v0.1.4

What's Changed

• docs: add Awesome ACE-Step link to README by @ChuxiJ in #715
• Fix auto-labelling crash when "auto" device is selected by @Copilot in #714
• refactor(gradio): decompose LLM actions and harden UI test isolation by @1larity in #710
• Automated clean up of issues by @schneidergithub in #723
• refactor(gradio): decompose batch management with facade by @1larity in #716
• feat(openrouter): add sample/audio2code endpoints and expand generation params by @ChuxiJ in #729
• refactor(openrouter): inline audio2code for cover mode, remove standalone endpoints by @ChuxiJ in #730
• Fixing the issue with _unwrap_decoder function import (see #719) by @arsenylosev in #720
• refactor(api): decompose model service and reinitialize routes by @1larity in #718
• Exclude macOS from flash-attn wheel dependencies in nano-vllm to stop ~480MB download on macOS ARM64 by @Copilot in #734
• feat: Add NVIDIA Jetson Docker support with GPU acceleration by @toolboc in #735
• refactor(api): decompose sample/query/release task routes by @1larity in #725
• refactor(api): decompose runtime/job helpers (API FD) by @1larity in #726
• fix(generation): add timeout, progress fallback, and VRAM pre-flight … by @Uni404x64 in #671
• feat: add codes for openrouter_adapter by @DumoeDss in #738
• refactor(api): decompose api_server into focused modules by @1larity in #741
• fix(lrc): always return intermediate tensors so LRC generation works for all task types by @ChuxiJ in #750

New Contributors

• @arsenylosev made their first contribution in #720
• @toolboc made their first contribution in #735
• @Uni404x64 made their first contribution in #671

Full Changelog: v0.1.3...v0.1.4

v0.1.3

What's Changed

• Refactor (gradio part1) events wiring contracts by @1larity in #652
• Refactor (gradio part2) generation wiring decomposition by @1larity in #653
• fix: AutoGen 'No next batch available' when next batch is ready by @ChuxiJ in #659
• Refactor (gradio part3) mode wiring decomposition by @1larity in #654
• refactor(gradio-events): extract generation run and batch navigation wiring modules (PR5) by @1larity in #655
• Refactor (gradio part 6) mode wiring decomposition by @1larity in #663
• Refactor (gradio part 7) generation interface decomposition by @1larity in #668
• fix(gradio): restore language and optional parameter interactivity by @1larity in #672
• Refactor: Dynamic i18n Language Management by @orlagno in #686
• Fix Gradio audio volume persistence and playback reset by @1larity in #680
• Improving and updating the Hebrew language and adding features by @start-life in #694
• refactor: clean up i18n discovery and CLI help by @orlagno in #692
• Revert "Fix Gradio audio volume persistence and playback reset" by @ChuxiJ in #696
• feat: remove prompt parse to sample_query by @ChuxiJ in #700
• Fix garbled audio after switching from Remix/Repaint to Custom mode by @Copilot in #701
• feat: add model init API, model inventory, and MLX CFG/APG support by @ChuxiJ in #703
• fix: respect ACESTEP_INIT_LLM=false in lazy load and prioritize expli… by @ChuxiJ in #707
• feat: add Streamlit-based ACE Studio UI by @pabbasian in #693
• fix(gradio): reapply audio volume persistence with sane default by @1larity in #697
• fix: handle MLX incompatibility in python_embeded on macOS by @ChuxiJ in #711

New Contributors

• @pabbasian made their first contribution in #693

Full Changelog: v0.1.2...v0.1.3

v0.1.2

What's Changed

• fix(training): skip torch.compile when PEFT LoRA adapters are active by @FeelTheFonk in #640
• (File permisions) Set executable permissions 755 on start_*.sh scripts for consistency. by @Red007Master in #641
• fix: resolve symlinks in safe_path to prevent false rejections by @ChuxiJ in #648
• feat: LoKr adapter support, LoRA status fix, and training docs update by @ChuxiJ in #649

New Contributors

• @FeelTheFonk made their first contribution in #640
• @Red007Master made their first contribution in #641

Full Changelog: v0.1.1...v0.1.2

v0.1.1

What's Changed

• Refactor(handler part 14): split initialize_service into focused init-service mixins and expand unit coverage by @1larity in #617
• Refactor(handler part 15): extract training preset switching into dedicated mixin by @1larity in #620
• Refactor(handler part 16): extract service_generate into dedicated mixin + normalize legacy encoding messages by @1larity in #621
• Refactor(handler part 17): extract MLX helper mixins by @1larity in #625
• Implemented new Studio UI by @goedzo in #627
• Refactor(handler): decompose generate_music orchestration (part 18) by @1larity in #626
• Refactor(handler part 19): final facade cleanup post 17/18 by @1larity in #628
• fix: pass project_root instead of checkpoint dir to initialize_service by @ChuxiJ in #637
• Add cross-platform interactive manual launch scripts by @orlagno in #636
• fix: clear stale UI state on mode switch to prevent remix noise bug by @ChuxiJ in #638

New Contributors

• @orlagno made their first contribution in #636

Full Changelog: v0.1.0...v0.1.1

v0.1.0

What's Changed

• Refactor(handler part 10): decompose generate_music orchestration into reques… by @1larity in #588
• Fix start scripts by @DumoeDss in #592
• Refactor(handler part 11): decompose generate_music orchestration and add upl… by @1larity in #590
• Chore(workflow): enforce independent PR branches with pre-push guard by @1larity in #595
• feat: add lora training tutorial by @DumoeDss in #593
• Add lycoris lokr load and load lora api by @sdbds in #582
• Refactor(handler part 13): vae encode by @1larity in #594
• Refactor(handler part 12): vae decode by @1larity in #591
• refactor: relocate gradio_ui → ui/gradio and remove old shim by @ChuxiJ in #599
• refactor: move mlx_dit and mlx_vae into models/mlx by @ChuxiJ in #602
• refactor: consolidate scoring modules into acestep/core/scoring/ by @ChuxiJ in #603
• docs: link LoRA Training Tutorials and Side-Step advanced training docs by @ChuxiJ in #605
• Docs/link lora training tutorials by @ChuxiJ in #607
• fix: manifest path double-nesting causes training to find no samples by @ChuxiJ in #611

New Contributors

• @sdbds made their first contribution in #582

Full Changelog: v0.1.0-rc.1...v0.1.0

v0.1.0-rc.1

What's Changed

• fix: remix strength info by @ChuxiJ in #546
• fix: add tiled VAE encoding and CPU offloading to training preprocessing by @ChuxiJ in #547
• Add Opus and AAC audio output formats by @Copilot in #554
• Add .env support to launcher scripts to preserve user customizations across updates by @Copilot in #552
• Relax Python constraint to support ROCm 7.2 Windows (requires 3.12) by @Copilot in #553
• feat: Side-Step -- corrected LoRA/LoKR fine-tuning with interactive wizard by @koda-dernet in #557
• Include LoRA state in audio file UUID generation by @Copilot in #556
• Refactor(handler part 8): extract lyric alignment timestamp/score mixins by @1larity in #565
• Refactor(handler part 9): service generate by @1larity in #568
• fix: escape Rich markup in user-provided paths to prevent MarkupError by @koda-dernet in #573
• i18n: add missing training/export translations for he, ja, zh by @jayvenn21 in #574
• Make batch size configurable and persistent across UI actions by @Copilot in #567
• Add batch size cap validation in service_generate by @Copilot in #570
• Enable multi-LoRA stacking with per-adapter scaling (addresses #338) by @jayvenn21 in #571
• (fix):DGX Spark dependencies by @tonyjohnvan in #575
• Support meta auto by @ChuxiJ in #580
• fix: differentiate LoRA adapters in UUID and load LM codes on JSON im… by @ChuxiJ in #583
• fix: prevent stale src_audio from leaking into text2music (Custom) mode by @ChuxiJ in #584

Full Changelog: v0.1.0-beta.3...v0.1.0-rc.1

v0.1.0-beta.3

What's Changed

• docs: add AGENTS.md guardrails for AI-assisted contributions by @1larity in #475
• Fix: {TRACK_NAME} not replaced by actual track name in API. by @goedzo in #474
• fix: seeds param by @DumoeDss in #496
• refactor(handler part 4): extract diffusion mixin and add validation tests by @1larity in #490
• Fix repaint mode not populating lyrics from generated audio metadata by @Copilot in #489
• [Fix] Eliminate redundant LLM load/offload cycles in PT batch mode by @agorevski in #492
• Fix LoRA memory bloat: replace deepcopy with CPU state_dict backup by @Copilot in #499
• fix: phase-aware max_new_tokens to fix misleading progress bar stopping early by @agorevski in #493
• Add CodeQL analysis workflow configuration by @schneidergithub in #502
• Use CPU for LM by @Jay4242 in #497
• Add minimal Copilot instructions for repository by @Copilot in #506
• Feat update skills simplemv by @DumoeDss in #500
• Fix YAML indentation in CodeQL workflow configuration by @Copilot in #503
• refactor(handler part 5+6 ): Extend decomposition and harden source-audio analyze validation by @1larity in #508
• fix(handler): address post-merge review feedback in io/prompt/task/padding mixins by @1larity in #510
• Potential fix for code scanning alert no. 35: Uncontrolled data used in path expression by @schneidergithub in #507
• fix: improve error handling in training and generation handlers by @ChuxiJ in #513
• Doc only: Give coderabbit explicit guidance on module loc limits. by @1larity in #512
• Demote vllm→PyTorch backend auto-selection from warning to info by @Copilot in #511
• feat: add dgx spark support by @NicasioSirvent in #485
• support cover (exp) by @ChuxiJ in #479
• feat: add Side-Step training v2 (corrected LoRA fine-tuning) by @koda-dernet in #478
• Revert "feat: add dgx spark support" by @ChuxiJ in #514
• Skip torch.distributed initialization in single-GPU mode by @Copilot in #477
• Fix CodeQL alert #59: Sanitize checkpoint paths to prevent deserialization attacks by @Copilot in #525
• Streamline path traversal security fix: remove verbose logging and defensive checks by @Copilot in #526
• fix for intel gpu support by @xushengyuan in #515
• Refactor(handler part 7): continue FD extraction for batch conditioning and embedding pipeline by @1larity in #516
• fix(training): guard pin_memory_device when None to prevent DataLoader crash by @jayvenn21 in #517
• (Feat)(mlx): Add dit progress bar and Fix the final-step `break by @tonyjohnvan in #519
• Potential fix for code scanning alert no. 59: Deserialization of user-controlled data by @schneidergithub in #524
• Support local LoKr safetensors in adapter loader by @riversedge in #530
• (feat)(api)Enhance audio result data structure to include per-audio metadata by @tonyjohnvan in #531
• Address review feedback: remove redundant path validations and tighten exception handling by @Copilot in #532
• Potential fix for code scanning alert no. 59: Deserialization of user-controlled data by @schneidergithub in #529
• Polishing tooltips in Gradio to de-clutter the GUI by @pedro16797 in #527
• Refactor tooltip implementation in Gradio to improve GUI clarity and … by @ChuxiJ in #534
• fix: vocal language ui by @ChuxiJ in #535
• (feat)(mlx): Supports MLX Pre-compile model for DiT by @tonyjohnvan in #522
• feat: enhances the Extract mode by @ChuxiJ in #536
• fix: only hide optional outputs in Extract mode, not all modes by @ChuxiJ in #537
• fix: Lego mode UI redesign and Extract/Lego mode switch cleanup by @DumoeDss in #538
• feat: add inline help buttons with modal documentation to Gradio UI by @ChuxiJ in #542
• Fix nanovllm CUBLAS error on Turing GPUs by detecting bfloat16 support by @Copilot in #541

New Contributors

• @schneidergithub made their first contribution in #502
• @Jay4242 made their first contribution in #497
• @NicasioSirvent made their first contribution in #485
• @koda-dernet made their first contribution in #478
• @pedro16797 made their first contribution in #527

Full Changelog: v0.1.0-beta.2...v0.1.0-beta.3

v0.1.0-beta.2

What's Changed

• fix: guard vLLM and CUDA graph usage on 16GB GPUs to prevent OOM by @jayvenn21 in #173
• fix: pin torchao version to >=0.14.1,<0.16.0 by @ChuxiJ in #440
• feat(mlx): Native MLX backend for DiT diffusion on Apple Silicon (2-3x speedup) by @tonyjohnvan in #439
• feat(training): auto-load .caption.txt and .lyrics.txt files for LoRA datasets by @jayvenn21 in #438
• feat(training): add validation-aware graph + best-checkpoint tracking to LoRA training by @jayvenn21 in #437
• Add torch compile logic to LoRA trainer by @fidel1234xdd in #422
• Fixed the Edited Draft Injection by @chigkim in #389
• fix(gradio): refresh audio results and reset batch navigation on new generation by @jayvenn21 in #178
• fix(gradio): initialize UI controls correctly for SFT models in servi… by @jayvenn21 in #207
• Enhance MPS LoRA training by enabling gradient checkpointing in DiT training path by @riversedge in #401
• (fix)Add CORS support for direct open browser-based frontends (studio.html) by @tonyjohnvan in #404
• fix: auto-detect Mac/MPS and apply optimal configuration by @ChuxiJ in #443
• refactor: streamline default backend configuration for device compati… by @ChuxiJ in #444
• enhance: implement caching for VAE audio encoding to improve efficiency by @ChuxiJ in #445
• refactor: optimize audio encoding logic with caching enhancements by @ChuxiJ in #446
• Normalize audio against clipping and 32bit wav support by @lutzkirschner64-dot in #406
• fix: API 400 - absolute audio file paths are not allowed for temp folder by @goedzo in #447
• refactor: improve audio encoding logic with enhanced caching by @ChuxiJ in #448
• refact: Reorganized Advanced Settings UI & Added latent shift and rescale by @ChuxiJ in #452
• (feat)(mlx): Native MLX VAE acceleration for Apple Silicon by @tonyjohnvan in #459
• fix(repaint): pass lyrics conditioning to repaint generation pipeline by @jayvenn21 in #461
• Fix #455: Restore missing training keys in en.json by @lutzkirschner64-dot in #463
• Refactor(handler part 2): Extract init utility mixin and add focused init-service tests by @1larity in #456
• Refactor(handler part 3): Feat/handler init service mixin v2 by @1larity in #464
• fix: reset audio playback position on regenerate & add GPU tier i18n … by @ChuxiJ in #467
• [Bugfix] Only use FlashAttention if GPU supports it by @agorevski in #469
• Refine Gradio UI by @ChuxiJ in #468
• Make torchcodec optional for ROCM and Intel XPU platforms by @Copilot in #465
• Add AdamW 8 bit by @fidel1234xdd in #449
• feature: support lokr by @ChuxiJ in #471
• i18n-zh-fix by @ChuxiJ in #473

New Contributors

• @fidel1234xdd made their first contribution in #422
• @lutzkirschner64-dot made their first contribution in #406
• @agorevski made their first contribution in #469

Full Changelog: v0.1.0-beta.1...v0.1.0-beta.2

v0.1.0-beta.1

What's Changed

• Refact add inference by @ChuxiJ in #1
• feat ✨ : add lyrics alignment scores by @keylxiao in #2
• Fix lrc bugs by @ChuxiJ in #3
• cover/repaint test ok by @ChuxiJ in #4
• refact understand_music by @ChuxiJ in #5
• test ok by @ChuxiJ in #6
• Add rewrite lyrics by @ChuxiJ in #7
• support input timesteps by @ChuxiJ in #8
• max_model_len 8192 -> 4096 by @ChuxiJ in #9
• reduce vRAM usage for vae decode by @ChuxiJ in #10
• fix: 修复 offload 模式下的设备不匹配错误 by @DumoeDss in #11
• fix language issue simple mode by @ChuxiJ in #12
• support lora trianing & inter by @ChuxiJ in #13
• fix cover logic for api_server by @ChuxiJ in #14
• add docs and serve mode by @ChuxiJ in #15
• add docs and readme by @ChuxiJ in #16
• add model zoo by @ChuxiJ in #17
• fix audio play position reset by @ChuxiJ in #18
• update hf demo by @ChuxiJ in #19
• fix duration str bug by @ChuxiJ in #20
• fix vllm bug by @ChuxiJ in #21
• Update abstract with specific performance metrics and unified messaging by @Copilot in #22
• Fix author names in BibTeX citation by @ChuxiJ in #23
• Verify author name order in BibTeX citation by @Copilot in #24
• fix mac nv install by @ChuxiJ in #25
• feat: update skills by @DumoeDss in #26
• improve lora training by @ChuxiJ in #28
• fix lora reload by @ChuxiJ in #29
• add model downloader by @ChuxiJ in #30
• add model downloader by @ChuxiJ in #31
• openrouter compatible by @seaniezhao in #32
• fix duration by @ChuxiJ in #33
• fix cfg kv block allocate by @ChuxiJ in #34
• fix_tiled_vae_encode_bug by @ChuxiJ in #35
• fix refer audio shape by @ChuxiJ in #36
• fix 4 vram by @ChuxiJ in #38
• tutorial by @ChuxiJ in #39
• Fix TypeError: AceStepConditionGenerationModel does not support len() by @saltchicken in #53
• Fix runtime crash with integer seeds and correct type hints by @iamgrootns in #48
• Fix torch.compile crash: add len to dynamically loaded models by @Copilot in #47
• Update Readme.md by @Saganaki22 in #62
• fix missing torchao by @ChuxiJ in #65
• Update release status for acestep-5Hz-lm-4B by @ChuxiJ in #71
• fix: reset linux torch version for nano vllm by @DumoeDss in #72
• fix max audio code id by @ChuxiJ in #78
• Revert "fix max audio code id" by @ChuxiJ in #81
• add support for intel gpu, tested on U9 285H, with and without 5hz llm by @xushengyuan in #88
• Fix max audio code by @ChuxiJ in #103
• Fix AttributeError when llm_handler is None by @fspecii in #109
• Fix brocken HuggingFace link in README by @oumaklaus in #118
• Optimize env installation by @DumoeDss in #128
• Fix a Windows error that was unexpected at this time. (Fixes #134) by @start-life in #137
• fix: bat error by @DumoeDss in #148
• feat: Add the enforced use of lm models by @DumoeDss in #152
• Add tensorboard to training dependencies for LoRA by @tlennon-ie in #155
• fix: Windows path compatibility and UI freeze issues (Issue #113) by @DumoeDss in #159
• docs: add Korean translation for core documentation by @acidsound in #187
• feat(i18n): add translation support to LoRA training UI by @jayvenn21 in #205
• Adding Hebrew language by @start-life in #201
• docs: clarify ROCm / AMD usage and uv behavior by @jayvenn21 in #179
• fix: avoid epoch-boundary stalls during LoRA training on Windows by @jayvenn21 in #177
• fix path traversal vulnerability in audio endpoint by @Albab-Hasan in #161
• fix: guard against missing reference audio in text2music by @jayvenn21 in #171
• docs: add responsible disclosure security policy by @jayvenn21 in #169
• fix: disable CUDA graph capture for 5Hz LM during LoRA training by @jayvenn21 in #172
• CLI by @chigkim in #94
• fix: add tolerance to 16GB VRAM detection by @jayvenn21 in #175
• fix: Wrap audio paths in FileData for Gradio 6.x compatibility by @DemetrionWare in #181
• feat: expose denoise control for audio-conditioned generation in Gradio by @jayvenn21 in #176
• Extended API to allow Analysis only without generating new music. by @goedzo in #197
• Fix Gradio Audio component error by passing file paths directly by @Copilot in #219
• Fix MPS tiled_decode conv1d output length limit on macOS 14.7 by @Copilot in #222
• Fix 4B model support for 16GB GPUs by increasing VRAM detection tolerance by @Copilot in #221
• fix: use per-socket timeout instead of global setdefaulttimeout by @Albab-Hasan in #226
• Fix bugs and add debuglog by @ChuxiJ in #230
• Add musician-friendly guide to documentation by @sigalarm in #238
• Position Audio time labels lower to avoid scrollbar overlap by @rkfg in #234
• Feat/progress UI: Prevent Gradio slider math error when no audio files are found by @1larity in #194
• Include torchcodec on MacOS to read audio files. by @chigkim in #243
• docs: clarify AMD / ROCm usage and uv behavior by @jayvenn21 in #244
• fix(song): ensure lyrics / vocal conditioning is preserved during song generation by @jayvenn21 in #241
• fix: resolve intermittent CUDA assertion error in concurrent serving … by @ChuxiJ in #257
• fix: remove nano vllm install in toml by @ChuxiJ in #260
• feat(ui): add experimental Studio UI for REST API by @jayvenn21 in #258
• feat(mps): Comprehensive Apple Silicon MPS backend support by @tonyjohnvan in #256
• fix(mps): Address memory management and performance issues for Apple … by @ChuxiJ in #261
• feat: Add AMD ROCm support for Windows (RX 7000/6000 series) by @clowerweb in #210
• [WIP] Fix uv sync regression with CUDA version mismatches by @Copilot in #268
• fix: add anti-clipping normalization to prevent audio clipping/overload by @ChuxiJ in #276
• fix: validate training-incompatible settings before starting LoRA training by @ChuxiJ in #277
• Fix check_update.bat: handle line ending changes and remove pager blocking by @Copilot in #273
• Add manual for rocm Linux (Cachy-OS) and additional requirements file by @Neresco in https://github.com/ac...