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
tools/dsv41/engram_fold.py --gate-kv-from: make the ple table the only variable, and the answer it gave
226d73b Niko1221 committed 27 days ago
dsv41: one runtime header for the tests on both backends (HIP or CUDA), so a CUDA box needs no shim on -I
5254376 Niko1221 committed 27 days ago
tools/dsv41/engram_fold.py: transpose their Engram layout, but the fold's reader is still wrong
3826885 Niko1221 committed 27 days ago
dsv41 win: RAM expert store (unbuffered reads, policy, lookahead fill), Windows direct reads, dense dequant for the chunked prompt, SMT option, store test
e7b3f15 Niko1221 committed 27 days ago
tests: the old-launcher test expects the 0.2.3.1 version line
7ddc17a Niko1221 committed 27 days ago
tools/dsv41: engram_fold.py (their k*q gate and wkv from their own shards) and pack_gguf accepts an override for the Engram tensors
d62ff4a Niko1221 committed 27 days ago
dsv41: --pack-info's device-type list is derived from iq::supported() instead of copied
488dd84 Niko1221 committed 27 days ago
dsv41: the Engram reader takes IST-DAS's IQ4_NL table, and a pack with an unknown Engram type stops cleanly instead of segfaulting
b2d2cf2 Niko1221 committed 27 days ago
Control 0.1.48.1 / Terminal 0.2.3.1: version bump for the Control-only hotfix of #3810 (app stays 0.2.3, engine packages stay 0.1.48); release notes
7b3e2ea Niko1221 committed 27 days ago
setup: the HIP ready-made-engine checks read gfx names with arch_names, not arch_numbers (an installed AMD engine was never recognised); tests for archs as a list and as a string
f5f205b Niko1221 committed 27 days ago
setup: BUILD.json archs / vision_archs are read as a list or as the ';'-separated string the packaging writes (setup stopped on the released Windows CUDA package: int(';'), #3810)
0edd96c Niko1221 committed 27 days ago
dsv41 win: pack_gguf reads the IST-DAS Engram parameters from tensors, IQ1_M type, packinfo 64-bit sizes, DS4_TRACE_AHEAD, cachesim
eb6c0a2 Niko1221 committed 27 days ago
dsv41: a CPU-only build path, and the intake's pack recipe fixed
c523890 Niko1221 committed 27 days ago
dsv41: device kernels for IQ1_S, IQ4_NL and IQ1_M (the three IST-DAS types that had none)
5c523d5 Niko1221 committed 27 days ago
dsv41: the CUDA build of the tree needs two portable edits (found on z440's sm_86 bring-up)
45dc5e3 Niko1221 committed 27 days ago
dsv41 on Windows+CUDA: IQ1_M device kernels (bit-exact vs ggml), IQ4_NL Engram rows, CUDA build fixes (strtok_r, shfl_xor, test language), IST-DAS pack names, hardware bench
c42f4f2 Niko1221 committed 27 days ago
serve: max_tokens -1 and 0 mean no cap again, as in 0.2.2 (clients send -1 for that); documented
5c8dd32 Niko1221 committed 28 days ago
release notes 0.2.3: hotfix heading, status/logs names, launch not setup, drop the unreleased-pack limit
5964d11 Niko1221 committed 28 days ago
launch config never keeps --mtp beside "mtp": false; one line says where the pack and draft layer are read from (B10, C08 note)
81a7260 Niko1221 committed 28 days ago
Save and launch saves the preset only after the restart question is answered Restart; the question names the running model and port
3b1ab28 Niko1221 committed 28 days ago
Folder step: models Strata already keeps stay where they are (the text said they were moved)
80a70b8 Niko1221 committed 28 days ago
a failed start ends its own runtime host, checks the port first and quotes only this attempt's logs (X01)
9222b6a Niko1221 committed 28 days ago
tests: the smoke test reads the engine's own words from .reason, the message is plain (X01)
d64ae1b Niko1221 committed 28 days ago
Merge f-023-tui1: TUI stress fixes (second launch reuse-or-ask with --restart, context change on the same port, status/logs find the launch, key kept, retune stops its own model, folder checks, Ctrl+C, --model-path wins)
27c0cab Niko1221 committed 28 days ago
Failed engine test start gives a plain cause and next step; rollback never says the same version twice; max_tokens below 1 is a 400 (TUI stress X01, #18, #15)
2f5f7c1 Niko1221 committed 28 days ago
Installer: an older feed shows its real reason; a folder with accented letters unpacks (TUI stress #11 A01b/A05, #12 A04)
0af0f9f Niko1221 committed 28 days ago
A missing shard is named with its file name in the CLI, the TUI and the app's plain issues (TUI stress #9, C05, #3478)
ea67f69 Niko1221 committed 28 days ago