This document describes how to run each KernelPort deployment (model + backend + optional worker).
| Name | Manifest | Backend | Worker image | How to run |
|---|---|---|---|---|
| identity-onnx | models/sample-manifest.yaml | onnx | — | MODEL_MANIFEST_PATH=... model_fetch then serve --backend onnx --model-path /models/<id>/model.onnx |
| helion-demo | — | helion | kernelport-helion:gpu | docker compose up (default) |
| luxtts | models/luxtts-manifest.yaml | helion | kernelport-luxtts:gpu | docker compose -f docker-compose.yml -f docker-compose.luxtts.yml up or Lambda deploy |
The authoritative mapping is in deployments/deployments.yaml.
- Set
HUGGINGFACE_HUB_TOKEN(orHF_TOKEN) so the LuxTTS worker can pull the model from Hugging Face. - Build and run with the LuxTTS override:
docker compose -f docker-compose.yml -f docker-compose.luxtts.yml up --build
- Inference endpoint: gRPC at
localhost:50051, methodkernelport.v1.InferenceService/Infer. See LuxTTS tensor contract for input/output tensors.
See Lambda validation for the automated provision → inference validation → benchmark → cleanup workflow, required secrets, artifacts, and manual teardown instructions. The default workload is Helion softmax; LuxTTS is a separate selectable contract test.
Inputs (all optional except text and prompt_audio; omitted use LuxTTS defaults):
| Name | Dtype | Shape | Description |
|---|---|---|---|
| text | U8 | [N] | UTF-8 bytes of the text to synthesize. |
| prompt_audio | U8 | [M] | Raw reference audio file bytes (WAV or MP3). |
| rms | F32 | [1] | Loudness (default 0.01). |
| t_shift | F32 | [1] | Sampling param (default 0.9). |
| num_steps | I32 | [1] | Sampling steps (default 4). |
| speed | F32 | [1] | Playback speed (default 1.0). |
| return_smooth | I32 | [1] | 0 = false, 1 = true (default 0). |
| ref_duration | I32 | [1] | Reference duration in seconds (default 5). |
Outputs:
| Name | Dtype | Shape | Description |
|---|---|---|---|
| audio | F32 | [S] | Mono waveform at 48 kHz. |
| sample_rate | I32 | [1] | 48000. |