Require at least one tensor to be marked dynamic with --dynamic-batch-only (#99620)

Signed-off-by: Edward Z. Yang <ezyang@meta.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/99620
Approved by: https://github.com/voznesenskym

benchmarks/dynamo/common.py

@@ -2412,14 +2412,19 @@ def run(runner, args, original_dir=None):
             # NB: This must be done late enough so that we don't do more
             # conversions on the inputs
             # NB: Assumes only the first batch-y like dimension is the batch
+            marked = False
+
             def detect_and_mark_batch(t):
+                nonlocal marked
                 for i, s in enumerate(t.size()):
                     if s == batch_size:
                         torch._dynamo.mark_dynamic(t, i)
+                        marked = True
                         break
 
             if args.dynamic_batch_only:
                 tree_map_only(torch.Tensor, detect_and_mark_batch, example_inputs)
+                assert marked, f"nothing in example_inputs had a dim with {batch_size}"
 
             if args.log_operator_inputs:
                 log_operator_inputs(

benchmarks/dynamo/torchbench.py

@@ -282,6 +282,10 @@ class TorchBenchmarkRunner(BenchmarkRunner):
         if self.args.accuracy and model_name in MAX_BATCH_SIZE_FOR_ACCURACY_CHECK:
             batch_size = min(batch_size, MAX_BATCH_SIZE_FOR_ACCURACY_CHECK[model_name])
 
+        # See https://github.com/pytorch/benchmark/issues/1560
+        if model_name == "speech_transformer":
+            batch_size = 10
+
         # workaround "RuntimeError: not allowed to set torch.backends.cudnn flags"
         torch.backends.__allow_nonbracketed_mutation_flag = True
         extra_args = []
@@ -317,6 +321,10 @@ class TorchBenchmarkRunner(BenchmarkRunner):
         # the right example_inputs
         if model_name == "yolov3":
             example_inputs = (torch.rand(batch_size, 3, 384, 512).to(device),)
+        # See https://github.com/pytorch/benchmark/issues/1561
+        if model_name == "maml_omniglot":
+            batch_size = 5
+            assert example_inputs[0].shape[0] == batch_size
         # global current_name, current_device
         # current_device = device
         # current_name = benchmark.name