Multi node multi gpu

src/qibotn/QiboCircuitConvertor.py

@@ -95,7 +95,8 @@ def init_intermediate_circuit(self, circuit):
             required_shape = self.op_shape_from_qubits(len(gate_qubits))
             self.gate_tensors.append(
                 (
-                    cp.asarray(gate.matrix).reshape(required_shape),
+                    cp.asarray(gate.matrix(), dtype=self.dtype).reshape(
+                        required_shape),
                     gate_qubits,
                 )
             )

src/qibotn/cutn.py

@@ -1,8 +1,40 @@
-# from qibotn import quimb as qiboquimb
 from qibotn.QiboCircuitConvertor import QiboCircuitToEinsum
 from cuquantum import contract
+from cuquantum import cutensornet as cutn
+from mpi4py import MPI  # this line initializes MPI
+import multiprocessing
+from cupy.cuda.runtime import getDeviceCount
+import cupy as cp
 
 
 def eval(qibo_circ, datatype):
     myconvertor = QiboCircuitToEinsum(qibo_circ, dtype=datatype)
     return contract(*myconvertor.state_vector_operands())
+
+
+def eval_tn_MPI(qibo_circ, datatype):
+    ncpu_threads = multiprocessing.cpu_count() // 2
+    n_samples = 8
+
+    comm = MPI.COMM_WORLD
+    rank = comm.Get_rank()
+    device_id = rank % getDeviceCount()
+    cp.cuda.Device(device_id).use()
+
+    handle = cutn.create()
+    cutn.distributed_reset_configuration(handle, *cutn.get_mpi_comm_pointer(comm))
+    network_opts = cutn.NetworkOptions(handle=handle, blocking="auto")
+
+    myconvertor = QiboCircuitToEinsum(qibo_circ, dtype=datatype)
+    operands_interleave = myconvertor.state_vector_operands()
+
+    network = cutn.Network(*operands_interleave, options=network_opts)
+    network.contract_path(
+        optimize={"samples": n_samples, "threads": ncpu_threads}
+    )  # Calculate optimal path, returns path and info
+
+    result = network.contract()
+
+    cutn.destroy(handle)
+
+    return result, rank