fix: Ensure data directory exists in LakhMIDIService

fix: Add missing backend/services/msd_importer.py

fix: Add missing backend/services/genre_profiler.py

fix: Add missing backend/services/msd_database_service.py

fix: Add missing backend/services/__init__.py

Phase 2: Update theory_constraint_service.py

Phase 2: Update lakh_midi_service.py

Fix crossfade to use proper equal-power overlap instead of fade-out-then-in

Add embedded metadata extraction from audio files

Enhanced error handling for ZeroGPU quota errors

Support T4 GPU and CPU - make ZeroGPU decorator conditional

Fix tensor serialization error - disable gradients for ZeroGPU

Fix ZeroGPU timeout - pre-initialize model on CPU and increase GPU duration

Add LoRA support to music generation - load from HF and local sources

Fix mastering presets syntax error and add preset description updates

Force upload clean hf_storage_service.py - fix syntax error

Delete corrupted file

Update backend/services/hf_storage_service.py for dataset repo ZIP uploads

Fix syntax error - rebuild file

Update backend/services/hf_storage_service.py for dataset repo ZIP uploads

Fix HuggingFace authentication for LoRA uploads - Add HF_TOKEN support from environment variables - Pass token to all HF API calls (create_repo, upload_folder, snapshot_download) - Better error messages when not authenticated - Return detailed upload results with URLs - Guide users to log in or duplicate Space for uploads

Fix missing shutil import in LoRA training service - Adds shutil and zipfile imports for LoRA import/export operations - Fixes NameError when uploading trained LoRAs

Add dataset import & LoRA collection sync - Adds import_prepared_dataset() method to support ZIP dataset imports with name conflict resolution - Adds sync_on_startup() to download LoRAs from HF collection on app startup - Enhanced upload_lora() with training_config for proper metadata - Implements numeric suffix naming for conflicts (_1, _2, etc.)

Fix LoRA import - Handle ZIP files with metadata at root level

Fix two errors in LoRA training and dataset preparation

Fix ZeroGPU compatibility - Dynamic device allocation

Switch to HF Collection model - Upload LoRAs as individual models

Fix dataset preparation bug and HF repo authentication

LEMM v1.0.0 - Rename and add HF repo storage persistence

Add ZeroGPU decorator to DiffRhythm2 generation for HF Spaces compatibility

Fix dataset refresh, add LoRA list function, improve empty dataset error handling

Add persistent dataset detection for HF and user datasets across sessions

Remove MusicCaps dataset (metadata only, no audio)

Remove MTG-Jamendo dataset (not found on Hub)

Fix dataset IDs: remove Common Voice & FLEURS, fix FSD50K & LibriSpeech

Remove MSD (no audio), add FSD50K and Common Voice

Add Million Song Dataset (10K subset)

Remove broken datasets that don't exist on HuggingFace Hub

Completely disable pin_memory and num_workers for ZeroGPU compatibility

Fix CUDA pin_memory error in training

Access audio file paths directly to bypass torchcodec

Fix audio decoding and duplicate method issue

Fix FFmpeg dependencies and improve audio decoding

Remove VCTK dataset (uses unsupported loading script)

Fix missing is_dataset_downloaded methods

Re-add LibriSpeech and implement download status tracking

Replace broken datasets with working alternatives

Add dataset preparation feature for training

Fix training error - validate HF datasets are prepared

Fix dataset loading with Parquet-based datasets