Repository navigation
fix(webui): LFM2.5-Audio conversation panel follow-ups - #864
Merged
Merged
Conversation
A review of the LFM2.5-Audio conversation panel after it merged found three things to fix in the page's hook for a panel with requestMode 'panel'. With UI management, the function that loads the entry for a panel read the selection again after each wait. Press Run on an S2S entry that is not loaded, then pick another entry in the Model select while it loads: the page unloaded the entry it had just loaded and loaded the newly selected one in streaming mode. The function now gets the entry of the request. After each wait it checks that this entry is still selected; if not, it stops with "The model changed while it loaded; press Run again." instead of loading the other one. The panel had this message for after the load; it now comes from the page. One short wait is left as it is: the page's doLoad, which every model's Run uses, reads the selection again after it unloads another loaded model. On a server with a config file the panel did not call the page's ensureLoaded. So after the user imported settings that differ from the configured model, the panel still ran a turn, where the page refuses other requests with "The imported model settings require a reload through the server configuration." The run context's function, now named ensureLoaded, calls the page's ensureLoaded on such a server, and the panel calls it on both kinds of server. For an entry that the server does not list, the panel now shows the page's message too. The hook was described as general, but it checked the speech-to-speech inputs itself: that source audio is chosen and that no recording is running. The panel makes these checks now, from the sourceFile and sourceRecording props the page already passes to panels. It throws the error that the context's new warning() makes, and the page shows it as a warning like its own. The run context no longer carries the question. The warnings read as before, and nothing changes for the other panels.
Fixes from a review of the conversation panel after it merged. They are all in the panel's own files. - The conversation is a Svelte store. It was a module object, and only the panel that ran a turn drew it again. With UI management, press Run on an entry that is not loaded, open the Models tab and come back before the load ends: the panel mounted on the way back kept showing "loading the model" after the run bar said "Cancelled.", until the next Run. Every mounted panel now sees each change, and the refresh calls are gone. - A Run with another entry, such as JP after EN, cleared the conversation before anything was checked. So a turn that failed at once lost it: the model did not load, the question did not decode, or the user pressed Cancel. The turns now stay until the new turn is done, and that turn then starts the new conversation. - For an S2S entry under an id that is not the catalog's, the panel takes the family's ASR options out of the request. It decided this from the parameter props after the load and the upload, when they can be those of another entry the user selected meanwhile. Then the ASR options went out and the server rejected the turn. The options are now made when Run is pressed. - Starting a turn stopped only the live player. A player of an earlier turn, or the Result column's, went on under the new reply, and the Result column's player did not stop a reply that played live. When the reply is to play as it streams in, Run now pauses the page's audio and video players, and any player that starts stops the live reply. - A stream event whose audio the live player could not use failed the turn, though the result has the whole reply. Live playback now stops for that turn and the turn finishes. With live playback on, the Runtime log says why, and the panel's note says that live playback is not available. The note used to say that the browser cannot play the stream. - With turns of one entry listed and another entry selected, the header said they go with the next question, while the note said Run starts a new conversation. The header does not count them now, and their Leave out and Include buttons are disabled until their entry is selected again. - Each stream event updated the turn list, read its layout and set the status line: about 50 times a second in a test on an M3 Ultra. The list is now updated when text arrives, when the first audio comes and when the turn's state changes. The status line changes with each whole second of the reply, as in "Streaming the reply: 12 s". - Three checks that could not fire are removed: for a stream that ends without a result (taskStreamEvents throws then), for a conversation reset during a turn (it cannot be while the page is busy), and the try around source.stop() (each source is started before it is kept).
Rebuilt from this branch's source with the lockfile (node 24.19.0, npm 11.17.0): cd webui/native npm ci npm run build 0xShug0#858 and 0xShug0#859 changed the Qwen3 TTS, Breeze TTS and Higgs Audio TTS model specs without rebuilding the page, so this rebuild also brings those changes into the page.
Owner
|
@ykhrustalev Thanks! Merged. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-up to #857. A second look at the LFM2.5-Audio conversation panel found a few small bugs. This fixes them. Only the panel and the page hook from #857 change.
What was wrong
What changes for users
Left for later
Testing
On an M3 Ultra with headless Chromium:
npm run checkandnpm run buildpassThe last commit rebuilds the page. It also brings in model spec changes from #858 and #859, which did not rebuild it.
Build and run
These are the commands tested on the M3 Ultra; on other machines pick your backend.
Then pick an LFM2.5-Audio S2S entry under Voice Conversion / S2S, record or pick a question, and press Run.