feat: LoRA adapter support for dit-vae

[Copilot]

Adds runtime LoRA adapter loading to dit-vae via PEFT-format adapter_model.safetensors, with deltas applied in-graph across all DiT self-attention, cross-attention, and MLP projections.

New files

• src/safetensors.h — zero-dependency safetensors reader (header length + JSON metadata + raw tensor reads by offset)
• src/dit-lora.cpp — dit_ggml_load_lora(): normalises all PEFT key-name variants, matches A/B pairs per slot, transposes PyTorch row-major → ggml column-major, uploads to backend

Core changes

• src/dit.h — DiTGGMLLayer gains 22 LoRA tensor pointer slots (SA/CA Q/K/V/O + MLP gate/up/down); DiTGGML gains lora_wctx (independent weight context) and lora_scale
• src/dit-graph.h — new dit_ggml_linear_lora() (W@x + scale·B@A@x); self-attn, cross-attn, and MLP builders apply deltas on both fused and separate weight paths
• src/request.h/request.cpp — custom_tag (trigger word auto-appended to caption before condition encoding) and genre fields; parsing aliases for formatted_lyrics, language, is_instrumental
• tools/dit-vae.cpp — --lora <path> and --lora-scale <float> (default 1.0); adapter loaded once, reused across all requests

Usage

./build/dit-vae \
    --request request0.json \
    --text-encoder models/Qwen3-Embedding-0.6B-Q8_0.gguf \
    --dit   models/acestep-v15-turbo-Q8_0.gguf \
    --vae   models/vae-BF16.gguf \
    --lora  lora/adapter_model.safetensors \
    --lora-scale 1.0

{
  "caption": "Nu-disco track with funky bassline",
  "custom_tag": "crydamoure",
  "inference_steps": 8,
  "shift": 3
}

custom_tag is appended to the caption so the trigger word reaches the condition encoder. See examples/lora.sh and examples/lora.json for a full runnable example.

Deltas are applied in the compute graph (base weights are not mutated). When lora_scale == 0 all LoRA paths are no-ops. All base weight fusion variants (full QKV, partial QK, separate) are handled.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.