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++

Fix #3368: verify_q8 no longer calls the device-only __half_as_ushort from host code (CUDA 12.0/12.1)

Yudhanjaya Wijeratne17b93dcb7c94
diff --git a/src/qwen35/q36_engine.cu b/src/qwen35/q36_engine.cu
index b0d7e75cb..03660de40 100644
--- a/src/qwen35/q36_engine.cu
+++ b/src/qwen35/q36_engine.cu
@@ -1652,11 +1652,14 @@ struct Engine {
std::memcpy(t, u, sizeof(u));
}
unsigned ds_a, ds_b;
+ const __half t0h = __float2half(t[0]); // host side: __half_as_ushort is __device__ only
+ unsigned short t0bits;
+ std::memcpy(&t0bits, &t0h, sizeof(t0bits));
std::memcpy(&ds_a, &a[o], 4);
std::memcpy(&ds_b, &b[o], 4);
std::fprintf(stderr, "VDEBUG block %zu: ds new %08x old %08x, qs bytes %s, cpu butterfly sum %.9g (half %04x)\n",
blk, ds_a, ds_b, std::memcmp(&a[o + 4], &b[o + 4], 32) == 0 ? "equal" : "DIFFER", (double) t[0],
- (unsigned) __half_as_ushort(__float2half(t[0])));
+ (unsigned) t0bits);
break;
}
}