do not clear UncachedCompile between node runs

Author: haifeng-jin

torch_xla/_dynamo/dynamo_bridge.py

@@ -643,6 +643,7 @@ def run_node(self, n: torch.fx.Node):
     # We need to restore this metric count later, so save it in a separate variable
     dynamo_extract_graph_helper_metric_count = metrics.counter_value(
         'DynamoExtractCompiledGraph')
+    uncached_compile_metric_count = metrics.counter_value('UncachedCompile')
 
     metrics.clear_counters()
     result = super().run_node(n)
@@ -681,6 +682,8 @@ def all_tensors_on_xla_device(value):
     # Restore this metric counter
     torch_xla._XLAC._xla_increment_counter(
         'DynamoExtractCompiledGraph', dynamo_extract_graph_helper_metric_count)
+    torch_xla._XLAC._xla_increment_counter('UncachedCompile',
+                                           uncached_compile_metric_count)
 
     return result
 
@@ -725,6 +728,7 @@ def allow_cpu_device(self, node: torch.fx.Node):
 
 def extract_compiled_graph(xla_model: torch.fx.GraphModule, xla_args):
   torch_xla._XLAC._xla_increment_counter('DynamoExtractCompiledGraph', 1)
+  torch_xla._XLAC._xla_increment_counter('UncachedCompile', 1)
 
   with torch_xla.experimental.eager_mode_context(False):
     return extract_compiled_graph_helper(xla_model, xla_args)