You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
image-max-tokens: the most tokens one image may take, for a projector whose count follows
the image's size (0, the default, keeps the projector's own cap). A camera photo took the cap's
thousands of tokens, each prefilled like a prompt token, which on a phone, where image tokens are
prefilled on the CPU, was twenty minutes for one picture; a few hundred reads the same scene in
under one. Rendered in every UI's settings from the parameter table, like every tunable.
Images on the phones (#484): the Android and iOS composers attach images, from the photos or
the camera, with the web UI's rules: at most 8 in a message and 16 MB each, as removable
thumbnails above the text until sent, a message of images alone allowed, the entries shown
disabled with the reason while the loaded model reads none (model.vision), and a message shows
the images it carried, which Edit and send again and Answer again send again unchanged. The
apps send what POST /api/conversations/<id>/messages already takes; a photo from the iOS
camera is encoded once as JPEG by the app, everything else goes as picked. The web UI is brought
to the same rules on the way: its Image button now stays, disabled with the reason, while the
model reads none; Answer again and an edit keep a message's images; a message of images alone can
be edited; the first refused file is the one reported.
Fixed
Stop during an image. A cancel reached the engine only at a decode boundary, and an image's
encoder has none: Stop pressed while a photo was encoded waited for the whole encoder. The
encoder now stops where it is through the library's eval callback, and the turn ends as
cancelled at once, on every UI and API.
An image paid again on every turn after it, on a model whose memory cannot be cut anywhere
(a recurrent or hybrid one, Qwen3.6 among them): the next turn diverges a few tokens past the
image, where the template drops the previous turn's empty think block, and the reuse fell back to
the checkpoint before the image, so the image was encoded and prefilled again. A checkpoint is now
taken after every image, as at the end of a prompt; the turn after a photo prefills its own tokens
only.
A projector that names another model is never paired with it. The pairing by header fell
back to the folder's only projector whatever its own header said, so a model with no projector of
its own (a Gemma beside a Qwen mmproj) loaded the wrong one and warned about it. The fallback
now holds only when one of the two headers carries no name; GET /api/models and the load agree.
A model is listed once, from where it really is. A scan directory inside the models directory,
or a models folder of links into shared storage (the phone's), listed the same file twice, once per
path. GET /api/models now keeps one row per file on disk, the models directory's listing first.
Changed
llama.cpp moves to upstream tag v0.9.0 (cd959ba of 2026-11-04), 164 commits past 8ca4060 (docs/seam.md). The fork branch bmoe/expert-ready-hook-0411 is that tag plus the one
expert-ready-hook commit, cherry-picked without a conflict: the hook still fires ahead of
upstream's tiled mul_mat_id path, the only seam file upstream touched carries the new FP8 E4M3
type and nothing near the scheduler's eval callback, and the backend directories are the same.
No engine change was needed. All 41 host tests pass, the overlap gates included, and on a real
35B model --auto prints the same greedy text as the previous pin at the same CPU time per token
(two runs per build). What the tag brings this engine's users: Clover-35B-A3B, Mistral Large 4,
Metis-1.6, DeepSeek V4.1 Flash, the LFM3 family, Nemotron Bolt 74B-A9B and Kolibri-1 among the
new architectures, the FP8 E4M3 tensor type, and Gemma 4 E2B/E4B prefill work upstream.
Four community rows in docs/community-benchmarks.md: a 96 GB mini-PC with Qwen3.8-Flash-Next Q2_K on the CPU-only and the Vulkan builds (#460, #462), and two phones from #177 (a 12 GB Snapdragon 6 Gen 1 class phone with flint 40B-A2.5B, an 18 GB phone with Qwen3.5-397B).