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++

generate.cpp: the --spec-split comment no longer calls its windows exact (they are not bit-exact against no-split: batch-shape divergence); comment only

Niko12218aee4538f2fd
diff --git a/src/program/generate.cpp b/src/program/generate.cpp
index d381227de..0b153ee16 100644
--- a/src/program/generate.cpp
+++ b/src/program/generate.cpp
@@ -709,7 +709,7 @@ struct Options {
/// Stop when the model emits an end-of-turn token (<|endoftext|> 248044, <|im_end|> 248046, or --eos-ids).
bool stop_eos = false;
std::vector<int64_t> eos_ids = {248044, 248046};
- bool spec_split = false; ///< opt-in split verify window (the overlap study: exact, ~7% slower)
+ bool spec_split = false; ///< opt-in split verify window (the overlap study: ~7% slower; not bit-exact against no-split, the batch shape of its windows differs)
/// --serve, multi-GPU layer split: "K" or "K1,K2,.." (the first layer of each later stage) or "auto" (placed
/// from each GPU's free VRAM); empty = one GPU
std::string layer_split;