intel版本，家里集群

docs/contest_runners.md

@@ -12,13 +12,32 @@ TORCH_THREADS=48 \
 SEARCH_REPEATS=2048 \
 SEARCH_TIME=300 \
 SCHEDULER_HOST=10.20.1.103 \
-WORKER_HOSTS="10.20.1.103 10.20.6.101" \
+WORKER_HOSTS="10.20.1.103 10.20.1.101" \
 DASK_ADDRESS="tcp://10.20.1.103:8786" \
 NWORKERS=84 \
 NTHREADS=1 \
 MPIEXEC_FULL="mpirun -np 4 -hostfile /home/yx/qibotn/hostfile -perhost 2" \
 tools/run_tn_dask_mpi_all.sh
 
+
+
+I_MPI_FABRICS=shm:ofi \
+I_MPI_OFI_PROVIDER=tcp \
+FI_PROVIDER=tcp \
+CASE=main1 \
+OBSERVABLES=long_z_string \
+NQUBITS=34 \
+NLAYERS=20 \
+TORCH_THREADS=48 \
+SEARCH_REPEATS=2048 \
+SEARCH_TIME=300 \
+SCHEDULER_HOST=127.0.0.1 \
+WORKER_HOSTS="127.0.0.1" \
+DASK_ADDRESS="tcp://127.0.0.1:8786" \
+NWORKERS=12 \
+NTHREADS=1 \
+MPIEXEC_FULL="mpirun -np 2 -hostfile /home/yx/qibotn/hostfile -perhost 2" \
+tools/run_tn_dask_mpi_all.sh
 # 单独缩并contract计算
 
 I_MPI_FABRICS=shm:ofi \
@@ -43,11 +62,16 @@ cd /home/yx/qibotn
 I_MPI_FABRICS=shm:ofi \
 I_MPI_OFI_PROVIDER=tcp \
 FI_PROVIDER=tcp \
-MPIEXEC_FULL="mpirun -np 4 -hostfile /home/yx/qibotn/hostfile -perhost 2" \
+MPIEXEC_FULL="mpirun -np 2 -hostfile /home/yx/qibotn/hostfile -perhost 2" \
 TORCH_THREADS=48 \
 OBS_FILTER=ring_xz \
 MAIN1_NQ=128 \
 MAIN1_LAYERS=24 \
-MAIN1_BOND=1024 \
+MAIN1_BOND=512 \
 tools/run_vidal_mpi_contest_cases.sh main1
-```
+```
+
+
+ 136.997
+
+ 

hostfile

@@ -1,2 +1,2 @@
 10.20.1.103:2
-10.20.6.101:2
+10.20.1.101:2

src/qibotn/backends/qmatchatea.py

@@ -12,6 +12,7 @@ from qibo.config import raise_error
 from qmatchatea.utils import MPISettings
 
 from qibotn.backends.abstract import QibotnBackend
+from qibotn.circuit_fusion import fuse_qiskit_one_qubit_gates
 from qibotn.observables import check_observable
 from qibotn.result import TensorNetworkResult
 
@@ -287,6 +288,7 @@ class QMatchaTeaBackend(QibotnBackend, NumpyBackend):
         if not preprocess:
             if compile_circuit:
                 qiskit_circuit = qmatchatea.tensor_compiler(qiskit_circuit)
+            qiskit_circuit = fuse_qiskit_one_qubit_gates(qiskit_circuit)
             return qiskit_circuit
 
         # Transpile the circuit to adapt it to the linear structure of the MPS,
@@ -297,6 +299,7 @@ class QMatchaTeaBackend(QibotnBackend, NumpyBackend):
                 tensor_compiler=compile_circuit
             ),
         )
+        qiskit_circuit = fuse_qiskit_one_qubit_gates(qiskit_circuit)
         return qiskit_circuit
 
     def _qiboobs_to_qmatchaobs(self, hamiltonian, observable_name="custom_hamiltonian"):

src/qibotn/backends/vidal_mpi_segment.py

@@ -124,7 +124,7 @@ class SegmentVidalMPIExecutor:
         gates = circuit.queue
         if compile_circuit:
             gates = _route_non_adjacent_gates(gates, circuit.nqubits)
-            gates = _fuse_one_site_blocks(gates)
+        gates = _fuse_one_site_blocks(gates)
         for batch in _disjoint_batches(gates):
             if _is_two_qubit_batch(batch):
                 self._apply_two_site_batch(batch, timings)

src/qibotn/backends/vidal_tebd.py

@@ -194,7 +194,7 @@ class VidalTEBDExecutor:
         gates = circuit.queue
         if compile_circuit:
             gates = _route_non_adjacent_gates(gates, circuit.nqubits)
-            gates = _fuse_one_site_blocks(gates)
+        gates = _fuse_one_site_blocks(gates)
         for batch in _disjoint_batches(gates):
             for gate in batch:
                 self._apply_gate(gate)

src/qibotn/circuit_fusion.py

@@ -0,0 +1,68 @@
+"""Circuit-level gate fusion helpers."""
+
+from __future__ import annotations
+
+import numpy as np
+
+
+def fuse_qiskit_one_qubit_gates(circuit):
+    """Fuse consecutive unconditional one-qubit gates in a Qiskit circuit.
+
+    The fused gate is emitted as a Qiskit ``UnitaryGate``.  qmatchatea's Qiskit
+    executor applies generic gates through ``operation.to_matrix()``, so this
+    reduces Python gate dispatch without changing the physical circuit.
+    """
+
+    from qiskit import QuantumCircuit
+    from qiskit.circuit.library import UnitaryGate
+
+    fused = QuantumCircuit(*circuit.qregs, *circuit.cregs, name=circuit.name)
+    pending = {}
+
+    def qubit_index(qubit):
+        return circuit.find_bit(qubit).index
+
+    def flush_qubit(index):
+        matrix = pending.pop(index, None)
+        if matrix is not None:
+            fused.append(UnitaryGate(matrix, label="fused1q"), [fused.qubits[index]])
+
+    def flush_all():
+        for index in sorted(tuple(pending)):
+            flush_qubit(index)
+
+    for instruction in circuit.data:
+        operation = instruction.operation
+        qubits = instruction.qubits
+        clbits = instruction.clbits
+        condition = getattr(operation, "condition", None)
+
+        can_fuse = (
+            len(qubits) == 1
+            and len(clbits) == 0
+            and condition is None
+            and hasattr(operation, "to_matrix")
+        )
+        if can_fuse:
+            try:
+                matrix = np.asarray(operation.to_matrix(), dtype=np.complex128)
+            except Exception:
+                can_fuse = False
+
+        if can_fuse:
+            index = qubit_index(qubits[0])
+            if index in pending:
+                pending[index] = matrix @ pending[index]
+            else:
+                pending[index] = matrix
+            continue
+
+        touched = {qubit_index(qubit) for qubit in qubits}
+        for index in sorted(touched):
+            flush_qubit(index)
+        if len(qubits) != 1:
+            flush_all()
+        fused.append(operation, [fused.qubits[qubit_index(q)] for q in qubits], clbits)
+
+    flush_all()
+    return fused

tools/tn_contest_runner.py

@@ -47,7 +47,7 @@ CASES = {
     "main1": CaseSpec(
         circuit_kind="rxx_rzz_chain",
         observables=("ring_xz",),
-        nqubits=34,
+        nqubits=37,
         nlayers=20,
         seed=31001,
         target_slices=None,