Refactors trace operations to be more self-contained, separates front…

…end/trace tensors more cleanly

- Refactors Trace operation so that it reports how many outputs it generates instead
     of requiring the caller to know.

     The trace op is now also responsible for creating its own output trace
     tensors. Additionally, `build`/`build_internal` have been removed, meaning
     the trace does *not* create frontend tensors anymore.

     Frontend tensors no longer create trace tensors directly but instead only
     interface with ops and wrap their outputs as needed.

- Consolidates and renames some frontend Tensor constructors to better reflect
     their purpose. For example, `create_directly` -> `fast_init`.

- Temporarily removes the "how to add ops" guide. A new version of this will be
     written once we have switched to the TRT dialect, which will signficantly affect
     how ops are added.

tripy/CONTRIBUTING.md

@@ -74,9 +74,6 @@ We've written developer guides to help you understand the codebase:
     [architecture](https://nvidia.github.io/TensorRT-Incubator/post0_developer_guides/architecture.html)
     documentation.
 
-- If you need to add a new operation, refer to
-    [this guide](https://nvidia.github.io/TensorRT-Incubator/post0_developer_guides/how-to-add-new-ops.html).
-
 
 ### Tests
 

tripy/notebooks/resnet50.ipynb

@@ -23,7 +23,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "!python3 -m pip install nvtripy -f https://nvidia.github.io/TensorRT-Incubator/packages.html"
+    "%pip install nvtripy -f https://nvidia.github.io/TensorRT-Incubator/packages.html"
    ]
   },
   {
@@ -39,7 +39,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "!pip install \"datasets==2.18.0\" \"matplotlib>=3.9.1\" \"pillow>=9.4.0\" \"transformers==4.46.2\" \"torch>=2.3.1\""
+    "%pip install \"datasets==2.18.0\" \"matplotlib>=3.9.1\" \"pillow>=9.4.0\" \"transformers==4.46.2\" \"torch>=2.3.1\""
    ]
   },
   {

tripy/nvtripy/backend/api/executable.py

@@ -229,7 +229,7 @@ def add(a, b):
 
             raise
 
-        output_tensors = [Tensor.create_directly(output, fetch_stack_info=False) for output in executor_outputs]
+        output_tensors = [Tensor.fast_init(output) for output in executor_outputs]
         if len(output_tensors) == 1:
             output_tensors = output_tensors[0]
         return output_tensors

tripy/nvtripy/backend/mlir/executor.py

@@ -91,7 +91,7 @@ def _get_output_tensor_info(self, outputs_runtime_shape, output_devices):
 
             output_device = output_devices[index]
             if not output_device:
-                output_device = device.create_directly(
+                output_device = device.fast_init(
                     "gpu" if memref.address_space == runtime.PointerType.device else "cpu", 0
                 )