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
setup: the Mac control path says plainly that the Mac engine has no picture reader yet
9a22b78 Niko1221 committed 1 month ago
packaging: the engine package carries strata-vision when the build has it
417cbab Niko1221 committed 1 month ago
launch: a missing picture reader file is fetched again instead of starting an engine that cannot read pictures
debe8dd Niko1221 committed 1 month ago
A tester build for another AMD card family: one plain line (This build is for AMD RX 9000 cards ... Ask for the RX 7000 build) in the engine status and the start error
7ecfe4a Niko1221 committed 1 month ago
Engine packages: the folders named in BUILD.json lib_dirs go on the library path with the package folder (the Windows HIP package keeps the ROCm DLLs in rocm/bin), for serve, the binary smoke test and tuning
4db866a Niko1221 committed 1 month ago
Windows AMD: the engine is picked from the display adapters when no GPU tool answers (an AMD card, not an integrated Radeon), the app lists AMD cards on Windows and no longer refuses the hip backend there
8c676a7 Niko1221 committed 1 month ago
Merge f-145-pp-volta into f-146-volta (0.1.46): the community Volta/Pascal perf-PR group (35 commits)
8b3a967 Niko1221 committed 1 month ago
qsa prompt attention: STRATA_QSA_WARP=1 only forces the Turing kernels on cards that have the Turing MMA (sm_75+)
c61364f Niko1221 committed 1 month ago
launch: turning pictures on for a model that is already installed says it downloads only the picture reader, not the whole model
e291123 Niko1221 committed 1 month ago
launch: Pictures on goes through setup's Control mode (setup keeps the choice when the engine component carries strata-vision); the screen warns only when the installed engine lacks the reader
8d65058 Niko1221 committed 1 month ago
launch: the Folder screen says that model files already stored are moved into a new folder
e819ba0 Niko1221 committed 1 month ago
test: the bundle keeps one real python and launchers for the other names
ee0f328 Niko1221 committed 1 month ago
strata launch on a Mac: a Terminal that carries strata_control/mac_fit.py installs the model through setup.py's Control mode instead of adopting prepared files
d4abe4b Niko1221 committed 1 month ago
launch: the Pictures screen says plainly that this Terminal cannot set up the picture reader yet
78fc321 Niko1221 committed 1 month ago
Control: the NVIDIA driver version also from /proc/driver/nvidia/version (Fedora has no nvidia-smi without its cuda package); an integrated AMD GPU beside a usable NVIDIA card is not offered
254ee2b Niko1221 committed 1 month ago
metal: the opt-in HIP/Turing entry points (fused_gr_persist_init, fused_gr_read_persist, f16_range_read, gemm_f16_range_read) answer 'nothing to do' instead of stopping the engine; e2e script waits for the bundled engine
a932648 Niko1221 committed 1 month ago
metal: stubs for the 0.1.45 entry points Metal does not implement yet (stubs_0145.cpp) and the Gemm::bf16 / bf16_converts_x signatures, so the engine links from the 0.1.45 tree
f2fe9a9 Niko1221 committed 1 month ago
launch: Pictures on means the CPU picture reader on an AMD card, and a plain warning when pictures stay off
47e2343 Niko1221 committed 1 month ago
metal: the CUDA-name compat layer declares the seven calls the 0.1.45 sources added (compute capability, stream priority, cudaMemcpyPeer, DisablePeerAccess), so the Metal engine builds from the 0.1.45 tree again
eeeeddd Niko1221 committed 1 month ago
Mac install plan: no card, KV or RAM-mode options; the context follows unified memory; a Mac too small for the model is refused in plain words
ccb33b7 Niko1221 committed 1 month ago
launch: Pascal/Volta cards are admitted when they are the only NVIDIA cards, as setup does (V100 showed 'no usable graphics card' and no Cards step)
6c4dbe3 Niko1221 committed 1 month ago