FX converter doc

docsrc/contributors/fx_converters.rst

@@ -0,0 +1,190 @@
+.. _conversion:
+
+FX Converters
+==================
+The converter library in Torch-TensorRT is located in ``TensorRT/py/torch_tensorrt/fx/converters``.
+They are categorized into - ``aten_ops_converters``, ``acc_ops_converters`` and ``nn_ops_converters``.
+The individual converters present are useful for the quantization workflow.
+The converters are registered using the ``tensorrt_converter``
+
+Steps
+==================
+
+Operation Sets
+-------------------
+Depending on whether the operation is generated using acc_trace, aten_trace or fx_trace, the converters are registered in one of the three operation sets
+``aten_ops_converters``, ``acc_ops_converters`` or ``nn_ops_converters``.  The converters are registered using ``tensorrt_converter`` decorator. The function decorated
+has the arguments - ``network, target, args, kwargs, name``,  which is common across all the operators schema.
+These functions are mapped in the ``aten`` converter registry dictionary, with key as the function target name.
+
+    * acc_ops_converters
+        *  acc_trace is produced by ``torch_tensorrt.fx.tracer.acc_tracer.acc_tracer.trace``.
+    * aten_ops
+        There are two options at present for this
+            #.   Dynamo: aten_trace is produced by ``torch_tensorrt.dynamo.backend.compile``. The second round of trace is produced by ``aot_torch_tensorrt_aten_backend`` by invoking ``aot_module_simplified`` from ``torch._functorch.aot_autograd``
+            #.   FX: aten_trace is produced by ``torch_tensorrt.fx.tracer.dispatch_tracer.aten_tracer.trace``. This flow is more common currently, but this will soon be deprecated in torch_tensorrt.
+    * nn_ops
+        *  symbolic_trace is produced by ``torch.fx._symbolic_trace``.
+
+The implementation of the above converter set is to be included in the common implementation library present in ``TensorRT/py/torch_tensorrt/fx/impl`` 
+This documentation focuses on the implementation of the ``aten_ops`` converters. There might be some steps The steps are discussed in more detail in the next section.
+
+Converter implementation
+------------------------
+In this section, we illustrate the steps to be implemented for writing a converter. We divide them according to activation, operator or lowering pass implementation.
+Each of them is detailed with the help of an example
+
+    * Registration
+
+        The converter needs to be registered with the appropriate op code in the ``tensorrt_converter``.
+
+        * Activation
+
+            Example: ``leaky_relu``
+
+            * acc_ops_converters
+
+                Define in ``py/torch_tensorrt/fx/converters/acc_ops_converters``. One needs to register the opcode generated in the trace with ``tensorrt_converter`` decorator. Op code to be used for the registration or the converter registry key in this case is ``acc_ops.leaky_relu``
+
+                .. code-block:: python
+
+                    @tensorrt_converter(acc_ops.leaky_relu)
+                    def acc_ops_leaky_relu(
+                        network: TRTNetwork,
+                        target: Target,
+                        args: Tuple[Argument, ...],
+                        kwargs: Dict[str, Argument],
+                        name: str,
+                    ) -> Union[TRTTensor, Sequence[TRTTensor]]:
+                        input_val = kwargs["input"]
+                        negative_slope = kwargs["negative_slope"]
+                        operation_type = trt.ActivationType.LEAKY_RELU
+                        return activation.leaky_relu(
+                        network, target, SourceIR.ACC, name, kwargs["input"], kwargs["negative_slope"]
+                    )
+            
+            * aten_ops_converters
+
+                Define in ``py/torch_tensorrt/fx/converters/aten_ops_converters``. One needs to register the opcode generated in the trace with ``tensorrt_converter`` decorator. Op code to be used for the registration or the converter registry key in this case is ``torch.ops.aten.leaky_relu.default``
+
+                .. code-block:: python
+            
+                    @tensorrt_converter(torch.ops.aten.leaky_relu.default)
+                    def aten_ops_leaky_relu(
+                        network: TRTNetwork,
+                        target: Target,
+                        args: Tuple[Argument, ...],
+                        kwargs: Dict[str, Argument],
+                        name: str,
+                    ) -> Union[TRTTensor, Sequence[TRTTensor]]:
+                        return activation.leaky_relu(network, target, SourceIR.ATEN, name, args[0], args[1])
+
+            The function decorated by ``tensorrt_converter`` has the following arguments which are automatically generated by the trace functions mentioned above.
+
+            #. network : Node in the form of ``call_module`` or ``call_function`` having the target as the key
+            #. target: Target key in the ``call_module`` or ``call_function`` above. eg: ``torch.ops.aten_.leaky_relu.default``
+            #. args: The arguments passed in the ``call_module`` or ``call_function`` above
+            #. kwargs: The kwargs passed in the ``call_module`` or ``call_function`` above
+            #. name: String containing the name of the target
+
+            As a user writing new converters, one just needs to take care that the approriate arguments are extracted from the trace generated to the implementation function in the implementation lib function ``activation.leaky_relu`` (which we will discuss below in detail). As one can see in the example above, the trace for ``acc_op`` and ``aten_op`` is different.
+            ``Acc_ops`` has arguments in the ``args`` whereas ``aten_ops`` has arguments in the ``kwargs`` in the trace.
+
+
+        * Operation type
+
+            Example: ``fmod``
+
+            It follows the same steps as the above converter. In this case the opcode is ``torch.ops.aten.fmod.Scalar`` or ``torch.ops.aten.fmod.Tensor``. 
+            Hence both the opcodes are registered in ``py/torch_tensorrt/fx/converters/aten_ops_converters``.  The opcode is ``acc_ops.fmod`` in ``py/torch_tensorrt/fx/converters/acc_ops_converters``.
+
+
+    * Implementation Library
+
+        The converters across all the above three opsets have the common implementation library ``py/torch_tensorrt/fx/converters/impl``
+
+        * Activation
+
+            Example: ``leaky_relu``
+
+            The implementation is to be placed in present in ``py/torch_tensorrt/fx/impl/activation.py``. This is where all the activation functions are defined and implemented.
+
+            .. code-block:: python
+
+                def leaky_relu(
+                    network: TRTNetwork,
+                    target: Target,
+                    source_ir: Optional[SourceIR],
+                    name: str,
+                    input_val: TRTTensor,
+                    alpha: Optional[Any],
+                ):
+                    #implementation
+
+            The implementation function has the following arguments.
+
+            #. network : ``network`` passed from the decorated function registration
+            #. target: ``target`` passed from the decorated function registration
+            #. source_ir: Enum attribute. ``SourceIR`` enum is defined in ``py/torch_tensorrt/fx/converters/impl/converter_utils``
+            #. name: ``name`` passed from the decorated function registration
+            #. input_val: Approriate arguments extracted from the decorated function registration from args or kwargs
+            #. alpha: Approriate arguments extracted from the decorated function registration from args or kwargs. If not None, it will set the alpha attribute of the created TensorRT activation layer eg: Used in leaky_relu, elu, hardtanh           
+            #. beta: Approriate arguments extracted from the decorated function registration from args or kwargs. If not None, it will set the beta attribute of the created TensorRT activation layer eg: Used in hardtanh
+            #. dyn_range_fn: A optional function which takes the dynamic range of a TensorRT Tensor and returns the output dynamic range
+
+            The implementation functions call the ``convert_activation`` function in ``py/torch_tensorrt/fx/impl/activation.py``. This function will add the approriate activation layer via ``network.add_activation``.
+
+        * Operator
+
+            The implementation is to be placed in ``py/torch_tensorrt/fx/impl/elementwise/ops.py``. This is where all the elementwise functions are defined and implemented.
+            For a new operator, one should identify the category to which it belongs. Following are some examples
+
+            #. Elementwise operators like ``fmod`` is present in ``py/torch_tensorrt/fx/impl/elementwise``. The ``py/torch_tensorrt/fx/impl/elementwise/base`` contains base functions for elementwise operator.
+            #. Unary operators like ``sqrt`` will be present in ``py/torch_tensorrt/fx/impl/unary``. The ``py/torch_tensorrt/fx/impl/unary/base`` contains base functions for unary operator.
+            #. normalization operators like softmax, layer_norm, batch_norm will be present in ``py/torch_tensorrt/fx/impl/normalization``. Since there are no base operations common to all, there is no base file. But one can choose to implement a base file, if there are common functions across all normalization operations
+            #. Individual perators like slice, select, where, embedding will be present in ``py/torch_tensorrt/fx/impl/*.py``. They will have individual operator implementation with the same API structure as above but with different individual arguments
+
+            Please note that the above operators would have common functions to be implemented which should be placed in 
+            ``py/torch_tensorrt/fx/impl/converter_utils.py``
+
+
+    * Lowering type
+
+        There are some converters which can be decomposed into suboperations and need not have seperate converter registration.
+        Such converters can be implemented via ``lowering passes``
+
+        Example: ``addmm``
+
+        The decompositions are registered via ``register_decomposition`` in ``py/torch_tensorrt/dynamo/backend/lowering/_decompositions.py``
+        We define ``addmm_replacement`` and replace it with the torch ops, which will have their corresponding converters called.
+
+        .. code-block:: python
+
+            @register_decomposition(torch.ops.aten.addmm, registry=DECOMPOSITIONS)
+            def addmm_replacement(
+                input_: torch.Tensor, mat1: torch.Tensor, mat2: torch.Tensor, *, beta=1, alpha=1
+            ) -> torch.Tensor:
+                return torch.add(
+                    torch.mul(input_, beta), torch.mul(torch.matmul(mat1, mat2), alpha)
+                )
+
+
+       
+Tests
+----------------
+
+    * FX testing
+        Implement the fx tests in ``py/torch_tensorrt/fx/test/converters/aten_op/test_<operator_name>_aten.py``. Derive the test class from ``DispatchTestCase``, with parameterized testing to implement different test cases. Check for the following two conditions
+            #. Compare the results for ``dispatch_tracer.aten_trace`` and torch.
+            #. Test the ``expected_op``. You can find examples in the above tests. This op will be called by the model and needs to be specified in the test so that the test checks that the approriate converter is invoked
+
+        The tests should fail if any of the above two conditions fail
+
+    * Dynamo testing
+        Dynamo tests are present for the lowering ops in ``py/torch_tensorrt/dynamo/backend/test/test_decompositions.py``. The above converters will soon be ported to dynamo tests
+            #. Compare the results for ``fx.symbolic_trace `` and ``torch_tensorrt.dynamo.compile``
+            #. The tests also test for the ``expected_op`` and the ``unexpected_op``. 
+                * ``expected_op``: Operations the operations are lowered to. eg: ``mul`` and ``add`` for ``addmm``
+                * ``unexpected_op``: Original operation. eg: ``addmm`` for ``addmm``
+
+        The tests should fail if any of the above two conditions fail