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
prefill: opt-in sparse copied events and deduplicated slot waits in the streamed prompt walk (#4005)
a11f8b0 Vop3p committed 24 days ago
setup: use the published size when the download length probes fail (#2287)
a1d2fd7 Keyan Hu committed 24 days ago
setup: --gguf-dir files for Qwen3.6 no longer need a .done mark made by setup (#3606)
560cc9f Günter committed 24 days ago
Fix the follow-up of #4014 (a variable name already used in Chat.tsx) and two of my tests (#3254 iterator, #3656 raw paths)
96b6068 Niko1221 committed 24 days ago
expert cache: STRATA_VRAM_FREE=device reads the whole card's free VRAM through NVML, Windows + NVIDIA, opt-in (#1911)
3100ca0 Xiaomi2008 committed 24 days ago
setup: an explicit --rope-scale must be finite and at least 1 (#3393)
a332cc9 Jakub Luwierski committed 24 days ago
tests: cuda_spill_guard fails the CUDA test build when a Strata kernel spills or keeps a per-thread array in local memory (#2683, #2682)
3a3f35c Avicennasis committed 24 days ago
serve: opt-in LaTeX-safe JSON recovery for structured output (#2893)
2cc6ea1 Niko1221 committed 24 days ago
serve: the serial decode timing line shows the adaptive waits in its sum (#1864)
2251053 SanghunYun committed 24 days ago
bench: tools/strata_bench.py (BENCH.bat, bench.sh) names an AMD card on ROCm 7+ too (#3630)
c471588 l33tm4st3r committed 24 days ago
serve: an idle batch slot's decoded conversation is parked before an admission overwrites it (#2526)
01e404c hotschmoe committed 24 days ago
serve: /v1/responses accepts max_output_tokens of 0 and -1 (#4036)
9c2d35d Niko1221 committed 24 days ago
docs: GR_V3 and PROMPT_ATTN_IMMA defaults in DETAILS.md and the 0.2.7 release-notes draft
407f659 Niko1221 committed 24 days ago
mtp: the drafter's Q8_0 draft head and projections on the Pascal GEMV, opt-in STRATA_MTP_Q8_SM60=1 (#3751)
f213f78 ruibeikaa committed 24 days ago
calibrate: an engine that dies mid-measurement drops only its candidate (#3317)
41188ec Dzmitry Senchankau committed 24 days ago
iq_kernels: GP100 decode down projection staged in shared memory, opt-in STRATA_DOWN_WS=1 (#3867)
6a6d58a ruibeikaa committed 24 days ago
bench: the standard report names an AMD card and its ROCm runtime when only amd-smi exists (#3519)
2ac3482 l33tm4st3r committed 24 days ago
App: the DeepSeek card names the free disk the set needs, and the version line says "{N} GB RAM + VRAM, plus about {disk} GB of disk"
9e3fd0f Niko1221 committed 24 days ago
docs/TROUBLESHOOTING.md: the Windows AMD driver-reset entry is back
f06bfc3 Niko1221 committed 24 days ago
docs: STRATA_VCVARS, /metrics/prometheus behind the API key, and the Monitor's prompt speed (#3244)
e6f598c Norm committed 24 days ago
conversation copies: the private copy streams and events are handed back to a pool when a thread ends (#3350, #3470)
3478382 Niko1221 committed 24 days ago
serve: a checkpoint whose copy runs out of memory is skipped, it no longer ends the engine (#3350)
f62589f Niko1221 committed 24 days ago