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
Host tests green on this PC: the GuidedSetup very-large-model sentence gets its catalog key in all 8 languages; the Control lookup test no longer reads the real components folder
621e660 Niko1221 committed 24 days ago
PLE residency policy: the load log also says the expert arena size the rule counted (the proof that the layout is initialised and the sum is not 0)
bc5f84f Niko1221 committed 24 days ago
Merge f-027-win (PLE residency: admit the table only when RAM holds table + expert arena + 8 GiB; fixes the IQ3_S load failure, error 1453, on a 64 GB Windows PC)
ad1660a Niko1221 committed 24 days ago
App: off switches for the three 0.2.7 speed features in the model's Expert settings (Native engine settings > Speed), on by default, in eight languages
ddd4d5d Niko1221 committed 24 days ago
Terminal: the three 0.2.7 speed features as rows of the Advanced table, with their off values and plain words
3b3ea88 Niko1221 committed 24 days ago
KVA on by default for a cold prompt of 4096+ tokens (Qwen3.8 on one CUDA/HIP device); STRATA_KVA=0 gives the exact 0.2.6 prompt read
750ca8c Niko1221 committed 24 days ago
DeepSeek HIP: the matrix-core prompt attention (DS4_FAST_ATTN) is on by default on gfx1200/gfx1201 only; DS4_FAST_ATTN=0 is the off switch, gfx11 and gfx1151 stay scalar unless =1
1f079f9 Niko1221 committed 24 days ago
docs: the miss split is the gfx12 start value again (the owner's decision 26 Oct; revert of the opt-in wording)
e5cfa4b Niko1221 committed 24 days ago
gfx12: native packs start the decode miss split at pcie_frac 0.15 (R9700: +9.9% decode at 8K, +25.6% at 32K, +10.4% under serve, first-token KL 0.0013 vs 0.55 with top-1 12/12); explicit --pcie-frac wins, STRATA_PCIE_FRAC_DEFAULT=old restores 0.55
8d56d41 Niko1221 committed 24 days ago
STRATA_LOGITS_DUMP: a post-window logits tap for KL quality measurement (off by default); the sampler itself runs inside a CUDA graph capture, so the copy lives in the verify window's host-side epilogue where STRATA_DBG_NAN already copies from
45c9e62 Niko1221 committed 24 days ago
KVA: say why a configured run fell back to the full path (the observability criterion: an operator must be able to tell a configured-and-declined run from one that never asked)
053a415 Niko1221 committed 24 days ago
KVA: keep gdn_recurrence_variant's public signature - the SYCL tree defines that symbol, so a new parameter would break its link; the state-only form goes through a private impl
5b469ee Niko1221 committed 24 days ago
KVA: what the opt-in approximate late-layer prefill does, how to enable and capture, and the measured frontier
4abef6a Niko1221 committed 24 days ago
KVA: state-only GDN recurrence (no gate projection, no output norm) behind STRATA_KVA_GDN_NORM for the A/B
c5cf218 Niko1221 committed 24 days ago
KVA: layer-type filter, per-layer transfer skip, stream-entry release on a skipped MoE half
011e5be Niko1221 committed 24 days ago
KVA: optional approximate late-layer prefill (split 24, chunk-aligned suffix, capture/oracle modes)
f4012d6 Niko1221 committed 24 days ago
Merge f-027-r9700 (Qwen on one R9700: the PLE table kept in RAM when the PC can afford it, four bit-identical gfx12 decode switches back on) into f-027-int
7dacaf8 Niko1221 committed 24 days ago
Merge f-026-miss (DeepSeek 0.2.7: the RAM store's spare + put, the 12 GB-cell prompt fixes, opt-in matrix-core attention on HIP) into f-027-int
ca7b3d2 Niko1221 committed 24 days ago
docs: the release notes carry the 12 GB routing default with its held-out evidence
7e1a7cb Niko1221 committed 24 days ago
markov bias in Q8: the drafter's head is quantised to Q8_0 blocks as it loads (66 MB -> 33 MB, the pack untouched) and read by a Q8 variant of markov_kernel - the markov segment 1.80 -> 1.34 ms, the drafter's round 13.46 -> 12.95, the drafts' acceptance 90-95% (the same band)
a266bdf Niko1221 committed 24 days ago
head: 8 columns a pass for a GGUF head (every window rows<=kWin walks the 129280x5120 head once instead of twice, the same logits per column) - the drafter's fixed T=5 head 4.55 -> 2.83 ms, its round 14.49 -> 13.46; the verify window's head 3.48 -> 2.68 ms
e432845 Niko1221 committed 24 days ago
DeepSeek engine (AMD): the whole chunk's attention slab where the cards hold most of the experts
0f66c31 Niko1221 committed 24 days ago
DeepSeek engine (AMD): a small staging ring where the cards hold most of the experts
4aa645c Niko1221 committed 24 days ago