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++
dsv41: Maya-port defaults from the owner-PC A/Bs (ON: win_submit, flow pump, Stop, lane-free; OFF with switch: timer, Engram readers, unbuffered Engram, cold prefetch, DSpark cum)
Niko12216c6267f33c90
diff --git a/src/dsv41/ds4_main.cpp b/src/dsv41/ds4_main.cppindex c3a140d94..0bf6f2dbc 100644--- a/src/dsv41/ds4_main.cpp+++ b/src/dsv41/ds4_main.cpp@@ -141,11 +141,11 @@ static int g_round_nd = 0;// The drafts' cut on the product of their confidences (DeepInfer's cumulative rule): a draft is verified while the chance// that it and every draft before it are kept stays at least this - a long window whose later drafts are likely lost is// not verified whole. STRATA_DS4_DSPARK_CUM=F, a request's dspark_cum key; 0: off, only the floor on each draft's own-// confidence (dspark_conf). Project Maya's default 0.5 (a Radeon 8065S: decode +6.5%, the same tokens)+// confidence (dspark_conf). Off by default (not measured faster on code: H100 / owner PC); Project Maya runs 0.5 (a Radeon 8065S: decode +6.5%, the same tokens)float dspark_cum_default() {static const float v = [] {const char* e = std::getenv("STRATA_DS4_DSPARK_CUM");- return e != nullptr ? (float) std::atof(e) : 0.5f;+ return e != nullptr ? (float) std::atof(e) : 0.0f;}();return v;}diff --git a/src/dsv41/ds4_model.cu b/src/dsv41/ds4_model.cuindex 411f93265..abbccf6c6 100644--- a/src/dsv41/ds4_model.cu+++ b/src/dsv41/ds4_model.cu@@ -2420,13 +2420,12 @@ struct Model::Impl {// the row cache a small share of the RAM free (DS4_ENGRAM_CACHE_MB: its size, 0 off; DS4_ENGRAM_THREADS)const size_t row = (size_t) pk.eg_hd + pk.eg_hd / 32;#if defined(_WIN32)- // (Windows: each thread's ReadFile is synchronous, so the threads are the queue depth - a 28K-token prompt- // in one chunk, 1.36M rows before layer 1: 10 threads 17.1 s, 32 14.1 s, 64 13.4 s on a 20-core PC)- int nth = 64;-#else+ // (Windows: each thread's ReadFile is synchronous, so the threads are the queue depth - Project Maya: a 28K-token+ // prompt in one chunk, 10 threads 17.1 s, 64 13.4 s on a 20-core PC. Here (6 cores, shipped profile) 64 readers+ // gave +4.5% on a 4K prompt but no decode gain in every pair: DS4_ENGRAM_THREADS=64 turns them on)+#endifconst unsigned hw = std::max(1u, std::thread::hardware_concurrency());int nth = (int) std::min(16u, std::max(4u, hw / 2));-#endifif (const char* e = std::getenv("DS4_ENGRAM_THREADS")) nth = std::max(1, std::atoi(e));const uint64_t avail = ram_avail();uint64_t mb = avail == ~uint64_t{0} ? 64 : std::min<uint64_t>(256, std::max<uint64_t>(16, (avail >> 20) / 64));@@ -2437,10 +2436,10 @@ struct Model::Impl {// Windows: the Engram tables' files that hold nothing else are unmapped and read unbuffered - an SSD's// queue kept by the reader threads instead of the page cache's (whose reads of one file from many threads// went ~255K rows a second here); unbuffered reads of a mapped file take a slow path, so only unmapped- // ones (DS4_EG_DIRECT=0: buffered, mapped, as before)+ // ones (off by default: -8% on a 4K prompt here; DS4_EG_DIRECT=1 turns it on)std::vector<char> eg_direct(pk.files.size(), 0);#if defined(_WIN32)- if (pk.gguf && (std::getenv("DS4_EG_DIRECT") == nullptr || std::atoi(std::getenv("DS4_EG_DIRECT")) != 0)) {+ if (pk.gguf && std::getenv("DS4_EG_DIRECT") != nullptr && std::atoi(std::getenv("DS4_EG_DIRECT")) != 0) {std::vector<char> eg_file(pk.files.size(), 0), other(pk.files.size(), 0);for (const auto& kt : pk.t)if (kt.second.file >= 0 && (size_t) kt.second.file < pk.files.size())@@ -5084,12 +5083,12 @@ struct Model::Impl {const float lim = g.lim;{ // (os::prefetch_cold, as above: the resident ones are left alone - on Windows the hint walks every page of a// range, ~0.13 ms an IQ2_XXS expert, on the lane's critical path before the first entry computes.- // DS4_IQ_PF_ALL=1: every range, as before)+ // off by default (the mmap path, owner PC: not faster in every pair); DS4_IQ_PF_COLD=1 turns it on)std::vector<std::pair<const void*, size_t>> want;for (const Ent& en : ents)if (!store && (l.ram.empty() || l.ram[(size_t) en.e] < 0))for (int r = 0; r < 3; ++r) want.push_back({mx_mat(l, en.e, r), mx_part_bytes(l, r)});- static const bool pf_all = std::getenv("DS4_IQ_PF_ALL") != nullptr && std::atoi(std::getenv("DS4_IQ_PF_ALL")) != 0;+ static const bool pf_all = std::getenv("DS4_IQ_PF_COLD") == nullptr || std::atoi(std::getenv("DS4_IQ_PF_COLD")) == 0;if (pf_all) os::prefetch(want);else os::prefetch_cold(want);}diff --git a/src/dsv41/ds4_os.hpp b/src/dsv41/ds4_os.hppindex 19d0722eb..916d993fd 100644--- a/src/dsv41/ds4_os.hpp+++ b/src/dsv41/ds4_os.hpp@@ -40,11 +40,12 @@ namespace strata::ds4::os {/// Windows sleeps in 15.6 ms ticks unless a process asks for finer ones: the CPU lane's service thread naps (sleep_for 20 us)/// between tokens, and the device's first wait of the next token then waits out the tick. 1 ms for the process's life-/// (Project Maya #358); DS4_TIMER_MS=0 leaves the system's resolution alone. Elsewhere nothing.+/// (Project Maya #358). Off by default (owner PC, shipped profile, 3 pairs: not faster in every pair); DS4_TIMER_MS=1 turns it on.+/// Elsewhere nothing.inline void fine_timer() {#if defined(_WIN32)const char* v = std::getenv("DS4_TIMER_MS");- const int ms = v != nullptr ? std::atoi(v) : 1;+ const int ms = v != nullptr ? std::atoi(v) : 0;if (ms > 0) timeBeginPeriod((UINT) ms);#endif}