Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
Commits on main
Fix #3510: on a card that shares the PC memory (Strix Halo) the Work profile no longer reserves a fifth of the shared pool as graphics memory
dc16e3a Niko1221 committed 1 month ago
Fix #3469, #3437: UPDATE.bat and setup fetch the engine from the signed update feed when the release has no flat zip, and the minimum engine is 0.1.46
5cdb59c Niko1221 committed 1 month ago
Merge main (0.2.1.4 Terminal update: RX 9000 engine choice, download checks); engine choice keeps the RX 6000 tie-break
0cf83f6 Niko1221 committed 1 month ago
Fix #3425: the recommendation says how much of the picked size would be read from the SSD when RAM + VRAM is only just the tier's floor
18b4aff Niko1221 committed 1 month ago
Gyro: decode speed on NVIDIA (docs, 0.2.2 notes), Gyro-M's slowdown against Q2_0 16-21%
aae92ce Niko1221 committed 1 month ago
Merge f-022-uma-linux: Strix Halo + discrete card on Linux (no exit with PF_PAD, iGPU never last stage, UMA-aware expert cache, picture reader on a card with its own memory, nicer engine on the display chip); ports PR #3484 by John-Pi
926e7a3 Niko1221 committed 1 month ago
Fix #3503: the Linux cuda12 engine package carries sm_60/61/70; the source Dockerfile builds it
366c8db Niko1221 committed 1 month ago
Fix #3441, #3075: a drifted tool-call opener is read as the call; a call for a tool the request did not declare stays text
70f4240 Niko1221 committed 1 month ago
setup: say when the BIOS gives a big carve-out to a Strix Halo this run does not use (the OS loses that RAM; the experts and the desktop need it)
30a6705 Niko1221 committed 1 month ago
Fix #3504: a two-value option (control-vector-layer-range) is written and removed as two argv entries, and an empty Advanced value is no longer a bare flag
b8b381a Niko1221 committed 1 month ago
docs: Strix Halo with a card of its own on Linux (RAM counted once, split order, picture reader, nice)
25d025f Niko1221 committed 1 month ago
gpu_fit: a chip that shares the system memory adds only its BIOS carve-out beside the RAM (a Strix Halo model was counted as VRAM and as RAM)
8895949 Niko1221 committed 1 month ago
prefill gemm: skip the padded dequant on a device that has no padded product (a gfx12 stage of a gfx1151 + gfx1201 split)
0da82e7 Niko1221 committed 1 month ago
Fix #3501: in_trash splits backslashes, so a Windows Recycle Bin path is found on Linux too
95cfbd4 TheHubUser4860938 committed 1 month ago
Fix #3507, #3424: first-run 'Use your own GGUF' goes to the local GGUF page instead of making the folder the data folder (which downloaded the files again); preparing own files never uses their folder as data_dir
c8b1436 Niko1221 committed 1 month ago
Control 0.1.46.3: version bump for the Terminal/Control-only update (engine packages unchanged)
9de04cd Niko1221 committed 1 month ago
Progress words: a file being repaired after a wrong SHA-256 says so in plain words
31ca248 Niko1221 committed 1 month ago
Downloads: every ranged answer is checked, a file with a wrong SHA-256 is repaired piece by piece instead of deleted
63214a5 Niko1221 committed 1 month ago
Updates: an AMD card with an unknown gfx never gets the newest HIP build by version; read the gfx from the Windows display adapters, choose by GPU name, keep the installed variant
8662083 Niko1221 committed 1 month ago
Fix #3490: tuning stops waiting for memory when none is being freed instead of ten minutes per candidate, and says how much is free and needed
75d0d0b Niko1221 committed 1 month ago
Fix #3502: a CUDA-init failure in a tuning run is not recorded as OOM (the memory guard info line no longer decides), and a command that passed before is not blacklisted by a later failure
6515aeb Niko1221 committed 1 month ago
Linux integrated GPU: the auto expert cache counts the RAM once (cache <= RAM available - desktop reserve); the APU warning uses the same reserve
f801953 Niko1221 committed 1 month ago
Fix #3488: a reused PID in a child marker without a start time no longer blocks every load; the engine hash check reads twice before refusing
5375240 Niko1221 committed 1 month ago
prefill: a padded activation is gfx11 only, and Gemm::native copies it unpadded instead of ending the engine
ab61f89 johnpi committed 1 month ago
AMD: built-in Radeon graphics (Strix Halo) is never the last stage of a layer split, the picture reader leaves it, --gpus all leaves it out beside two cards, engine nice +5 on Linux with such a chip, split note
034422c Niko1221 committed 1 month ago
generate: --spec on a native pack starts the speculative loop for a 1-token prompt (spec_pos -1 = unused, 0 is a real position)
aece46d Niko1221 committed 1 month ago
dsv41: tq_convert --act (input-column weighted encode from the gate-3 calibration sums)
da70196 Niko1221 committed 1 month ago
dsv41: portable (FMA) tile test target, docs for the trellis experts
12a3c07 Niko1221 committed 1 month ago