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++

DeepSeek on arm64 Linux: the same two compile fixes as the Metal branch (amd_family() on a non-x86 CPU, iq256x stubs in portable.cpp)

Niko1221e3d8823e6c8e
diff --git a/src/dsv41/ds4_gg_cpu.cpp b/src/dsv41/ds4_gg_cpu.cpp
index 78853b12c..72891fc66 100644
--- a/src/dsv41/ds4_gg_cpu.cpp
+++ b/src/dsv41/ds4_gg_cpu.cpp
@@ -169,7 +169,7 @@ static int amd_family() {
if (!__get_cpuid(1, &a, &b, &c, &d)) return 0;
const unsigned eax = a;
#else
- const unsigned eax = 0; // not x86 (arm64, a DGX Spark): no AMD family, the base family below reads 0
+ const unsigned eax = 0; // (not an x86 host: no family)
#endif
const unsigned base = (eax >> 8) & 0xF;
return (int) (base == 0xF ? base + ((eax >> 20) & 0xFF) : base);
diff --git a/src/kernels/cpu/portable.cpp b/src/kernels/cpu/portable.cpp
index 09cc3271a..a49d358f2 100644
--- a/src/kernels/cpu/portable.cpp
+++ b/src/kernels/cpu/portable.cpp
@@ -121,6 +121,10 @@ bool iq512x_supported(int) noexcept { return false; }
void iq512x_rows(int, const uint8_t*, size_t, int, const void* const*, int, float* const*, int, int) {
unsupported("the AVX-512 exact i-quant kernel");
}
+bool iq256x_supported(int) noexcept { return false; } // (the DeepSeek CPU lane's multi-token rows: ggml-cpu's NEON dots instead)
+void iq256x_rows(int, const uint8_t*, size_t, int, const void* const*, int, float* const*, int, int) {
+ unsupported("the AVX2 multi-token i-quant rows");
+}
int iq256_variant() noexcept { return 0; }
int iq256_variants() noexcept { return 0; }
void iq256_gu_rows_v(int, int, const uint8_t*, size_t, size_t, int, const void* const*, int, float* const*, int, int) {