Fix bug on NCCL reduce. Need to change size to 2x for complex128 dtype

src/qibotn/eval.py

@@ -219,11 +219,20 @@ def reduce_result(result, comm, method="MPI", root=0):
         return comm.reduce(sendobj=result, op=MPI.SUM, root=root)
     elif method == "NCCL":
         stream_ptr = cp.cuda.get_current_stream().ptr
+        if result.dtype == cp.complex128:
+            count = result.size * 2   # complex128 has 2 float64 numbers
+            nccl_type = nccl.NCCL_FLOAT64
+        elif result.dtype == cp.complex64:
+            count = result.size * 2   # complex64 has 2 float32 numbers
+            nccl_type = nccl.NCCL_FLOAT32
+        else:
+            raise TypeError(f"Unsupported dtype for NCCL reduce: {result.dtype}")
+
         comm.reduce(
             result.data.ptr,
             result.data.ptr,
-            result.size,
-            nccl.NCCL_FLOAT64,
+            count,
+            nccl_type,
             nccl.NCCL_SUM,
             root,
             stream_ptr,