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++

layer-major prefill on a split: the PLE layer is fine in a first stage that holds layers 0 and 1 (#3976)

Norihiro Nishimurac6f6db857b2a
diff --git a/src/prefill/prefill.cpp b/src/prefill/prefill.cpp
index 67a932f7e..f51490539 100644
--- a/src/prefill/prefill.cpp
+++ b/src/prefill/prefill.cpp
@@ -5641,7 +5641,12 @@ int64_t Prefill::lm_prepare(int64_t n, std::string& why, int64_t cap) {
const bool split = stage_lb_ != 0 || stage_le_ != g.n_layers || next_ != nullptr;
if (hand_in_ != nullptr) { why = "a layer split's chunk hand-off is in use"; return 0; }
if (m.pp || helper_ != nullptr || core::peer_portable()) { why = "a peer GPU"; return 0; }
- if (split && m.ss->ple.ready()) { why = "a layer split with PLE layers"; return 0; }
+ // the PLE block (layer 1) reads the token history: on a split it runs in a first stage that holds layers 0 and 1
+ // (a later stage never reaches it)
+ if (split && m.ss->ple.ready() && stage_lb_ <= 1 && !(stage_lb_ == 0 && stage_le_ >= 2)) {
+ why = "a layer split whose PLE layer is not in the first stage";
+ return 0;
+ }
if (embd_rows != nullptr) { why = "an image prompt"; return 0; }
// run_impl's debug hooks on the routing and the rows have no layer-major form
if (std::getenv("STRATA_DBG_FEAT") || std::getenv("STRATA_DBG_ROUTE_OVERRIDE") || std::getenv("STRATA_PREFILL_DUMP_R_ALL")) {