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
sm_80 defaults of #3779 (wide dense dequant, MMVQ-IL sm_80 tables) follow STRATA_DEFAULTS=026: back to the 0.2.6 kernels and the shared table; explicit =1 and =0 still win; arch_defaults_test checks the rule; release-notes line
2579635 Niko1221 committed 24 days ago
Merge f-027-gh-eng-n (#3779 sm_80 tuning: wide dequant, MMVQ-IL sm_80 tables and the opt-ins)
1e278a2 Niko1221 committed 24 days ago
docs: the sm_80 defaults and the opt-in sm_80 switches of #3779, STRATA_GRAPH_PRIO and STRATA_MMVQ_IL_PF (#3779, #3712)
56d98f1 linxuhao committed 24 days ago
Decode: opt-in STRATA_MMVQ_IL_PF=1 hoisted-load, exact-integer interleaved GEMV kernels for Q6_K (#3779, #3712)
8b95bff linxuhao committed 24 days ago
mmvq_il_parity: the L2-residency arms (--hot / --gap / --persist / --ab), the clock warm-up and MMVQ_IL_TMIN (#3779, #3712)
c2d3fe5 linxuhao committed 24 days ago
Decode window: opt-in STRATA_GRAPH_PRIO=1 gives the window graph's kernel nodes their stream's priority (#3779, #3712)
0049fa7 linxuhao committed 24 days ago
docs: conversation cache copies on a slow host link, STRATA_CKPT_ASYNC and the park switches (#3779, #3712)
0e63d59 linxuhao committed 24 days ago
Conversation cache: opt-in STRATA_PARK_VRAM_MIB=N keeps parked conversations' K/V in VRAM in front of the host copy (#3779, #3712)
85e3dde linxuhao committed 24 days ago
Conversation cache: opt-in ready checkpoint pool STRATA_CKPT_POOL_READY=N, gathered small checkpoint copies STRATA_CKPT_GATHER=1, park traces (#3779, #3712)
ca37867 linxuhao committed 24 days ago
Conversation cache: opt-in STRATA_PARK_ASYNC=1 parks the K/V in the background and moves the dropped checkpoints (#3779, #3712)
f73b3c8 linxuhao committed 24 days ago
Conversation checkpoints: opt-in STRATA_CKPT_ASYNC (background device-to-host copy) and STRATA_CKPT_VRAM_SLOTS (#3779, #3712)
0f7224d linxuhao committed 24 days ago
Prefill GDN: opt-in STRATA_GDN_TC=1 runs the DeltaNet recurrence as a chunked scan on the tensor cores, sm_80 (#3779, #3712)
8c2a846 linxuhao committed 24 days ago
Prefill fused IQ experts: opt-in f16 mma path STRATA_PF_FUSED_F16=1 and warp-specialized kernel STRATA_PF_FUSED_WS=1 (#3779, #3712)
10c8cc6 linxuhao committed 24 days ago
batch slots: MTP drafts in the pipelined groups, STRATA_PIPED_MTP=N opt-in with --batch-mtp (#4121, #3794)
17878ac wannagoxuf committed 24 days ago
Merge f-027-rp15: route-prior default 1.5 on discrete NVIDIA cards of 13 GiB or less (owner decision)
1d45da3 Niko1221 committed 24 days ago
Split serving: check tool uses cold prompts per request and a single-card arm on the shared arena
1e9678c Niko1221 committed 24 days ago
Merge f-027-shs: STRATA_SH_STREAM off by default on Windows gfx1151 (owner decision)
b1bea3a Niko1221 committed 24 days ago
docs: Windows Strix Halo leaves STRATA_SH_STREAM off, with the token-id check; DeepSeek numbers say which build they come from
f932e41 Niko1221 committed 24 days ago
gfx1151 on Windows: the shared-expert fork (STRATA_SH_STREAM) is off by default
c77292f Niko1221 committed 24 days ago
Docs: Strix Halo Windows guide, plain emphasis and a stray blank line
e240850 Niko1221 committed 24 days ago
Merge f-027-gh-ctl: GitHub setup, Docker, docs and bench items for 0.2.7
538acd0 Niko1221 committed 24 days ago
docs/STRIX_HALO_WINDOWS.md: a Strix Halo Windows guide, measured again on 0.2.7 (#3029)
ee1a6ef Niko1221 committed 24 days ago
fused_gr: the persistent hyper-connection read's barrier only calls __nanosleep on sm_70 and newer (the sm_60 build of f-027-int2 failed on it)
f1dadd1 Niko1221 committed 24 days ago
Tests: the new catalog is checked in nine languages, French included
3b908cc Niko1221 committed 24 days ago