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++
gfx1151 on Windows: a bare --prefill auto may reach 8192 again (the 6144 cap of #1630 stays for other cards); STRATA_GFX1151_CHUNK=0 keeps 6144
Niko1221051db74cdfec
diff --git a/src/program/generate.cpp b/src/program/generate.cppindex 4cd39ac5a..053b9f293 100644--- a/src/program/generate.cpp+++ b/src/program/generate.cpp@@ -2798,6 +2798,17 @@ int main(int argc, char** argv) {// So its expert arena stays unregistered. STRATA_UMA_REGISTER=1 (or STRATA_ARENA_PIN_GIB=0) registers it as before.if (strata::core::igpu_uma_arena_default(ad.gcnArchName, all_igpu, win_host))strata::core::arena_uma_no_register_flag().store(true);+ // gfx1151 on Windows (#1630 capped a bare `--prefill auto` at 6144 for a 7900 XTX's commit charge): the Strix Halo has its+ // 8192 chunk back. IQ3_S, 8K prompt: 691 -> 787 tok/s (one chunk streams 1,785 experts instead of 2,782), 64K 713 -> 748;+ // 4K is one chunk already. STRATA_GFX1151_CHUNK=0 (or STRATA_GFX1151_DEFAULTS=0) keeps 6144.+ if (win_host && all_igpu && o.prefill_auto_bare && o.prefill_auto_cap > 0 && std::strncmp(ad.gcnArchName, "gfx1151", 7) == 0) {+ const char* c1 = std::getenv("STRATA_GFX1151_CHUNK");+ const char* d1 = std::getenv("STRATA_GFX1151_DEFAULTS");+ if (!(c1 != nullptr && c1[0] == '0') && !(d1 != nullptr && d1[0] == '0')) {+ o.prefill_auto_cap = 0;+ o.prefill_chunk = std::max<int64_t>(o.prefill_chunk, 8192);+ }+ }} else (void) cudaGetLastError();}#endif