Hi,
I've built a working patch for runtime LoRA swapping and want to confirm it's within scope before opening a PR.
What it does: writes W ← W + scale × (B @ A) in-place on an existing StorageView buffer. This allows swapping LoRA adapters at runtime without reloading the model.
Motivation: serving N language-specific adapters on a single base model. Measured on L4 GPU (Whisper-large-v3 float16): cross-adapter swap ~54 ms vs ~1,520 ms for full model reload.
Patch size: 142 lines across 3 files. No existing API changed.
Questions:
Is this within scope for CTranslate2?
Should it be guarded by an experimental build flag?
Where should the Python binding live — currently on WhisperWrapper, but we've verified it works on OmniASR (Encoder/Generator) with the same pattern. Should it be on the base Model instead?
Happy to open a PR if the direction looks good.
Related: #2046
Hi,
I've built a working patch for runtime LoRA swapping and want to confirm it's within scope before opening a PR.
What it does: writes W ← W + scale × (B @ A) in-place on an existing StorageView buffer. This allows swapping LoRA adapters at runtime without reloading the model.
Motivation: serving N language-specific adapters on a single base model. Measured on L4 GPU (Whisper-large-v3 float16): cross-adapter swap ~54 ms vs ~1,520 ms for full model reload.
Patch size: 142 lines across 3 files. No existing API changed.
Questions:
Is this within scope for CTranslate2?
Should it be guarded by an experimental build flag?
Where should the Python binding live — currently on WhisperWrapper, but we've verified it works on OmniASR (Encoder/Generator) with the same pattern. Should it be on the base Model instead?
Happy to open a PR if the direction looks good.
Related: #2046