Ryanontheinside/feat/independent lora merge strategies

[ryanontheinside]

Depends upon #171 .

Per-LoRA Merge Strategy Support

Enables independent merge strategy selection for each LoRA adapter instead of forcing a global strategy.

What Changed

• Frontend: Added per-LoRA strategy dropdown in LoRA Manager
• Backend: LoRAConfig.merge_mode optional field with fallback to global lora_merge_mode
• Manager: Groups LoRAs by strategy, loads in order (permanent_merge → runtime_peft), routes runtime updates to appropriate handler

Benefits

• Mix permanent_merge (max FPS) and runtime_peft (live adjustments) LoRAs simultaneously
• Fine-grained control per adapter without pipeline-wide tradeoffs
• Backward compatible: defaults to global strategy if not specified

Technical Details

• LoRAManager refactored: consolidated mode resolution, generic grouping helper
• Frontend UI consistency: both Strategy and Scale labels use LabelWithTooltip
• Timeline export/import automatically preserves per-LoRA mergeMode
• Scale slider tooltip dynamically reflects effective merge mode

Example Use Case

Some LoRAs are desirable for performance but do not require scale updates. Users that want to usually/always include these with maximum performance, but still want to runtime updates of other LoRA will need this feature.

Example performance LoRA for:

• 1.3b
• 14b

Next Steps

A more explicit/elegant pattern for merge order can be revisited when moving scale updates to modular block is considered in place of Mixin.

[yondonfu]

LGTM!

[yondonfu]

Squashing to avoid dealing with rebase conflicts