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: kEntMax declared before Ent

Niko1221b43b0bb518cd
diff --git a/src/dsv41/ds4_model.cu b/src/dsv41/ds4_model.cu
index 07219aaf9..91278f20e 100644
--- a/src/dsv41/ds4_model.cu
+++ b/src/dsv41/ds4_model.cu
@@ -4848,13 +4848,13 @@ struct Model::Impl {
// in four rounds (the gate / up split-K pieces, each preparing its two input blocks of every token; each band's
// sums / codebook / Hadamard / swiglu and that block of the down input, which both nodes then read whole; the down
// pieces; the down bands). yd[slot] gets each row's expert output.
+ static constexpr int kEntMax = kWin > kc::kExl3CpuMaxN ? kWin : kc::kExl3CpuMaxN;
struct Ent {
int e = 0, n = 0, slot0 = 0;
int tok[kEntMax] = {};
};
// tokens an entry holds: the i-quant GGUF experts take a whole window's tokens of an expert in one call (the CPU
// lane's exact multi-token rows, 7+ tokens an expert: decoded once for all), the others (MXFP4, EXL3) up to 8
- static constexpr int kEntMax = kWin > kc::kExl3CpuMaxN ? kWin : kc::kExl3CpuMaxN;
static int ent_cap(const Layer& l) {
return l.lay.fmt == gg::kFmtMXFP4 && !gg::mx_experts(l.lay) ? kEntMax : kc::kExl3CpuMaxN;
}