YuE2 Studio 4.0.0 — restoring compressed tracks, reports without a GitHub account, models kept loaded, KV cache in Q8

Added

• System diagnostics and a bug-report pack, as in ACE-Step Studio and MiniMax Music3 Studio: Settings - Local engine has System diagnostics - the version, the system, the graphics card against the CUDA builds, memory, which models, the own model and the optional parts are in place, which libraries the engine lacks, and the latest log lines - with Copy diagnostics, Make a bug-report pack (a zip of the diagnostics, the journal and the logs in the data folder's reports, with Show in folder), Report a problem and Logs folder. Keys never get in. Agents have diagnostics_get and diagnostics_report.
• A track compressed to MP3 can be restored, as in MiniMax Music3 Studio: Tools - Restore after compression runs Apollo on a library track or on an audio file from this computer and removes the artefacts of compression, giving 44.1 kHz in the channels of the source. The result is a new library track beside the source, "<title> · restored", with its words and cover; it plays, downloads and goes to the tools like any track, and restoring the same track again replaces it. The run shows its time and can be stopped. Apollo (66 MB, weights CC BY-SA 4.0) runs in audio.cpp's server on the stems' device and is downloaded when you ask for it, with its size shown first; it is also listed in Settings - Models. Agents have restore_status, restore_start, restore_install, restore_remove and restore_cancel.
• Mel-Band RoFormer separates on every card the studio computes on, as in Dub Studio: an NVIDIA card or driver that runs CUDA 12 only (Maxwell to Volta, GTX 16, a driver older than 580) takes RoFormer's CUDA 12 build with the engine's CUDA 12 libraries instead of the CUDA 13 build, which has no code for it, and an AMD or Intel card takes its Vulkan build instead of the processor, on Windows and Linux. Karaoke's Parakeet on such an NVIDIA card runs transcribe.cpp's CUDA 12 build instead of Vulkan, with the same libraries.
• A problem can be reported without a GitHub account: Report a problem (Settings - About and System diagnostics) opens a window - what happened, how to repeat it and what you expected, with the engine log attached - without the styles, lyrics and scores of your songs, which the bug-report pack leaves out of its engine logs too - unless you clear the box - that shows exactly what will be sent, and the studio files it through Studio Hub, whose bot opens a public GitHub issue and gives its number and link. The window has Bug, Idea and Question, each with the fields of its GitHub form. Its main button, Send from your own GitHub account, opens the GitHub form with what you typed already in its fields and its title, and answers come to your GitHub notifications; Send anonymously beside it is for those without an account. When the hub does not take an anonymous message now - too many messages, sending switched off or the hub out of reach - the window says so. The GitHub forms are now one per language: English, Russian, Chinese, Japanese and Korean. Agents have problem_preview and problem_report.
• What a stopped download left on disk is shown and can go: a model set whose download was cancelled or cut off says "Stopped: X of Y downloaded" on its card under Settings - Models, with Continue, which resumes it, and Delete, which takes the unfinished files away so the next download starts from the beginning; Your own YuE2 model shows the same for a community model. X is what really arrived - the unfinished file is reserved at its whole size. Agents read it as partial_bytes in models_catalog and models_own_list, and have models_own_discard.
• The models stay loaded between songs where the card has room: Settings - Local engine has Keep models loaded - Auto, Yes or No - in place of the switch. Auto keeps both halves of the backbone and the VAE on the card when the set kept whole, a full-context cache for each track and a reserve fit the card's free memory at the engine's start, so the next song does not load them again; kept models are unloaded before stems, Audio to MIDI, karaoke, recognised lyrics or the writing assistant on the card, and after 10 minutes without a song. A switch saved on reads as Yes, off as Auto. Under the choice "Now: ..." says whether the running engine keeps them and why, with the GB it weighed.
• The KV cache goes to Q8 by itself when F16 does not fit the card: Settings - Local engine has KV cache - Auto, F16 or Q8 - with Auto by default. Auto keeps the cache in F16 while the model set and a full-context cache for each track of a request fit the card's free memory at the engine's start, and takes Q8 when they do not, before any track is given up. Q8 takes half the memory of the cache: in the studio's runs a 600-second song in BF16 peaked 1.3 GB lower and took 177 seconds instead of 188, a 120-second one 0.5 to 0.6 GB lower in about the same time; it works only with flash attention on an NVIDIA card (CUDA), and on the processor, Vulkan or with flash attention off the cache stays F16 with that reason. The same seed on another type of cache gives another song, so the song's parameters name its cache. The engine log says which cache a start took and why, in gigabytes, Settings shows it under the choice, and the window says once in a run of the studio when Auto has switched to Q8. Agents set it with engine_options_set and read it in engine_options_get.
• Four more LoRA in the catalog: 1980s Italo Disco (trigger italodisco) and 1980s Two-Tone Ska (trigger twotone) by James E. Petts, both CC0 sound adapters; Keygen Music by monsterovich, a composition and sound pair at AR 0.3 and NAR 0.5 trained on instrumentals; and Instrumental, no voice by stevensu (CC BY-NC), a composition adapter that keeps songs without lyrics from singing, humming or chanting. A LoRA whose strength lies in its file's metadata is now read at that strength.
• Your own YuE2 model sings in place of the set's (#208): a YuE2-3B fine-tune in GGUF for yue2.cpp, such as goldhub/YuE2-RS-Finetunes, put in models/yue2-cpp/own is chosen under Settings - Models - Your own YuE2 model, and the engine runs it in place of the set's model with the set's VAE; Do not use goes back to the set's. The same list has Community models - the nine merges of Ruach Studio's YuE2-RS-Finetunes in Q8_0 (3.8 GB) and Q6_K (2.9 GB), the authors' choice first - which the studio downloads into that folder when one is chosen, with the percent, Cancel and the Hugging Face mirror, checked against the SHA-256 of a pinned revision; a downloaded one is removed to the recycle bin from the same panel, and the download shows in the panel the moment one is chosen. Create - More and the side panel name the own model when it sings. A file of your own the studio neither downloads nor checks, a chosen file that is gone stops the engine's start by its name, the decoder companion stays off in Auto with it, and LoRA training and dataset preparation still run on the set's model. The song's parameters name the file. Agents have models_own_list, models_own_select, models_own_cancel, models_own_remove and models_own_reveal.
• On Linux songs are made on an NVIDIA card through CUDA, as on Windows: the Linux engine had only Vulkan for every card. It now carries the CUDA 13 build (Turing and newer, driver 580 or newer) and the CUDA 12 build (Maxwell, Pascal, Volta and GTX 16, driver 525 or newer), chosen by the same table as on Windows, so Flash Attention and the KV cache in Q8 work there too; other cards stay on Vulkan. The CUDA runtime and cuBLAS of the build are downloaded from NVIDIA's packages the first time the engine starts on it, with the percent on the start screen, into one folder the audio.cpp server of stems, karaoke and audio to MIDI shares (CUDA 13: about 440 MB, CUDA 12: about 585 MB).
• A computer without a graphics card to compute on is told what that means, as in MiniMax Music3 Studio (#224): where songs would be made on the processor - no card, or built-in graphics below the smallest set - the first screen's recommendation and Settings - Local engine say that this is many times slower, tens of minutes and hours, and that a discrete card is faster.
• A crash of the studio leaves a trace on disk, as in ACE-Step Studio and MiniMax Music3 Studio: logs/yue2-studio.log in the data folder records each start, a panic of any thread and, on Windows, an exception nobody handles (an engine library running an instruction this processor lacks, a bad access) with its module and offset, kept to 8 MB with the previous start's beside it; keys are masked.
• The style's parts explain their fields (#211): under the singing language, the tempo, the voice and the sound of Edit style parts a short line in the window's language says what the field sets, what Automatic leaves to the model and what words to type there.
• Training shows the part of the step it is working on (#166): under the bar of a run, "Step N of M · part k of K" names what the trainer does now, from the composition's passes to updating the adapter, and the bar moves between whole steps as the parts go by; a slow step no longer looks stuck. Agents read the same in the training status as step_stage.
• A song that ran out of video memory says so: on Windows with an NVIDIA card a song whose card memory is full does not fail - the driver lends it the computer's memory and it goes on tens of times slower, which looked like a song hanging for hours. The song's card now says "Out of video memory" with what to lower, and the engine log notes it once, while the song goes on. Before that, a request of several tracks makes only as many as the card's free memory holds for its length - a minute's song takes far less than a ten-minute one - and its card says "Fewer tracks" with how many were asked.
• The MIDI editor splits and joins notes (#146): Split Notes (Ctrl+E) cuts the chosen notes at the playhead, or where the right-click menu was opened; Join Notes (Ctrl+J) makes one note of the chosen notes of each pitch, from the start of the first to the end of the last, and leaves a pitch alone when a note of it outside the choice lies between. Both are in the right-click menu and in Edit, and each undoes in one step.
• The writing assistant follows your own rules, as in ACE-Step Studio (#209): Settings - Models - Writing assistant has Your own rules for the whole song, the style and the lyrics, added after the studio's guide and followed where the two differ, and Words to avoid, typed or loaded from a .txt file; editing a score and laying out recognised or published lyrics keep their words. Agents set them with assistant_set.
• A long song keeps its level after the seam of its sound (#183): a song longer than the KV cache holds gets its sound in chunks, and the next chunk used to come out 1 to 3 dB quieter for some 20 seconds after the seam. Each next chunk now hears 30 seconds of the song again before it is faded in, which keeps the level after the seam within half a decibel of a song made in one piece, for about 5 seconds more on a 4-minute song. Settings - Local engine has the overlap beside the KV cache, from 0 to 45 seconds.
• Your own subtitles become a song's karaoke, and its lines can be edited (#218): on the song panel, Load .srt / .lrc takes a .srt, .ass, .ssa or .lrc file as the song's karaoke in place of the recognised timings, asking first when the song has karaoke already; Edit lists the karaoke's lines with their starts, and a line changed there is saved as written while the lines left alone keep the times of their words. A file reads in the encoding it was saved in: UTF-16 or UTF-8 with its mark, plain UTF-8, or the system's code page such as Windows-1251, GBK or Shift-JIS. Agents have karaoke_set.
• A deleted song goes to a bin and comes back from there, as in ACE-Step Studio (#175): Delete puts a song in the Bin tab of the library with its stems, its karaoke and its versions, the message after it offers Undo, and Restore on the tab brings it all back for 30 days; after that, or with Delete for good and Empty the bin, the song leaves the library and its files go to the system recycle bin. A cover, a re-render, stems, MIDI, karaoke, a cover image or a video of a song in the bin is refused by its name until it is restored, in the window's language. Agents have library_trash_list, library_trash_restore, library_trash_delete and library_trash_empty.
• Lyrics are heard from a recording on this computer, as in ACE-Step Studio (#222, #207): Tools - Recognise speech can use the recogniser chosen for karaoke (Parakeet or Whisper) as well as OpenRouter, the karaoke recogniser first when one is chosen; and in Cover a button above the lyrics hears what the source sings and puts it in the lyrics to edit, asking first when the lyrics are not empty. The words are laid out in lines: a sentence's end closes a line, a pause opens one, a long line is cut after its last comma. Agents have lyrics_recognise.
• Flash attention can be switched on by hand (#228): Settings - Local engine has Auto, On and Off in place of the switch that could only turn it off. Under Auto flash attention is on again on RTX 50 cards, as in 3.8.3, where turning it off took twice the time and memory; if in the second half of a song the words turn into meaningless syllables, choose Off.
• Settings - Local engine says which KV cache the engine runs with and why (#203, #228): an empty field read 24576 while a smaller cache was in use. The field now reads Auto, a line under it names the cache of the running engine (the model's full context, or set by you), and a saved value that has not reached the engine yet says it applies at the next start.
• The minimum of the sound stage explains itself and warns before it cuts a song off (#225): a Min tokens close to the length limit kept the song from ending - the model sang the lyrics, then filled the rest with near silence and was cut off at the limit. The field now says what the number means (25 frames are one second; empty lets the model end the song itself), and a minimum of 90 % of the length or more is flagged under it.

Changed

• Karaoke and the dataset lyrics hear with CrisperWhisper, as in ACE-Step Studio and MiniMax Music3 Studio: the Whisper choice is now CrisperWhisper 2.0 turbo, a Whisper large-v3 turbo trained for the time of every word, run in the audio.cpp server the stems and Audio to MIDI already use - on the card's CUDA, Vulkan on another card, the processor or Metal - in place of Faster-Whisper and its CUDA 11 libraries. Segments the model fills a gap with are left out. The update screen downloads CrisperWhisper and then removes Faster-Whisper; a Whisper size left in the settings no longer names a model.

• Audio to MIDI shows its notes as it hears them (#185): the player draws the notes while MuScriptor works and says which chunk it is on, instead of staying empty until the end; the finished MIDI is the same as before.

• The where-it-runs table says what the studio picks on AMD and Intel cards: HT-Demucs and Mel-Band RoFormer separate the stems through Vulkan there and MuScriptor runs on the processor; the table under Settings - Models and the README said the stems were separated on the processor.

• The README's tables of files to download by hand name what the studio downloads: transcribe.cpp 0.2.1 archives, audio.cpp's server for HT-Demucs and MuScriptor, and HT-Demucs as audio.cpp's GGUF, where they still named the ONNX HT-Demucs and HOT-Step's MIDI transcriber.

• The MIDI editor is updated to the latest signal, as in ACE-Step Studio: rewind after recording, the direction of curves and lines, the shortest note length and the rest of upstream's work since our last update, with the studio's own mode, chord and section lanes kept.

• The library search finds a song by any of its words, as in ACE-Step Studio: the words of a search may come in any order, and it reads the lyrics, your note on the song and the settings it was made with - the seed, the model and the LoRA - besides the title, the style and the tags.

• The form's and the song panel's blocks fold and stay folded, as in ACE-Step Studio (ACE-Step Studio #63): in the form the style, the lyrics, the score, LoRA and quality fold by their title, and in the song panel Style, Note, Parameters and Lyrics do too; the choice is kept across restarts, and Parameters opens folded until you open it.

• Tools opens on the song that stems or MIDI are working on, as in ACE-Step Studio: the progress is in view at once, and the chosen song always stays in the Tools list, first when it would be past the twelve songs the list shows.

• The agent's guide names every platform the studio ships for, as in ACE-Step Studio: the first step of the MCP skill said the studio is a Windows program that needs an NVIDIA card; it now points an agent to the Windows, macOS (Apple silicon) and Linux (experimental) packages and says the engine also runs on the processor and on AMD and Intel cards through Vulkan. The README adds the owners' reports of a Radeon RX 7900 XT under Windows and an RX 7900 XTX under Ubuntu 26 (#196).

• A download that a proxy does not let through names the proxy, as in MiniMax Music3 Studio: when the studio goes out through the system's proxy or the one set in Settings - Providers and that proxy does not take the connection at all, the message names the proxy by its address (never its login) and says what to do - start the proxy program, or choose No proxy - instead of suggesting HF-Mirror, which sits behind the same proxy. A proxy that takes the connection but cannot reach the site is not blamed, and the mirror is still tried through it. A model download that failed now shows its reason in Settings - Models too, not only on the first start.

• Engines and tools up to date: the assistant on llama.cpp b12248 (b11900 before; CUDA 13.5 on Windows and Linux, the same Ubuntu 22.04 limit for its CUDA build), tempo, key and the listening pack on ONNX Runtime 1.31.1 (1.30.0 before) with its Rust binding ort 2.0.0-rc.15, Parakeet on transcribe.cpp engine-0.2.1 and stem separation, karaoke and audio-to-MIDI on the audio.cpp server server-31cb1a4a. The update screen offers them as a components update. On Windows the assistant's CUDA 13 build now comes with the CUDA 13.5 libraries of the same llama.cpp release, so a computer that runs the assistant on CUDA 13 downloads 426 MB of them again.

• On the processor the sound of a song is computed about 1.2 times faster: the engine now uses ggml's tinyBLAS matrix kernel on the processor for the Q8_0 and BF16 sets, as llama.cpp does; the same seed on the processor gives another song than in 3.8.5.

• The tempo of a song is measured on the processor: in 5 to 7 seconds a song, for Describe by ear and dataset preparation alike. ONNX Runtime no longer runs on the graphics card, and the update screen removes the CUDA, cuDNN and DirectML files of earlier versions: about 3.1 GB on Windows with an NVIDIA card, about 36 MB with DirectML.

• The writing assistant lets llama.cpp place its layers: the studio no longer asks for every layer on the graphics card, which on a card of 8 GB or less could stop Gemma from loading, as it did in Dub Studio.

• Studio Hub notices follow what you do at once: a notice closed or clicked, and a song that ends, bring the next notice without waiting for the minute's refresh; a link in a notice is counted by its address, and the hub's requests wait while a song is made.

• A score lost to a strong LoRA says so and makes the song again in one click: when the LoRA on Composition add up to more than 1.2 and the model could not write the score, the song's error names their strength and offers to make it again with Composition at 1.0 (two LoRA share it, 0.5 each; the sound keeps its strengths and the form keeps yours). The form warns under the LoRA when Composition is set that high. In the studio's own runs one LoRA at 1.2 wrote the score and at 1.5 did not, and two at 1.0 each lost the key every time.

• A score that lost only its header is kept more often: the engine also drops the stray text the model wrote in the header before filling in the meter, length and key, where such a score used to fail.

• Karaoke lines start where they are sung more often, as in ACE-Step Studio: the alignment counts a recognised word it has to skip, the words heard before the first line included, so a line is no longer pulled onto an earlier word that merely sounds alike.

• About is three blocks on one screen: the studio - its name, version and description, with Report a problem and the studio's GitHub; Author - who the author is in his own words and his links as buttons, each "Network · name" (Telegram, X, GitHub); Support the author - a bank card, Boosty, and All the ways, the donation page of the studio's site in the window's language (Apple Pay, Google Pay, crypto and more). The list of the other studios is gone from it. The README and the donation page point to x.com/nerualdreming as well.

Fixed

• A stopped Audio to MIDI run says it was stopped, as in ACE-Step Studio: a run you stopped ends as stopped, said plainly in the window instead of as a failure, and a karaoke made from the song panel that fails says why in the window's language.
• A song lost to an engine that crashed says what happened in the window's language, as in ACE-Step Studio: the meaning of the exit code (an access violation, an instruction the processor does not have, a library of the engine that did not load, the computer out of memory and others) and the engine's last line now reach the song's error in all five languages, not only in English.
• Tools says why Split waits when its model is in, as in ACE-Step Studio: when the stems model is downloaded but the audio.cpp server it runs in is missing or from an earlier release, the panel says so, with a button to Settings - Models.
• A screen reader can play a song from the library, as in MiniMax Music3 Studio (#171): the play control on a song's cover in the library, a playlist and Liked is a real button named by the song ("Play: <title>", "Pause: <title>" while it plays), reached with Tab and pressed with Enter or Space.
• On Linux the audio.cpp server finds its CUDA libraries, as in ACE-Step Studio and MiniMax Music3 Studio: the CUDA 13 and CUDA 12 servers behind stems, karaoke and audio to MIDI had no CUDA libraries beside them on Linux and could not start on an NVIDIA card; each now downloads its cuFFT from NVIDIA's packages into its folder and takes the CUDA runtime and cuBLAS from the engine's (see the CUDA engine on Linux above), and is started with both folders on its library path.
• An error stays on screen until you close it, as in ACE-Step Studio (#230): an error message vanished after three seconds, often before it could be read; it now stays until closed, a long one scrolls and its text can be selected and copied. Ctrl with the wheel scales the interface over every panel, and an assistant that writes too slowly to finish in three minutes says so instead of reporting that it cannot be reached.
• A song over a chord bed is sung to a melody (#223): the bed went to the model as the song's whole score, with the Vocal voice at rest in every bar, so the words were chanted on two or three notes and the song's score showed rests only. A bed under words to sing now opens the plan with its tempo, key and meter and holds each section to its chords, and the model writes the melody over them: the song keeps the score with that melody. An instrumental is played from the bed as before. The line under Write the bed names the chords each section plays - a two-line section plays four chords of its progression - in place of the whole progressions.
• A cloud recogniser that returns no timings no longer makes a dataset song an instrumental, as in MiniMax Music3 Studio: an OpenRouter answer without timed words is heard again with the vocal lifted out, as a local "no words" is, and then says to pick a model that returns timings instead of marking the song instrumental; while a LoRA trains, the vocal is lifted out on the processor, a separator that fails leaves no temporary file, and a second hearing that fails for another reason says so.
• Tools says when the library does not answer, as in MiniMax Music3 Studio: the song list shows a spinner until the library answers, and an error with Refresh instead of "No songs yet" when it does not.
• The window keeps up with a slow service, as in MiniMax Music3 Studio: the setup status, the engine start, the tools, LoRAs, training, the agent panel and the engine log were asked again on a timer whether or not the last answer had come; each is now asked again a pause after its answer, one at a time.
• A song no longer comes out silent when the card computes something other than sound: the engine computes such a step again up to three times, and a song still left with numbers that are not sound fails with that reason and a Make again button instead of being saved as silence. How often a step had to be computed again is noted in the studio's log and counted in the anonymous statistics. Agents have song_job_again.
• A training run the studio closed on goes on to where it was meant to end (#221): Train further on such a run now offers the steps it was started for, not four checkpoints more, and says "The studio closed at step N. Training goes on from checkpoint M up to K."
• Training on an RTX 20xx card says why it is slow (#199): when the trainer computes the layers that card has no BF16 for in F32, the run's card says training goes several times slower there and still finishes.
• A scaled studio in a browser tab fits the window, as in ACE-Step Studio (#165): at an interface scale other than 100 % the studio and its dialogs fit the tab, with the player at the bottom and no scrolling of the page, and the "slide to turn on" knob and a click on the MIDI player's notes follow the pointer.
• What the studio starts ends with it, and what it opens for you stays, as in ACE-Step Studio: the engine, the assistant's server and the other helpers end when the studio closes, also on Linux and macOS and before the update installer runs, while a file manager the studio opened stays open after it closes.
• Updating karaoke's engine while karaoke has run no longer breaks it, as in MiniMax Music3 Studio: an install over the engine the studio has loaded stopped at its first file, having removed some of the earlier ones, and karaoke reported the model missing until the next start. Files a running studio holds are now moved aside and removed at the next start, the new engine lands whole, and the update screen says it is used from the next start of the studio. A reserved port range is also found on a Windows in another language.
• The karaoke of a video moves up to 30 seconds either way, in hundredths (#231): the lyrics offset in the video window reached only 3 seconds in tenths; it now takes from -30.00 to +30.00 seconds, on the slider and typed in its field.
• Karaoke on lyrics longer than the song times only what is sung: when the text repeats verses the song never reaches, every lyric line was put on one moment of the song together with the sung ones. The sung lines now get their own times from the first copies of the text, and lines the song does not reach get none; a first line sung unclearly keeps its place from the first word heard.
• An instrumental writes a new score each time and keeps No score (#217): the score written for an instrumental stayed in the form, so the next instrumental was made from the same score, and No score was switched back to Melody + chords. The score written for an instrumental now belongs to that song only, a score you put in the form stays as you wrote it, and No score makes the instrumental without one.
• Lyrics edited after Reuse prompt stay when you come back to Create (#216): leaving Create and coming back loaded the reused song's lyrics over your edits; a Reuse now loads the song once.
• Cloud requests ride out a busy moment, as in MiniMax Music3 Studio: an OpenRouter request (the assistant, descriptions, covers) that the service answers with 408, 409 or 429 is sent again after the wait it asks for, at most a minute, up to four times; a server that cannot be reached is named by its host.
• A graphics card chosen in Settings that is no longer in the computer is said so, as in MiniMax Music3 Studio: the studio no longer points CUDA at a card number that is not there (which left every CUDA part of the studio without a card, and the engine went to the processor without a word); the engine start says the chosen card is not in this computer, Settings - Local engine says so too and asks to choose one, and the engine starts once one is chosen. When nvidia-smi does not answer, the choice is left as it was.
• The first start says when the engine computes on the card through Vulkan, as MiniMax Music3 Studio does: next to the card the recommendation is sized for, an AMD or Intel card on Windows and every card on Linux are named as computing through Vulkan.
• No Windows Terminal tab is left open beside the studio: on Windows 11 24H2 and later the studio started from a shortcut gets no console of its own and makes one without a window, which the engine and the other programs it starts share, so neither a console window nor a Terminal tab appears; started from a terminal, it still writes its log there. On Windows 11 22H2 and 23H2 with Windows Terminal as the default terminal the tab still opens, as before: those versions have no way to start a console without one.
• Stems and Audio to MIDI at the same time no longer break each other: audio.cpp's server at work is never replaced under a run - a run takes it as it is when it carries its model, else waits for the other run to end - and stopping one panel's download of the server leaves the other's going; a run stopped while it waits ends at once.
• A song is no longer refused because its style, lyrics and score leave too little room for the length asked (#203, #205, #206, #213, #182): 3.8.4 and 3.8.5 held cards of 8 GB or less, and RTX 50 cards of up to 16 GB, to a 9216-token context and stopped a song whose prompt and score left less room than its maximum duration. Every card starts on the model's full 24576-token context again, as in 3.8.3, and when a smaller context is set the engine sings the song as far as the room goes instead of refusing it.
• A song that stopped at its limit is marked, not mistaken for a finished one: the library shows "Cut short" on it, with the reason - the length limit (with a click to make it longer) or the style, lyrics and score taking too much of the model's memory.
• A long song without flash attention keeps its length without running out of memory (#201, #228): without flash attention the sound is now rendered in pieces sized to the card's memory (on a 16 GB card about three minutes each) instead of shrinking the whole context, so the song keeps its full length; a 240-second song rendered in 120-second pieces peaked at 8.2 GB of card memory instead of 14.3 GB in one piece.
• A song the graphics card runs out of memory on is made once more instead of lost: the studio starts the engine again with the next smaller KV cache (16384, then 12288, then 8192), makes the song again and keeps that size for this card and model set. The journal says so, Settings - Local engine shows the size as chosen after the card ran out of memory with a button to go back to the full context, and a cache set by hand is never changed. A song stopped by you, or lost while training or stem separation shared the card, is not made again.
• Describe by ear keeps to the processor when you chose it (#204): with the processor chosen in Settings - Engine the captioner still ran on the graphics card with its CUDA libraries; it now runs on the processor then, without them.
• A program of the studio that crashes without a word says where it crashed (#204, #215, #181): when the engine fails to start, or the captioner or the trainer stops on an exit code such as 0xc0000005 or 0xc000001d before writing anything, the studio reads Windows' record of the crash and adds the module it happened in, the exception and the offset to the error and the engine log - the file name only, never a path.