Repository navigation
FreeVideo 0.4.15 - #342
Merged
Merged
FreeVideo 0.4.15#342
Conversation
… memory Making room for the decoder after sampling released the idle text encoder through the library's own unload, which copies every weight to the CPU first: 13 GiB for the video text encoder, at the moment memory was already short. Windows already let the encoder's tensors go where they are; Linux now does the same. The Mac keeps its own encoder release. The native encoder tests also run when the library is found next to the engine.
… white model Keep only the cat ready-made scene; answer progress questions about a white model; say clearly when it times out
…on an int8 install
Generated images keep readable names and are listed newest first in the editor's generation tool, with place, use-as-reference, delete and open-folder actions. A standard/HD size choice sits beside the aspect ratio on blank canvases.
…s kept on the card
…y of the weights When the planner puts every DiT block and encoder layer on the card, the weights load straight to it and are not pinned, so host memory keeps only the vocabulary and temporary buffers. A later request whose budget no longer fits them all copies the weights back to the host and streams as before.
…when locking pages fails If locking pages fails while the weights return to the host, they stream from ordinary memory; a failed move leaves the layers on the card, and the runner leaves whole-model mode whichever stack stopped. Tests cover the three cases.
… to retry it When the weights cannot be copied back to the host, the runner stays in whole-model mode so the next request tries again, and a stack that already streams is not converted twice. A test with a real card checks the layers stay on it.
… to defaults Optional Generate inputs appended in later versions (step cache, upscale factor, voice reference and others) could be empty in an older save, and ComfyUI refused the whole request. The prompt hook now resets them to their defaults and notes the names in the diagnostic report; prompt, size, duration, seed and step counts still report their own errors. Fixtures rebuild each release's saved node by script, with two real saves.
Report a problem next to failures everywhere; say why a chat cannot be deleted and offer to stop and delete it One report button across the studio, editor, agent pages and launcher, shown once per failure and not during automatic retries. A chat that cannot be deleted says why and can be stopped and deleted.
The Mac engine takes the same continuation inputs as the CUDA engine: the previous part's last frames and audio lead-in, overlap frames, hidden tail frames and first-frame or character references. Messages about continuing a video now say it needs an NVIDIA GPU or a Mac.
…source delta updates Windows checks a domestic mirror for release files and resumes interrupted downloads; mirror manifests count only with the release signature. Source updates apply as verified overlays with fallback to the full package. DMG and AppImage stay on GitHub.
Below 640 px the sidebar collapses to icons, the review page uses one column, long paths are elided in the middle, and the footer buttons stay inside the window. Desktop widths are unchanged.
…ode folder as a package
…card placement used
The workspace record now carries cudaMemGetInfo free, the WDDM local budget and usage, allocated and reserved, the process budget and the bound that decided, before and after making room. The compute and conditioning snapshots include the Windows counters, and the summary shows the workspace rows and the resident-model decisions of the request.
…g-only workspace The decoder runs after sampling releases its tables, so it is charged only the part of its workspace that the sampling allocations do not already cover. Full-size requests whose sampling and decode each fit the card are no longer refused. The low-memory reserve reduction still keeps the whole decoder workspace in its trigger.
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
1 participant
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.
0.4.15: Image generation keeps a history of every picture and needs less memory on both small and large graphics cards, error messages everywhere now come with a button to report the problem, and Windows updates download faster in mainland China.