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++

fused_gr: the persistent hyper-connection read's barrier only calls __nanosleep on sm_70 and newer (the sm_60 build of f-027-int2 failed on it)

Niko1221f1dadd171056
diff --git a/src/kernels/cuda/fused_gr.cu b/src/kernels/cuda/fused_gr.cu
index 6f6f32321..fe8672989 100644
--- a/src/kernels/cuda/fused_gr.cu
+++ b/src/kernels/cuda/fused_gr.cu
@@ -2797,7 +2797,9 @@ __device__ __forceinline__ bool grid_sync(unsigned* bar, unsigned long long tick
while (vb[32] == gen) {
if (vb[48] != 0u) { ab = 1; break; }
if (clock64() - t0 > (long long) ticks) { atomicExch(&bar[48], 1u); ab = 1; break; }
- __nanosleep(64);
+#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 700
+ __nanosleep(64); // (sm_60 has no nanosleep: the barrier spins)
+#endif
}
}
if (vb[48] != 0u) ab = 1;