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++
Fix #3480: the HIP queue priming stands aside when GPU_MAX_HW_QUEUES is 1 or 2
Niko1221f05725c33257
diff --git a/src/program/generate.cpp b/src/program/generate.cppindex 3cef8621c..ab1e4480a 100644--- a/src/program/generate.cpp+++ b/src/program/generate.cpp@@ -2233,7 +2233,7 @@ static double ple_direct_rows_per_s(strata::kernels::PleTable& t) {// it is the fast one). Streams are destroyed again; the runtime keeps the queues.// Two queues are enough (R9700, "nn": 0 slow of 32 fresh starts, unprimed 10 of 32); five (nnnnh) fixed the starts too but cost// decode 2 to 3.6% and two expert slots, four normal ones did not even fix the starts.-// STRATA_HIP_QUEUE_PRIME unset: gfx1200 / gfx1201 on Linux only, "nn"; "0": off; any other string is a spec+// STRATA_HIP_QUEUE_PRIME unset: gfx1200 / gfx1201 on Linux only, "nn" (not when GPU_MAX_HW_QUEUES is 1 or 2, #3480); "0": off; any other string is a spec// ('n' = a stream at the default priority, 'h' = at the highest) for experiments on other cards.static void prime_hw_queues() {const char* env = std::getenv("STRATA_HIP_QUEUE_PRIME");@@ -2252,6 +2252,16 @@ static void prime_hw_queues() {(void) cudaGetLastError();return;}+ // #3480: with the runtime capped to two hardware queues (GPU_MAX_HW_QUEUES=1 or 2, the user's own setting) the two+ // primed queues are all of them: the high-priority draft stream and the side stream share them and a 4K prompt+ // reads 39% slower (RX 9060 XT, 681-753 against 1188-1238 tok/s). Not primed then; a spec in+ // STRATA_HIP_QUEUE_PRIME still is.+ const char* cap = std::getenv("GPU_MAX_HW_QUEUES");+ if (cap != nullptr && std::atoi(cap) >= 1 && std::atoi(cap) <= 2) {+ std::fprintf(stderr, "strata generate: GPU queues not primed: GPU_MAX_HW_QUEUES=%s leaves none for the draft stream "+ "(STRATA_HIP_QUEUE_PRIME=nn primes them anyway)\n", cap);+ return;+ }spec = "nn";#endif}