fix(distributed): stage backend companion assets to remote nodes

[localai-bot]

Problem

In distributed inference, a model whose ModelFile is a single file (e.g. sherpa-onnx VITS/piper, where parameters.model is the .onnx) fails to load on a remote worker. The backend resolves sibling assets from the model directory — tokens.txt, lexicon.txt, the espeak-ng-data / dict directories, Kokoro's voices.bin — but staging only ships the declared ModelFile. The worker then hits:

failed to create sherpa-onnx TTS engine from /models/.../it_IT-paola-medium.onnx

and TTS produces no audio. (countStageableFiles returns 1 for a regular file; whole-directory expansion only happens when ModelFile itself is a directory.)

Fix

Lean on the existing option-path staging (vae_path:-style) rather than hardcoding filenames in the router:

• stageGenericOptions resolves an option value relative to the model's own directory as well as the frontend models dir, so a shared config can declare companions with bare names regardless of whether Model includes a subdirectory (new resolveOptionPath helper).
• It now expands directory-valued options (e.g. espeak-ng-data) file-by-file via stageOptionDir, instead of handing a directory fd to the stager.
• gallery/sherpa-onnx-tts.yaml declares the companion assets as option paths once for all sherpa TTS voices (tokens, lexicon, espeak-ng-data, voices.bin, dict, per-language lexicons). The backend ignores these keys and keeps resolving siblings from the model dir; they exist only so distributed staging ships them. Files a given voice doesn't ship are skipped.

Tests

• New core/services/nodes/router_optionstage_test.go covers staging an option-declared sibling file and expanding a directory option via the model-dir fallback.
• Full core/services/nodes suite passes; go vet clean.

🤖 Generated with Claude Code