Skip to content

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 serve: with slots off the loop is as before (no anchor copy, same continuation test on an empty engine)

Niko1221fa3252ad5de1
diff --git a/src/dsv41/ds4_main.cpp b/src/dsv41/ds4_main.cpp
index 92c15d0fa..a55b717e8 100644
--- a/src/dsv41/ds4_main.cpp
+++ b/src/dsv41/ds4_main.cpp
@@ -678,7 +678,7 @@ int serve(strata::ds4::Model& m, const std::string& eos_list, bool spec, float c
// the conversation continues, reused in place - a conversation has to be in the model first: a fresh engine
// (the start of a run, so every restart) holds 0 tokens, and every prompt "continues" that to token 0; taken
// as a continuation, the take below would never run for the folder's conversations, the case they are for
- if (n_state > 0 && c == n_state && n_state < r.ids.size()) {
+ if ((n_state > 0 || slot_max <= 0) && c == n_state && n_state < r.ids.size()) { // (slots off: as before)
from = n_state;
continued = true;
}
@@ -790,7 +790,7 @@ int serve(strata::ds4::Model& m, const std::string& eos_list, bool spec, float c
}
held = std::move(pics);
m.checkpoint(0);
- if (!continued) m.pin(); // (a new run of prompts: where it may be rendered again from)
+ if (!continued && slot_max > 0) m.pin(); // (a new run of prompts: where it may be rendered again from)
const double t1 = now_ms();
long long produced = 0;
const char* finish = "length";