Changes done internally at Facebook (#1172)

b18ad449f9cbcad8d9e13c74c7605dc4dcca53bc Jason Park <jasonjk@fb.com> AccOpGraph and pattern matcher from GraphModule
0888aecb9f5a6add2d855300f574592587ee8484 Jason Park <jasonjk@fb.com> Dependency checker
e935e26f2ace98e9ef251ad8328d6a07ddca828f Jason Park <jasonjk@fb.com> Sort topologically.
d6b5a5187cd22a7a7340be7bf021edbaf178b981 Jason Park <jasonjk@fb.com> Grouped op fusion for layer norm
c41054bec9cf617248d08363e15992bc24eb0ce0 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.eisum
9f9505f320c87e9069658349fcc6040d8fdc77a2 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.elu
fec2cab3bb76c22f0eb660d6dd9a4f3ae5523cfd Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.embedding
1a07ef201654adb752160e8e47e2f8369abe142a Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.eq
7f110836cd4e36c04954591698aac4addac8fcb9 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.tanh
c1b61c387833adcc602036605886a05df1d18152 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.size
3b3ae7efd15a599a2ef48e98ff2c0b7c1303d187 Kefei Lu <kefeilu@fb.com> fx2trt: remove some comments
bc4724e645d10dd75b47cef084a680f5eaa9d0bf Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.gelu
252a70960be30e037bee2d712ff286743015df66 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.getitem
80039709fde7cdcdf81abdbfd2199adfe9df215b Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.gt
d6c9202ad3c1d9f35884cd189edad289e36a6787 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.hardsigmoid
6b1b7ba4690c8d5724ff6784f2f56a7cafad2c78 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.hardtanh
37747aec5f7ec8d1e1552f9280dae41910c966ba Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.isinf
6eb5891faae2e95ac7b6918f5a88b3a0a84a5140 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.leaky_relu
df212a34c238cd60a6e13f508400468cec539961 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.linear
8dc800208a7482d4b89a37526d37cdd0934a2402 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.logical_and
edfd5e9f6336700ebc91e907a9ee207d8fe25336 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.logical_or
588f7801308ee2129cb67276fe40e00f9078af1d Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.xor
cba566389cc4b310dbccceb9dd265ada4aded201 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.lt
67497e1c99ef796e5ad88b3ecf5751d596877932 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.max
c91cc6dda5221e55b2e960a6cd546f551d01cc7b Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.maximum
fd95c60b850bff8e00ba844da38c5d04c8bc25ba Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.maxpool
e7a00066f73231c466f79f15d7f48b917c855104 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.min
5f9fcfd27a4d1f7d6a8f034ea2b7c6c5c8a51809 Jason Park <jasonjk@fb.com> grouped swish LN
249ce82c30f0d70dd6de952c4e85ce273b09cd45 Jason Park <jasonjk@fb.com> Handling different eps for layer norms.
48907b6557d57569545d926b94f96c0d58ae2fcb Kefei Lu <kefeilu@fb.com> fx2trt: log input specs

Author: Wei

examples/fx/quantized_resnet_test.py

@@ -49,7 +49,7 @@ def build_int8_trt(rn18):
         # uncomment to check per channel quant works
         weight=torch.quantization.default_per_channel_weight_observer,
     )
-    prepared = prepare_fx(rn18, {"": qconfig})
+    prepared = prepare_fx(rn18, {"": qconfig}, data)
     for _ in range(10):
         prepared(data)
     quantized_rn18 = convert_to_reference(prepared)

py/torch_tensorrt/fx/lower.py

@@ -104,10 +104,13 @@ def __call__(self, mod, input, split_name) -> TRTInterpreterResult:
                     ),
                     self.lower_setting.opt_profile_replica,
                 )
-                if self.lower_setting.explicit_batch_dimension and self.lower_setting.dynamic_batch
+                if self.lower_setting.explicit_batch_dimension
+                and self.lower_setting.dynamic_batch
                 else InputTensorSpec.from_tensors(input)
             )
         )
+        logger.info(f"{split_name=} {input_specs_val=}")
+
         # Prepare algorithm selector and timing_cache for TRTInterpreter
         algo_selector = None
         if self.lower_setting.algo_selector:

py/torch_tensorrt/fx/lower_setting.py

@@ -64,11 +64,7 @@ class LowerSetting(LowerSettingBasic):
     cache file is provided.
     cuda_graph_batch_size (int): Cuda graph batch size, default to be -1.
     preset_lowerer (str): when specified, use a preset logic to build the
-    instance of Lowerer. Refer to
-    `caffe2.torch.fb.model_transform.fx2trt.presets.LowererPresetsManager` on
-    how presets are applied. Refer to
-    `caffe2.torch.fb.model_transform.fx2trt.presets.ESUHMLowererPreset` on how
-    to add a preset.
+    instance of Lowerer.
     opt_profile_replica (int): the number of opt profile set for TensorRT engine, this field is
     only used by explicit batch dim with dynamic shape mode.
     dynamic_batch: enable the dynamic shape in TRT with dim=-1 for the 1st dimension.

py/torch_tensorrt/fx/passes/pass_utils.py

@@ -63,14 +63,17 @@ def pass_with_validation(
                     y = y.cpu()
                 accuracy_check = torch.allclose(x, y, **kwargs)
                 if not accuracy_check:
+                    _LOGGER.error(
+                        f"Pass {pass_} failed correctness check, get original model output as {x} and processed model output as {y} for output {kk}."
+                    )
                     if suppress_accuracy_check_failure:
                         _LOGGER.error(
-                            f"pass {pass_} failed correctness check due to output {kk}, escape current pass."
+                            f"Pass {pass_} failed correctness check due to output {kk}."
                         )
                         return processed_module
                     else:
                         raise RuntimeError(
-                            f"pass {pass_} failed correctness check due to output {kk}"
+                            f"Pass {pass_} failed correctness check due to output {kk}"
                         )
             return processed_module
 

py/torch_tensorrt/fx/test/converters/acc_op/test_dequantize.py

@@ -45,6 +45,24 @@ def forward(self, x):
             TestModule(), input_specs, expected_ops={acc_ops.dequantize}
         )
 
+    def test_dequantize_with_dynamic_shape_four_dimensions(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                x = torch.quantize_per_tensor(x, 1, 0, torch.quint8)
+                return x.dequantize()
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (1, 2, 3, 3), (3, 3, 3, 3))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            TestModule(), input_specs, expected_ops={acc_ops.dequantize}
+        )
+
 
 if __name__ == "__main__":
     run_tests()

py/torch_tensorrt/fx/test/converters/acc_op/test_einsum.py

@@ -3,7 +3,7 @@
 import torch_tensorrt.fx.tracer.acc_tracer.acc_ops as acc_ops
 from parameterized import parameterized
 from torch.testing._internal.common_utils import run_tests
-from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase
+from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase, InputTensorSpec
 
 
 class TestConverter(AccTestCase):
@@ -30,6 +30,37 @@ def forward(self, x, y):
             test_implicit_batch_dim=False,
         )
 
+    @parameterized.expand(
+        [
+            ("4d_dim", "bcwd,bcdh->bcwh", (2, 3, 4, 5), (2, 3, 5, 6)),
+            ("4d_dim_ext", "bcxd,bcyd->bcxy", (2, 3, 4, 5), (2, 3, 6, 5)),
+            # TRT does not support ellipsis or diagonal operations
+        ]
+    )
+    def test_einsum_with_dynamic_shape_four_dimensions(
+        self, _, equation, x_size, y_size
+    ):
+        class Einsum(nn.Module):
+            def forward(self, x, y):
+                return torch.einsum(equation, x, y)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 3, 3), (1, 2, 3, 3), (3, 3, 3, 3))],
+            ),
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 3, 3), (1, 2, 3, 3), (3, 3, 3, 3))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            Einsum(), input_specs, expected_ops={acc_ops.einsum}
+        )
+
 
 if __name__ == "__main__":
     run_tests()

py/torch_tensorrt/fx/test/converters/acc_op/test_elu.py

@@ -30,6 +30,23 @@ def forward(self, x):
             TestModule(), input_specs, expected_ops={acc_ops.elu}
         )
 
+    def test_elu_with_dynamic_shape_four_dimensions(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.elu(x)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (1, 2, 3, 5), (3, 3, 3, 5))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            TestModule(), input_specs, expected_ops={acc_ops.elu}
+        )
+
 
 if __name__ == "__main__":
     run_tests()

py/torch_tensorrt/fx/test/converters/acc_op/test_embedding.py

@@ -5,7 +5,7 @@
 import torch_tensorrt.fx.tracer.acc_tracer.acc_ops as acc_ops
 from parameterized import param, parameterized
 from torch.testing._internal.common_utils import run_tests
-from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase
+from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase, InputTensorSpec
 
 
 @unittest.skip(
@@ -62,6 +62,46 @@ def forward(self, indices, weights):
             test_explicit_batch_dim=True,
         )
 
+    def test_embedding_with_dynamic_shape_four_dimensions(
+        self,
+        test_name,
+        indices_tensor,
+        weights_tensor,
+        padding_idx=None,
+        max_norm=None,
+        norm_type=2.0,
+        scale_grad_by_freq=False,
+        sparse=False,
+    ):
+        class TestEmbedding(torch.nn.Module):
+            def forward(self, indices, weights):
+                return torch.nn.functional.embedding(
+                    input=indices,
+                    weight=weights,
+                    padding_idx=padding_idx,
+                    max_norm=max_norm,
+                    norm_type=norm_type,
+                    scale_grad_by_freq=scale_grad_by_freq,
+                    sparse=sparse,
+                )
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (2, 3, 4, 5), (2, 3, 10, 10))],
+            ),
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (2, 3, 4, 5), (2, 3, 10, 10))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            TestEmbedding(), input_specs, expected_ops={acc_ops.embedding}
+        )
+
 
 if __name__ == "__main__":
     run_tests()

py/torch_tensorrt/fx/test/converters/acc_op/test_eq.py

@@ -2,7 +2,7 @@
 import torch_tensorrt.fx.tracer.acc_tracer.acc_ops as acc_ops
 from parameterized import parameterized
 from torch.testing._internal.common_utils import run_tests
-from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase
+from torch_tensorrt.fx.tools.common_fx2trt import AccTestCase, InputTensorSpec
 
 
 class TestEqConverter(AccTestCase):
@@ -184,6 +184,28 @@ def forward(self, x, y):
         )
 
 
+class TestEqOperatorSimpleConverterWithDynamicShape(AccTestCase):
+    def test_eq(self):
+        class Eq(torch.nn.Module):
+            def forward(self, x, y):
+                return x == y
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (2, 3, 4, 5), (2, 3, 10, 10))],
+            ),
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (2, 3, 4, 5), (2, 3, 10, 10))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(Eq(), input_specs, expected_ops={acc_ops.eq})
+
+
 class TestEqOperatorConstantConverter(AccTestCase):
     @parameterized.expand(
         [
@@ -243,5 +265,25 @@ def forward(self, x):
         )
 
 
+class TestConstInputConverterWithDynamicShape(AccTestCase):
+    def test_eq(self):
+        class Eq(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                return x.shape[0] == 4
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (1, 2, 3, 3), (3, 3, 3, 3))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(Eq(), input_specs, expected_ops={acc_ops.eq})
+
+
 if __name__ == "__main__":
     run_tests()

py/torch_tensorrt/fx/test/converters/acc_op/test_gelu.py

@@ -35,6 +35,23 @@ def forward(self, x):
             TestModule(), input_specs, expected_ops={acc_ops.gelu}
         )
 
+    def test_gelu_with_dynamic_shape_four_dimensions(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return nn.functional.gelu(x)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (1, 2, 3, 3), (3, 3, 3, 3))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            TestModule(), input_specs, expected_ops={acc_ops.gelu}
+        )
+
 
 if __name__ == "__main__":
     run_tests()