): TensorRT engine build + inference via FEATURES/arcface_tensorrt/

Key Skills

Embed faces with the configured backend

Run embedding with the configured backend (same interface as the pipeline).

Copyfrom tools.episode_run import get_embedding_backend

embedder = get_embedding_backend(
    backend_type="pytorch",  # or "tensorrt"
    device="cpu",
    tensorrt_config="config/pipeline/arcface_tensorrt.yaml",
    allow_cpu_fallback=True,
)
embedder.ensure_ready()
embeddings = embedder.encode(face_crops)  # (N, 512) L2-normalized

Build a TensorRT engine from ONNX

Copypython -m FEATURES.arcface_tensorrt --mode build --onnx-path models/arcface_r100_v1.onnx

Compare TensorRT vs PyTorch embeddings (parity + speedup)

Copypython -m FEATURES.arcface_tensorrt --mode compare --n-samples 100

This uses FEATURES/arcface_tensorrt/src/embedding_compare.py and reports cosine similarity + L2 distance stats.

Config Reference

File: config/pipeline/embedding.yaml

File: config/pipeline/arcface_tensorrt.yaml

Engine Storage

TensorRT engines are GPU-architecture specific. Stored in S3:

Copys3://screenalytics-models/engines/
├── arcface_r100-fp16-sm75.plan   # Ampere (RTX 30xx)
├── arcface_r100-fp16-sm80.plan   # A100
├── arcface_r100-fp16-sm86.plan   # Ada (RTX 40xx)
└── arcface_r100-fp16-sm89.plan   # Hopper (H100)

Naming convention: {model_name}-{precision}-sm{arch}.plan

Common Issues

"Engine not found" / TensorRT backend won’t load

Cause: No engine built for the current GPU / config mismatch

Fix: Build locally:

Copypython -m FEATURES.arcface_tensorrt --mode build --onnx-path models/arcface_r100_v1.onnx

Embedding drift too high

Cause: FP16 quantization or TRT optimization changes

Check: Run parity compare:

Copypython -m FEATURES.arcface_tensorrt --mode compare --n-samples 100

Fix: Use FP32 precision:

Copytensorrt:
  precision: fp32  # default is fp16

TensorRT slower than expected / falling back

Cause: Not batching, engine built with suboptimal shapes/precision, or backend fell back

Check: Ensure config/pipeline/embedding.yaml has embedding.backend: tensorrt and re-run with --mode benchmark.

Fix: Increase batch size, ensure GPU backend:

Copytensorrt:
  opt_batch_size: 32
  max_batch_size: 64

Out of GPU memory

Cause: Engine workspace too large

Check: nvidia-smi during inference

Fix: Reduce workspace:

Copytensorrt:
  workspace_size_mb: 512  # default is 1024

Benchmark Reference

Diagnostic Output

Copy{
  "backend": "tensorrt",
  "engine_path": "~/.cache/screenalytics/engines/arcface_r100_v1-sm86.trt",
  "precision": "fp16",
  "batch_size": 32,
  "embedding_dim": 512,
  "throughput_fps": 245.3,
  "latency_ms": 130.5,
  "vram_mb": 1024,
  "validation": {
    "drift_vs_pytorch": 0.9995,
    "regression_test": "passed"
  }
}

Key Files

Related Skills

• pipeline-insights - General pipeline debugging
• face-alignment - Alignment before embedding