Update on "Fix Half X86_F16 CUDA build failure"

swolchok · swolchok · commit 2bc42d0f9fd4 · 2024-11-19T07:52:42.000-08:00
It passed PyTorch CI, but internally we saw failures from this. Differential Revision: [D66137897](https://our.internmc.facebook.com/intern/diff/D66137897/) [ghstack-poisoned]
diff --git a/c10/util/Half.h b/c10/util/Half.h
@@ -349,6 +349,10 @@ inline uint16_t fp16_ieee_from_fp32_value(float f) {
 #endif // C10_X86_F16
 }
 
+#ifdef C10_X86_F16
+#undef C10_X86_F16
+#endif // C10_X86_F16
+
 #if defined(__aarch64__) && !defined(__CUDACC__)
 inline float16_t fp16_from_bits(uint16_t h) {
   return c10::bit_cast<float16_t>(h);
