Resolve merge conflicts: resync with upstream while preserving fork features

[Copilot]

PR #5 (ServeurpersoCom → audiohacking) had dirty merge state because the fork had diverged with its own additions (LoRA, cover/repaint mode, VAE encoder, reference audio). Blind upstream adoption would have severed those features. Each of the 16 conflicts was resolved by merging both sides' intent.

Brought in from upstream

• Flash attention toggle — use_flash_attn: bool added to CondGGML, DetokGGML, Qwen3GGML, Qwen3LM; qwen3_attn_f32() pure-F32 fallback added to qwen3-enc.h; threaded through all qwen3_build_layer / qw3lm_build_attn call chains
• --no-fa CLI flag — ace-qwen3 and dit-vae both accept --no-fa to disable flash attention at runtime
• qwen3_embed_lookup() via ggml_get_rows — replaces the old CPU mmap dequant approach (qwen3_cpu_embed_lookup); qw3lm_forward / qw3lm_forward_batch switch to ggml_backend_sched_alloc_graph instead of ggml_gallocr; galloc, gf_mmap, embed_mmap_data, embed_type fields removed from Qwen3LM
• ggml submodule pointer advanced to upstream (55e062ab)
• tests/debug-dit-cossim.sh rewritten to build and test CUDA / Vulkan / CPU in sequence
• README documentation updates

Preserved from fork

• LoRA — DiTGGMLLayer adapter tensors, dit_ggml_linear_lora(), dit_ggml_load_lora(), lora_wctx / lora_scale in DiTGGML, LoRA CLI args in dit-vae, dit-lora.cpp in CMakeLists
• Cover / repaint mode — task_type, reference_audio, src_audio, audio_cover_strength, repainting_start/end in AceRequest; custom_tag / genre for LoRA trigger words
• VAE Encoder — VAEEncoderGGML + vae_encoder_load() for reference audio timbre encoding
• FSQ codeword helpers — detok_ggml_build_codeword_table() + latent_frames_to_codes() kept in fsq-detok.h (used by file-based cover mode)
• .gitignore — !tests/fixtures/ exception preserved so fixture JSON files remain tracked
• CMakeLists — audio_loader.cpp and src/dit-lora.cpp remain in the build

Example: flash attention now toggleable per model

dit_ggml_init_backend(&model);
if (!use_fa) model.use_flash_attn = false;   // DiT

qwen3_init_backend(&text_enc);
if (!use_fa) text_enc.use_flash_attn = false; // text encoder

cond_ggml_init_backend(&cond);
if (!use_fa) cond.use_flash_attn = false;     // condition encoder

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.