rcli-device-e2e

# RCLI device modality e2e Do not write per-engine tests. The harness is `scripts/e2e-modalities.sh`, called from `scripts/e2e.sh`. Keys are **primitives** (`llm`, `stt`, `tts`, `vlm`, `embed`, `image`, `vad`, `rerank`, `segment`). rcli picks the engine from catalog framework, local path, or plugin priority. `--engine` is an override (`RCLI_E2E_ENGINE`), never a required test input. ## Run ```bash # Public CI (modelless): skip every modality bash scripts/e2e.sh /path/to/rcli # Device: discover whatever is already on disk, then run each primitive export RUNANYWHERE_HOME=/path/to/home # already-pulled OSS models export RCLI_E2E_MODEL_ROOTS=/path/to/hnpu:/path/to/coreml bash scripts/e2e-modalities.sh /path/to/rcli # Or pin one primitive (path or catalog id) RCLI_E2E_LLM=/path/to/lfm2_5_230m_HNPU \ RCLI_E2E_STT=/path/to/whisper_base_HNPU \ RCLI_E2E_TTS=/path/to/kitten_micro_0_8_HNPU \ RCLI_E2E_EMBED=/path/to/embeddinggemma_300m_HNPU \ bash scripts/e2e-modalities.sh /path/to/rcli ``` `RCLI_E2E_AUTO=1` pulls small OSS catalog defaults the **registered** backends can run (`smollm2`, `whisper-tiny`, `piper`, `minilm`, `silero`, `mlx-qwen3`, …). Never enable AUTO in public CI. ## Why a device "only has LLM" QHexRT and NeuRT implement STT/TTS/VLM/embed/diffusion. The Windows ARM64 box often only has LFM `*_HNPU` trees under `Downloads\hnpu` because those were copied for LLM smoke — not because the engine is LLM-only. Catalog ids: | Modality | QHexRT id (local `*_HNPU`) | NeuRT id (local Core ML tree) | |---|---|---| | LLM | `lfm2_5_230m` | `lfm2_5_230m_ane` | | STT | `whisper_base`, `moonshine_tiny` | `parakeet_tdt_0_6b_v2_ane` | | TTS | `kitten_micro_0_8` | — | | Embed | `embeddinggemma_300m` | — | | VLM | `internvl3_5_1b` (~10 GB) | — | | Image | `cosmos3_edge_diffusion` | `sd15` | `rcli pull` of a Hugging Face **repo page** is HTML. Pass the expanded directory to `-m`. Download `v81/*` only on Hexagon v81. Skip with a clear "no bundle" when the tree is missing. Fail only when a model was selected and the command failed. ## Overlay gotchas - Public bottles never list `neurt` / `qhexrt`. Rebuild product `rcli` against an overlay kit (`RCLI_SDK_KIT` pointing at that prefix). - **QHexRT:** QAIRT **2.48** on Snapdragon X2 Elite / Hexagon v81. `ADSP_LIBRARY_PATH` must be the fully expanded `...\lib\hexagon-v81\unsigned` path. Nested `%QNN_SDK_ROOT%` in `cmd /c set` does not expand. Copy `QnnHtp*.dll` next to `rcli.exe`. FastRPC ~90s then user-driver fallback is normal. Use a `.bat`, not nested `cmd /c`. - **NeuRT image:** `--prompt` and `--out` required; `--steps 4` for smoke. Compiled zip, not the HF repo HTML. Tree needs `TextEncoder.mlmodelc` / `Unet.mlmodelc` / `VAEDecoder.mlmodelc`. - **llama.cpp VLM:** do **not** add a literal `<image>` in the prompt — the SDK inserts `mtmd_default_marker()`. An extra `<image>` makes `mtmd_tokenize` see 0 media markers. A tiny PNG can `bad_alloc` in SmolVLM2 after the 512×512 warmup; skip or pass a real photo via `RCLI_E2E_VLM`. - **segment:** binary P6 PPM, not PNG. - STT has no `--engine` flag; put `-m` before the wav. See `rcli-e2e` for bottle/backends assertions and Apple MLX host link flags.