[FX] refactor the fx path in compile function (#1141)

* compile interface

* add compile method

* update

* update

* Update lower_setting.py

* update fx2trt_example

* add docstring

* update dynamic_batch default to False

* add docstring

* add save/load module

Author: Wei

py/torch_tensorrt/_compile.py

@@ -3,9 +3,11 @@
 import torch_tensorrt.ts
 from torch_tensorrt import logging
 import torch
-from torch import fx
+import torch.fx
 from enum import Enum
-from torch_tensorrt import fx
+import torch_tensorrt.fx
+from torch_tensorrt.fx.lower import lower_to_trt
+from torch_tensorrt.fx.utils import LowerPrecision
 
 class _IRType(Enum):
     """Enum to set the minimum required logging level to print a message to stdout
@@ -108,78 +110,14 @@ def compile(module: Any, ir="default", inputs=[], enabled_precisions=set([_enums
             ts_mod = torch.jit.script(module)
         return torch_tensorrt.ts.compile(ts_mod, inputs=inputs, enabled_precisions=enabled_precisions, **kwargs)
     elif target_ir == _IRType.fx:
-        from torch_tensorrt.fx.tracer.acc_tracer import acc_tracer
-        from torch_tensorrt.fx import InputTensorSpec
-        from torch_tensorrt.fx import TRTInterpreter
-        from torch_tensorrt.fx.passes.lower_basic_pass import transform_setitem
-        from torch_tensorrt.fx.tools.trt_splitter import TRTSplitter
-        from torch_tensorrt.fx.tools.trt_splitter import TRTSplitterSetting
-        from torch_tensorrt.fx.trt_module import TRTModule
-        from torch_tensorrt.fx.utils import LowerPrecision
-        acc_model = acc_tracer.trace(module, inputs)
-
-        splitter_setting = TRTSplitterSetting()
-        splitter_setting.use_implicit_batch_dim = False
-        splitter = TRTSplitter(acc_model, inputs, settings=splitter_setting)
-        splitter.node_support_preview()
-        split_mod = splitter()
-        num_piece = 0
-        for name, _ in split_mod.named_children():
-            print(f"graph is split into {name}")
-            num_piece += 1
-
-        # if the graph module is split into pieces larger than 8, we consider its perf
-        # is not good and fall back to non-TRT
-        if num_piece > 8:
-            print(
-                f"The graph module is split into {num_piece} which is large than the \
-                threshold=8. Fall back to non-TRT module."
-            )
-            return None
-
-        if torch.float16 in enabled_precisions or torch.half in enabled_precisions:
-            precision = LowerPrecision.FP16
+        if torch.float16 in enabled_precisions or torch_tensorrt.dtype.half in enabled_precisions:
+            lower_precision = LowerPrecision.FP16
+        elif torch.float32 in enabled_precisions or torch_tensorrt.dtype.float in enabled_precisions:
+            lower_precision = LowerPrecision.FP32
         else:
-            precision = LowerPrecision.FP32
-
-        def get_submod_inputs(mod, submod, inputs):
-            acc_inputs = None
-
-            def get_input(self, inputs):
-                nonlocal acc_inputs
-                acc_inputs = inputs
-
-            handle = submod.register_forward_pre_hook(get_input)
-            mod(*inputs)
-            handle.remove()
-            return acc_inputs
-
-        for name, _ in split_mod.named_children():
-            if "_run_on_acc" in name:
-                submod = getattr(split_mod, name)
-                # Get submodule inputs for fx2trt
-                acc_inputs = get_submod_inputs(split_mod, submod, inputs)
-
-                # fx2trt replacement
-                interp = TRTInterpreter(
-                    submod,
-                    InputTensorSpec.from_tensors(acc_inputs),
-                    explicit_batch_dimension=True,
-                )
-                r = interp.run(
-                    max_workspace_size=20 << 30,
-                    lower_precision=precision,
-                    # profiling_verbosity=trt.ProfilingVerbosity.DETAILED, #For profile
-                )
-                # For profile
-                # from torch_tensorrt.fx.tools.trt_profiler_sorted import profile_trt_module
-                # profile_trt_module("", trt_mod, acc_inputs)
-                trt_mod = TRTModule(*r)
-
-                setattr(split_mod, name, trt_mod)
-            else:
-                submod = getattr(split_mod, name)
-        return split_mod
+            raise ValueError(f"Precision {enabled_precisions} not supported on FX")
+
+        return lower_to_trt(module, inputs, lower_precision=lower_precision, max_batch_size=inputs[0].size(0), explicit_batch_dimension=True)
     else:
         raise RuntimeError("Module is an unknown format or the ir requested is unknown")
 

py/torch_tensorrt/fx/example/fx2trt_example.py

@@ -3,11 +3,11 @@
 import torch
 import torch.fx
 import torch.nn as nn
+from torch_tensorrt.fx.utils import LowerPrecision
 import torch_tensorrt.fx.tracer.acc_tracer.acc_tracer as acc_tracer
 from torch_tensorrt.fx import InputTensorSpec, TRTInterpreter, TRTModule
 from torch_tensorrt.fx.tools.trt_splitter import TRTSplitter
 
-
 # The purpose of this example is to demonstrate the overall flow of lowering a PyTorch
 # model to TensorRT via FX with existing FX based tooling. The general lowering flow
 # would be like:
@@ -30,11 +30,12 @@ def forward(self, x):
         x = self.linear(x)
         x = self.relu(x)
         x = torch.linalg.norm(x, ord=2, dim=1)
+        x = self.relu(x)
         return x
 
 
-inputs = [torch.randn(1, 10)]
-model = Model().eval()
+inputs = [torch.randn((1, 10), device=torch.device('cuda'))]
+model = Model().cuda().eval()
 
 # acc_tracer is a custom fx tracer that maps nodes whose targets are PyTorch operators
 # to acc ops.
@@ -64,20 +65,23 @@ def forward(self, x):
 # Split.
 split_mod = splitter()
 
-# After split we have two submodules, _run_on_acc_0 and _run_on_gpu_1.
+# After split we have three submodules, _run_on_acc_0 and _run_on_gpu_1.
 print(split_mod.graph)
 """
 graph():
     %x : [#users=1] = placeholder[target=x]
     %_run_on_acc_0 : [#users=1] = call_module[target=_run_on_acc_0](args = (%x,), kwargs = {})
     %_run_on_gpu_1 : [#users=1] = call_module[target=_run_on_gpu_1](args = (%_run_on_acc_0,), kwargs = {})
-    return _run_on_gpu_1
+    %_run_on_acc_2 : [#users=1] = call_module[target=_run_on_acc_2](args = (%_run_on_gpu_1,), kwargs = {})
+    return _run_on_acc_2
 """
 
 # Take a look at what inside each submodule. _run_on_acc_0 contains linear and relu while
-# _run_on_gpu_1 contains linalg_norm which currently is not supported by fx2trt.
+# _run_on_gpu_1 contains linalg_norm which currently is not supported by fx2trt. _run_on_acc_3
+# is the another submodule supported.
 print(split_mod._run_on_acc_0.graph)
 print(split_mod._run_on_gpu_1.graph)
+print(split_mod._run_on_acc_2.graph)
 """
 graph():
     %x : [#users=1] = placeholder[target=x]
@@ -90,32 +94,51 @@ def forward(self, x):
     %relu_1 : [#users=1] = placeholder[target=relu_1]
     %linalg_norm_1 : [#users=1] = call_function[target=torch_tensorrt.fx.tracer.acc_tracer.acc_ops.linalg_norm](args = (), ...
     return linalg_norm_1
+graph():
+    %linalg_norm_1 : [#users=1] = placeholder[target=linalg_norm_1]
+    %relu_3 : [#users=1] = call_function[target=torch_tensorrt.fx.tracer.acc_tracer.acc_ops.relu](args = (), kwargs = {input: %linalg_norm_1, inplace: False})
+    return relu_3
 """
 
-# Now let's lower split_mod._run_on_acc_0. If we know the model can be fully lowered,
-# we can skip the splitter part.
-interp = TRTInterpreter(split_mod._run_on_acc_0, InputTensorSpec.from_tensors(inputs))
-r = interp.run()
-trt_mod = TRTModule(r.engine, r.input_names, r.output_names)
-split_mod._run_on_acc_0 = trt_mod
-
-cuda_inputs = [input.cuda() for input in inputs]
-split_mod.cuda()
-lowered_model_output = split_mod(*cuda_inputs)
+def get_submod_inputs(mod, submod, inputs):
+    acc_inputs = None
+
+    def get_input(self, inputs):
+        nonlocal acc_inputs
+        acc_inputs = inputs
+
+    handle = submod.register_forward_pre_hook(get_input)
+    mod(*inputs)
+    handle.remove()
+    return acc_inputs
+
+# Since the model is splitted into three segments. We need to lower each TRT eligible segment.
+# If we know the model can be fully lowered, we can skip the splitter part.
+for name, _ in split_mod.named_children():
+    if "_run_on_acc" in name:
+        submod = getattr(split_mod, name)
+        # Get submodule inputs for fx2trt
+        acc_inputs = get_submod_inputs(split_mod, submod, inputs)
+
+        # fx2trt replacement
+        interp = TRTInterpreter(
+            submod,
+            InputTensorSpec.from_tensors(acc_inputs),
+            explicit_batch_dimension=True,
+        )
+        r = interp.run(lower_precision=LowerPrecision.FP32)
+        trt_mod = TRTModule(*r)
+        setattr(split_mod, name, trt_mod)
+
+lowered_model_output = split_mod(*inputs)
+
+# Save and load model
+torch.save(split_mod, "trt.pt")
+reload_trt_mod = torch.load("trt.pt")
+reload_model_output = reload_trt_mod(*inputs)
 
 # Make sure the results match
-model.cuda()
-regular_model_output = model(*cuda_inputs)
+regular_model_output = model(*inputs)
 torch.testing.assert_close(
-    lowered_model_output, regular_model_output.to(torch.float16), atol=3e-3, rtol=1e-2
+    reload_model_output, regular_model_output, atol=3e-3, rtol=1e-2
 )
-
-# We can utilize the trt profiler to print out the time spend on each layer.
-trt_mod.enable_profiling()
-trt_mod(*cuda_inputs)
-"""
-Reformatting CopyNode for Input Tensor 0 to LayerType.FULLY_CONNECTED_acc_ops.linear_linear_1: 0.027392ms
-LayerType.FULLY_CONNECTED_acc_ops.linear_linear_1: 0.023072ms
-PWN(ActivationType.RELU_acc_ops.relu_relu_1): 0.008928ms
-"""
-trt_mod.disable_profiling()

py/torch_tensorrt/fx/example/lower_example.py

@@ -198,6 +198,6 @@ def run_configuration_benchmark(
 
 
 if __name__ == "__main__":
-    test_model = torchvision.models.resnet101()
-    input = [torch.cuda.FloatTensor(1024, 3, 224, 224)]  # type: ignore[attr-defined]
-    benchmark(test_model, input, 100, 1024)
+    test_model = torchvision.models.resnet18(pretrained=True)
+    input = [torch.rand(128, 3, 224, 224)]  # type: ignore[attr-defined]
+    benchmark(test_model, input, 50, 128)

py/torch_tensorrt/fx/example/test_fx2trt.py

@@ -0,0 +1,57 @@
+import torch
+import copy
+import torchvision
+import torch_tensorrt
+from torch_tensorrt.fx import InputTensorSpec
+
+
+def test_torch_tensorrt(model, inputs):
+    # torchscript path
+    model_ts = copy.deepcopy(model)
+    inputs_ts = copy.deepcopy(inputs)
+    # fp32 test
+    with torch.inference_mode():
+        ref_fp32 = model_ts(*inputs_ts)
+    trt_ts_module = torch_tensorrt.compile(
+        model_ts, inputs=inputs_ts, enabled_precisions={torch.float32}
+    )
+    result_fp32 = trt_ts_module(*inputs_ts)
+    assert(torch.nn.functional.cosine_similarity(ref_fp32.flatten(), result_fp32.flatten(), dim=0)>0.9999)
+    # fp16 test
+    model_ts = model_ts.half()
+    inputs_ts = [i.cuda().half() for i in inputs_ts]
+    with torch.inference_mode():
+        ref_fp16 = model_ts(*inputs_ts)
+    trt_ts_module = torch_tensorrt.compile(
+        model_ts, inputs=inputs_ts, enabled_precisions={torch.float16}
+    )
+    result_fp16 = trt_ts_module(*inputs_ts)
+    assert(torch.nn.functional.cosine_similarity(ref_fp16.flatten(), result_fp16.flatten(), dim=0)>0.99)
+
+    # FX path
+    model_fx = copy.deepcopy(model)
+    inputs_fx = copy.deepcopy(inputs)
+    # fp32 test
+    with torch.inference_mode():
+        ref_fp32 = model_fx(*inputs_fx)
+    trt_fx_module = torch_tensorrt.compile(
+        model_fx, ir="fx", inputs=inputs_fx, enabled_precisions={torch.float32}
+    )
+    result_fp32 = trt_fx_module(*inputs_fx)
+    assert(torch.nn.functional.cosine_similarity(ref_fp32.flatten(), result_fp32.flatten(), dim=0)>0.9999)
+    # fp16 test
+    model_fx = model_fx.cuda().half()
+    inputs_fx = [i.cuda().half() for i in inputs_fx]
+    with torch.inference_mode():
+        ref_fp16 = model_fx(*inputs_fx)
+    trt_fx_module = torch_tensorrt.compile(
+        model_fx, ir="fx", inputs=inputs_fx, enabled_precisions={torch.float16}
+    )
+    result_fp16 = trt_fx_module(*inputs_fx)
+    assert(torch.nn.functional.cosine_similarity(ref_fp16.flatten(), result_fp16.flatten(), dim=0)>0.99 )
+
+
+if __name__ == "__main__":
+    model = torchvision.models.resnet18(pretrained=True).cuda().eval()
+    inputs = [torch.ones((32, 3, 224, 224), device=torch.device('cuda'))]  # type: ignore[attr-defined]
+    test_torch_tensorrt(model, inputs)

py/torch_tensorrt/fx/example/torch_trt_simple_example.py

@@ -164,6 +164,21 @@ def run(
         timing_cache=None,
         profiling_verbosity=None,
     ) -> TRTInterpreterResult:
+        """
+        Build TensorRT engine with some configs.
+        Args:
+            max_batch_size: set accordingly for maximum batch size you will use.
+            max_workspace_size: set to the maximum size we can afford for temporary buffer
+            lower_precision: the precision model layers are running on (TensorRT will choose the best perforamnce precision).
+            sparse_weights: allow the builder to examine weights and use optimized functions when weights have suitable sparsity
+            force_fp32_output: force output to be fp32
+            strict_type_constraints: Usually we should set it to False unless we want to control the precision of certain layer for numeric reasons.
+            algorithm_selector: set up algorithm selection for certain layer
+            timing_cache: enable timing cache for TensorRT
+            profiling_verbosity: TensorRT logging level
+        Return:
+            TRTInterpreterResult
+        """
         TRT_INTERPRETER_CALL_PRE_OBSERVER.observe(self.module)
 
         # For float outputs, we set their dtype to fp16 only if lower_precision == LowerPrecision.FP16 and