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++
nv_defaults: the sm_86 speculative start-up adaptation is opt-in (STRATA_SPEC_ADAPT_SM86=1); owner round 9 measured 0.995 on the A4000s and a length-confounded 3060, so the earlier gain did not reproduce
Niko1221a4c503d7cec9
diff --git a/include/strata/core/nv_defaults.hpp b/include/strata/core/nv_defaults.hppindex 4d0a591b5..21a0bfa72 100644--- a/include/strata/core/nv_defaults.hpp+++ b/include/strata/core/nv_defaults.hpp@@ -31,7 +31,7 @@ std::vector<std::string> apply_nv_defaults(int major, int minor);namespace strata::core {-/// dflash2-kern: the speculative runs' faster start-up adaptation of the expert tier (STRATA_SPEC_ADAPT_EVERY / _WARM, read by+/// dflash2-kern: the speculative runs' faster start-up adaptation of the expert tier (opt-in on sm_86 with STRATA_SPEC_ADAPT_SM86=1 since owner round 9; STRATA_SPEC_ADAPT_EVERY / _WARM, read by/// generate.cpp): sm_86 only (z440 A4000 Q2_0; the P100 is opt-in through the two variables, notes-dflash2-kern.md). Not part of/// the exact table above: the residency it changes decides which experts the GPU rounds, as the adaptive tier always does. The user's/// own settings and the cards' off switches (STRATA_SM86_PERF=0) is kept.diff --git a/src/core/nv_defaults.cpp b/src/core/nv_defaults.cppindex bc90f6623..05799f836 100644--- a/src/core/nv_defaults.cpp+++ b/src/core/nv_defaults.cpp@@ -73,8 +73,12 @@ std::vector<std::string> apply_nv_defaults(int major, int minor) {void apply_spec_adapt_default(int major, int minor) {const int mj = strata::cc_major_of(major), mn = strata::cc_minor_of(minor);// sm_86 only: the P100 (sm_60) gained 4-16% on JSON/code/chat/tool calls but lost 3% on the 4K agent transcript (5/5 pairs), so there it stays opt-in+ // Owner round 9 (2026-10-22): on the z440 A4000s the 5-pair A/B against 0.2.2 was 0.995 (code 0.996, json 1.019, tool 0.973,+ // agent 0.991) and the 3060 was length-confounded, so the earlier +9..17% did not reproduce: opt-in (STRATA_SPEC_ADAPT_SM86=1)+ // until a clean A/B shows it.const bool sm86 = mj == 8 && mn == 6;if (!sm86) return;+ if (const char* v = std::getenv("STRATA_SPEC_ADAPT_SM86"); v == nullptr || v[0] != '1') return;if (off("STRATA_SM86_PERF")) return;const std::pair<const char*, const char*> kv[2] = {{"STRATA_SPEC_ADAPT_EVERY", "2"}, {"STRATA_SPEC_ADAPT_WARM", "200"}};for (const auto& p : kv) {