docs: New documentation on the to_backend api integration

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>

Author: narendasan

docsrc/index.rst

@@ -25,6 +25,7 @@ Getting Started
 * :ref:`getting_started`
 * :ref:`ptq`
 * :ref:`trtorchc`
+* :ref:`use_from_pytorch`
 
 
 .. toctree::
@@ -36,6 +37,7 @@ Getting Started
    tutorials/getting_started
    tutorials/ptq
    tutorials/trtorchc
+   tutorials/use_from_pytorch
    _notebooks/lenet
 
 .. toctree::

docsrc/py_api/trtorch.rst

@@ -17,9 +17,11 @@ Functions
 
 .. autofunction:: check_method_op_support
 
+.. autofunction:: get_build_info
+
 .. autofunction:: dump_build_info
 
-.. autofunction:: get_build_info
+.. autofunction:: TensorRTCompileSpec
 
 Enums
 -------

docsrc/tutorials/use_from_pytorch.rst

@@ -0,0 +1,62 @@
+.. _use_from_pytorch:
+
+Using TRTorch Directly From PyTorch
+====================================
+
+Starting in TRTorch 0.1.0, you will now be able to directly access TensorRT from PyTorch APIs. The process to use this feature
+is very similar to the compilation workflow described in :ref:`getting_started`
+
+Start by loading ``trtorch`` into your application.
+
+.. code-block:: python
+
+    import torch
+    import trtorch
+
+
+Then given a TorchScript module, you can lower it to TensorRT using the ``torch._C._jit_to_tensorrt`` API.
+
+.. code-block:: python
+
+    import torchvision.models as models
+
+    model = models.mobilenet_v2(pretrained=True)
+    script_model = torch.jit.script(model)
+
+Unlike the ``compile`` API in TRTorch which assumes you are trying to compile the ``forward`` function of a module
+or the ``convert_method_to_trt_engine`` which converts a specified function to a TensorRT engine, the backend API
+will take a dictionary which maps names of functions to compile to Compilation Spec objects which wrap the same
+sort of dictionary you would provide to ``compile``. For more information on the compile spec dictionary take a look
+at the documentation for the TRTorch ``TensorRTCompileSpec`` API.
+
+.. code-block:: python
+
+    spec = {
+        "forward": trtorch.TensorRTCompileSpec({
+            "input_shapes": [[1, 3, 300, 300]],
+            "op_precision": torch.half,
+            "refit": False,
+            "debug": False,
+            "strict_types": False,
+            "allow_gpu_fallback": True,
+            "device_type": "gpu",
+            "capability": trtorch.EngineCapability.default,
+            "num_min_timing_iters": 2,
+            "num_avg_timing_iters": 1,
+            "max_batch_size": 0,
+        })
+    }
+
+Now to compile with TRTorch, provide the target module objects and the spec dictionary to ``torch._C._jit_to_tensorrt``
+
+.. code-block:: python
+
+    trt_model = torch._C._jit_to_tensorrt(script_model._c, spec)
+
+To run explicitly call the function of the method you want to run (vs. how you can just call on the module itself in standard PyTorch)
+
+.. code-block:: python
+
+    input = torch.randn((1, 3, 300, 300).to("cuda").to(torch.half)
+    print(trt_model.forward(input))
+