Benchmark: Update overlap and sharding matmul benchmarks (#19)

- Enable `computation-communication-overlap` and `sharding-matmul` in
some configs through the existing PyTorch distributed mode.
- Use `torchrun --standalone` for single-node `torch.distributed` runs
to avoid fixed rendezvous port conflicts on 29500.
- Update runner command-generation test expectation for the new
single-node torchrun behavior.

superbench/config/hygon_bw1000.yaml

@@ -535,3 +535,7 @@ superbench:
           - mobilenet_v2
         precision: float16
         batch_size: 1
+    computation-communication-overlap:
+      <<: *default_pytorch_mode
+    sharding-matmul:
+      <<: *default_pytorch_mode

superbench/config/hygon_bw150.yaml

@@ -424,3 +424,7 @@ superbench:
           - mobilenet_v2
         precision: float16
         batch_size: 1
+    computation-communication-overlap:
+      <<: *default_pytorch_mode
+    sharding-matmul:
+      <<: *default_pytorch_mode

superbench/runner/runner.py

@@ -176,7 +176,7 @@ def __get_mode_command(self, benchmark_name, mode, timeout=None):
             # TODO: replace with torch.distributed.run in v1.9
             # TODO: only supports node_num=1 and node_num=all currently
             torch_dist_params = (
-                '' if 'node_num' in mode and mode.node_num == 1 else
+                '--standalone ' if 'node_num' in mode and mode.node_num == 1 else
                 '--nnodes=$NNODES --node_rank=$NODE_RANK --master_addr=$MASTER_ADDR --master_port=$MASTER_PORT '
             )
 

tests/runner/test_runner.py

@@ -142,7 +142,7 @@ def test_get_mode_command(self):
                 },
                 'expected_command': (
                     'torchrun '
-                    '--no_python --nproc_per_node=8 '
+                    '--no_python --nproc_per_node=8 --standalone '
                     f'sb exec --output-dir {self.sb_output_dir} -c sb.config.yaml -C superbench.enable=foo '
                     'superbench.benchmarks.foo.parameters.distributed_impl=ddp '
                     'superbench.benchmarks.foo.parameters.distributed_backend=nccl'