addressing review comments- making dynamic=None in torch.compile and changing the option description

Author: apbose

examples/distributed_inference/tensor_parallel_simple_example.py

@@ -31,7 +31,7 @@ def compile_tp_model(tp_model, backend):
 
     try:
         return torch.compile(
-            tp_model, backend=backend, options=compile_options, dynamic=False
+            tp_model, backend=backend, options=compile_options, dynamic=None
         )
     except RuntimeError as e:
         if (
@@ -43,7 +43,7 @@ def compile_tp_model(tp_model, backend):
             )
             compile_options["use_distributed_mode_trace"] = True
             return torch.compile(
-                tp_model, backend=backend, options=compile_options, dynamic=False
+                tp_model, backend=backend, options=compile_options, dynamic=None
             )
         else:
             logger.debug("The distributed model fails with the following error")

py/torch_tensorrt/dynamo/_settings.py

@@ -92,7 +92,7 @@ class CompilationSettings:
         enable_weight_streaming (bool): Enable weight streaming.
         enable_cross_compile_for_windows (bool): By default this is False means TensorRT engines can only be executed on the same platform where they were built.
             True will enable cross-platform compatibility which allows the engine to be built on Linux and run on Windows
-        use_distributed_mode_trace (bool): Using aot_autograd to trace the graph. Enable this only if the model includes distributed operations
+        use_distributed_mode_trace (bool): Using aot_autograd to trace the graph. This is enabled when DTensors are present in distributed model
     """
 
     enabled_precisions: Set[dtype] = field(default_factory=lambda: ENABLED_PRECISIONS)

tests/py/dynamo/distributed/test_distributed_simple_example.py

@@ -27,7 +27,7 @@ def compile_tp_model(tp_model, backend):
 
     try:
         return torch.compile(
-            tp_model, backend=backend, options=compile_options, dynamic=False
+            tp_model, backend=backend, options=compile_options, dynamic=None
         )
     except RuntimeError as e:
         if (
@@ -39,7 +39,7 @@ def compile_tp_model(tp_model, backend):
             )
             compile_options["use_distributed_mode_trace"] = True
             return torch.compile(
-                tp_model, backend=backend, options=compile_options, dynamic=False
+                tp_model, backend=backend, options=compile_options, dynamic=None
             )
         else:
             logger.debug("The distributed model fails with the following error")

tests/py/dynamo/distributed/test_nccl_ops.sh

@@ -88,7 +88,6 @@ fi
 URL="https://pypi.nvidia.com/tensorrt-llm/$FILE"
 echo "Downloading $FILE from $URL..."
 
-echo "Downloading here...."
 #Installing wget
 ensure_installed wget