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
Web page: use wide windows and fit narrow screens (#3176)
ad1a310 Matthias Nott committed 24 days ago
docs/AMD_HIP.md: why STRATA_SH_STREAM goes both ways, and leave it off on Windows with KV streaming (#3895)
c2490b4 Vop3p committed 24 days ago
AMD on Windows: the shared-expert fork with KV streaming is not measured, never picked, and warned about when set by hand (#3895)
5f81454 Niko1221 committed 24 days ago
docs/TROUBLESHOOTING.md: every entry of v0.2.4.2 is back, updated for 0.2.7
d98d8f9 Niko1221 committed 24 days ago
qsa_pattn: sm_6x prompt attention without partials, opt-in STRATA_PATTN_SM60=1 (#3943)
a338990 ruibeikaa committed 24 days ago
docs: the two master switches in the 0.2.7 release-notes draft
9db4970 Niko1221 committed 24 days ago
Master switches DS4_DEFAULTS=026 (DeepSeek) and STRATA_DEFAULTS=026 (Qwen): every default changed since v0.2.6 gives v0.2.6's behaviour; explicit per-feature variables and flags still win; every rule is a pure function with a v026 argument and a test
a3ab37f Niko1221 committed 24 days ago
App: "Faster answers on 12 GB NVIDIA cards" row; DeepSeek speed-ups Off also writes DS4_DEFAULTS=026
6f1bd21 Niko1221 committed 24 days ago
Shared speed-switch table and Terminal: DS4_DEFAULTS=026 with the DeepSeek off switch, the 12 GB NVIDIA row, and a Qwen "0.2.6 behaviour" row (Advanced only)
211fca8 Niko1221 committed 24 days ago
OpenAI logprobs, top_logprobs and allowed_token_ids on /v1/chat/completions (#3200)
9ea0449 xiezhipeng-git committed 24 days ago
native_mmvq: sm_89 MMVQ-IL tables opt-in with STRATA_MMVQ_IL_89=1 (#3453)
2299c8f Niko1221 committed 24 days ago
batch-groups: opt-in adaptive-tier rounds in the pipelined loop, STRATA_PIPE_ADAPT_EVERY (#4010)
a3bb38e SuperDistance committed 24 days ago
hip-<package> is accepted wherever the backend is named (#4013)
557aad0 Niko1221 committed 24 days ago
expert source: STRATA_RAM_ADAPT copies the incoming expert out of its stage buffer when it is still there, and moves up to 4 pairs a layer (#3060)
2b13a44 SanghunYun committed 24 days ago
docs/AMD_HIP.md: a hand-built engine has no hipBLASLt tuning table unless it is pointed at one (#4034)
a3b7d9d Niko1221 committed 24 days ago
setup: Maxwell gets a warning with the build recipe, not an env switch (#3218)
c966d7a Niko1221 committed 24 days ago
bench: tools/strata_bench.py reads the CUDA UMD version, the storage, the engine's build toolkit and GPU power and PCIe detail (#3250, #2111)
c0845b7 JoShMiQueL committed 24 days ago
docs: a crashed engine start names itself, and a setup run again keeps the installed size (#3243)
7783b1c Norm committed 24 days ago
prefill: the copy-wait dedupe skips a wait only when both slots were released in the current walk and the ring is longer than one group (#4005)
f217b23 Niko1221 committed 24 days ago
serve: a slot's shorter checkpoints come along at a give-back only when they are prefixes of the tokens copied back (#2526)
cea3ed4 Niko1221 committed 24 days ago
docs: the xe expert tier has no SSD fallback under the wrapper default, and a native pack needs --spec 2 or more (#2153, #1796)
05c1dab Yuu committed 24 days ago
docs/BATCHING.md: the cache settings that let a long solo request move into a batch slot without reading its prompt again (#3334)
e21a187 oscarlius committed 24 days ago
kq_cpu_bench: i-quant pairs no longer crash, and Q2_0 is timed on the CPU pool's path (#3871)
5bc1e24 Heinz N. Gies committed 24 days ago
nvfp4 local pack: read the checkpoint with cached sequential reads, not a process-lifetime np.memmap (#3377, #3378)
0fbbf06 Highlander committed 24 days ago
tools/gguf_reader.py: bound the header reads and report malformed metadata as ValueError (#2288, #1919)
409f246 Keyan Hu committed 24 days ago
serve: Chat Completions tool results keep the order of the calls they answer (#3395)
11237f8 Jakub Luwierski committed 24 days ago
native_mmvq: sm_89 (Ada) MMVQ-IL table from the cells two cards agree on (#3453)
bdecbff rodi4eg committed 24 days ago
serve: return 503 when the elastic image encoder cannot start (#2435)
092c7d1 i-Vladimir-i committed 24 days ago
verify: a window graph that does not fit frees the graphs of other window sizes instead of ending the engine (#3702)
a8c1d4b Niko1221 committed 24 days ago