Changes done internally at Facebook (#1161)

91196dedd10ed4061bef719b01fe917192a6fc17 Shiyan Deng <dsy842974287@fb.com> fix EngineHolder
608acce060d91ac93099f59e6165d43ffd79d50a Shiyan Deng <dsy842974287@fb.com> delete unused file
f88cb1566ae1c72634668fa26fadc3ef368b9727 Alex Beloi <alexbeloi@fb.com> [fx] add common_subexpression_elimination graph opt
b079de89bce912c29284809722cae63cb185c481 Yinghai Lu <yinghai@fb.com> handle list of integers in SliceOp check
e8f8dbda72625872a2657cdfd09dd59209f98090 Shiyan Deng <dsy842974287@fb.com> Add a pass to eliminate unsqueeze + cat + getitem pattern
675b28120c432c63bf4d451a4479b2376f73201d Kefei Lu <kefeilu@fb.com> Back out "[fx] add common_subexpression_elimination graph opt"
a3fb38bb6193b86a0ef0795a9104c073f5044598 Shiyan Deng <dsy842974287@fb.com> Back out "Back out "[fx] add common_subexpression_elimination graph opt""
0ea2b5782d70ed07b2158a96a45082aba68c55f6 Andrew Or <andrewor@fb.com> [Quant][fx][bc-breaking] Replace is_reference with convert_to_reference
2e4384397e73e99a3d56303864dc09e23420dda1 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape suport for acc_ops.sigmoid
78f7505cbe18850f43261b5f92e67262127e229e Mor Tzur <mortzur@fb.com> uncomment accidentally commented out acc_op tests
a7c0210a5c6468d978fc55a541b08da4e1782df7 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.avgpool and error message change
8aab78b134c778bcafe56f0fd5baa9bf51d7fef9 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.any
0591b671542c186a257152f8ed82aecbd2322feb Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.avgpool
5d36837498f72f5ad3432eecf3f4e4091ac23071 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.batchnorm
9a3134742872040b31a70bf216c81de5dcd82c45 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.binary_ops
4802100968d39418d6d1ce75c0407b3344e5cfd9 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.as_strided
6ccf77223a40e16d635d453b6342658bdc44062e Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.cat
0dd29ac935705926cc29bf03f5da922d93adae17 Shreyansh Prajapati <shreyanshp@fb.com> Test dynamic shape support for acc_ops.chunk
81737ebcecb0a57a683971d9052b3f3262030d20 Shreyansh Prajapati <shreyanshp@fb.com> Added test case for testing dynamic shape support for acc_ops.clamp
0c2604c6db94309c8a76de00e8c69ce21d2114a8 Yinghai Lu <yinghai@fb.com> [fx2trt] Fix constant tensor conversion dtype bug
f0edd3db9b8f74103b536e1f2af81ce8558d2f88 Lu Fang <lufang@fb.com> [Not for landing] Unblock ifr_unified_1_with_cover_dhen's lowering
e610d365fa21f495d9b651e519ce8a407d3aa744 Lu Fang <lufang@fb.com> [Not for landing] Unblock ifr_unified_1_with_cover_dhen's lowering
eec55e9c3bab30dea9cde52feb9cc74f30a43035 Yinghai Lu <yinghai@fb.com> [fx2trt] Fix constant tensor conversion dtype bug
851ca556610bd396645a9822932939d12d6ed915 Yinghai Lu <yinghai@fb.com> [fx2trt] Fix constant tensor conversion dtype bug

Author: Yinghai Lu

README.md

@@ -213,7 +213,7 @@ bazel build //:libtorchtrt --compilation_mode opt
 ```
 
 ### FX path (Python only) installation
-If the user plan to try FX path (Python only) and would like to avoid bazel build. Please follow the steps below.
+If the user plans to try FX path (Python only) and would like to avoid bazel build. Please follow the steps below.
 ``` shell
 cd py && python3 setup.py install --fx-only
 ```

examples/fx/quantized_resnet_test.py

@@ -6,7 +6,11 @@
 
 import torch_tensorrt.fx.tracer.acc_tracer.acc_tracer as acc_tracer
 import torchvision.models as models
-from torch.ao.quantization.quantize_fx import convert_fx, prepare_fx
+from torch.ao.quantization.quantize_fx import (
+    convert_fx,
+    convert_to_reference,
+    prepare_fx,
+)
 from torch.fx.experimental.normalize import NormalizeArgs
 from torch.fx.passes import shape_prop
 from torch_tensorrt.fx import InputTensorSpec, TRTInterpreter, TRTModule
@@ -48,7 +52,7 @@ def build_int8_trt(rn18):
     prepared = prepare_fx(rn18, {"": qconfig})
     for _ in range(10):
         prepared(data)
-    quantized_rn18 = convert_fx(prepared, is_reference=True)
+    quantized_rn18 = convert_to_reference(prepared)
     ref_res = quantized_rn18(data)
     print("quantized model:", quantized_rn18)
 

py/torch_tensorrt/fx/converters/acc_ops_converters.py

@@ -2222,7 +2222,7 @@ def acc_ops_adaptive_avg_poolnd(
     extend_len = 2 if target == acc_ops.adaptive_avg_pool2d else 3
     assert all(
         input_val.shape[-(i + 1)] != -1 for i in range(extend_len)
-    ), "AdaptiveAvgPool2d currently doesn't support dynamic shapes for last two dims."
+    ), "AdaptiveAvgPool2d and AdaptiveAvgPool3d currently doesn't support dynamic shapes for last two dims."
 
     output_size = cast(
         Sequence[int], extend_attr_to_tuple(kwargs["output_size"], extend_len)

py/torch_tensorrt/fx/converters/converter_utils.py

@@ -415,9 +415,8 @@ def add_binary_elementwise_layer(
     This function adds a TensorRT elementwise layer. We allow both operands to be
     constant (not a trt tensor) because in implicit batch dimension mode, we could
     introduce constant via .size() op. Other scenario should be const folded first.
-    If any operand is not a trt tensor, we make it a trt constant layer which has
-    the same type as the other trt tensor. Then we broadcast these two inputs to
-    have the same number of dimensions.
+    If any operand is not a trt tensor, we make it a trt constant layer while preserve
+    its dtype. Then we broadcast these two inputs to have the same number of dimensions.
 
     Limitation:
         If we are using implicit batch dim mode, the operand that is not a trt
@@ -436,14 +435,16 @@ def add_binary_elementwise_layer(
     Returns:
         The output of TensorRT Elementwise layer.
     """
-    dtype = None
+    lhs_dtype = None
+    rhs_dtype = None
     is_lhs_trt_tensor = False
     is_rhs_trt_tensor = False
+
     if isinstance(lhs_val, TRTTensor):
-        dtype = torch_dtype_from_trt(lhs_val.dtype)
+        lhs_dtype = torch_dtype_from_trt(lhs_val.dtype)
         is_lhs_trt_tensor = True
     if isinstance(rhs_val, TRTTensor):
-        dtype = torch_dtype_from_trt(rhs_val.dtype)
+        rhs_dtype = torch_dtype_from_trt(rhs_val.dtype)
         is_rhs_trt_tensor = True
 
     if not is_lhs_trt_tensor and not is_rhs_trt_tensor:
@@ -463,10 +464,14 @@ def add_binary_elementwise_layer(
     # this way the shape will become [1], and then will be properly squeezed
     # into [], meaning that the result will have shape [], which is what we
     # expect.
+    #
+    # Note that the dtype here is supposed to be the same as the scalar
+    # dtype but we don't have a way to detect whether it makes sense for the
+    # scalar to be float or half. Hence we go with the lhs dtype.
     if is_lhs_trt_tensor and isinstance(rhs_val, (float, int)):
-        rhs_val = torch.tensor([rhs_val], dtype=dtype)
+        rhs_val = torch.tensor([rhs_val], dtype=lhs_dtype)
     if is_rhs_trt_tensor and isinstance(lhs_val, (float, int)):
-        lhs_val = torch.tensor([lhs_val], dtype=dtype)
+        lhs_val = torch.tensor([lhs_val], dtype=rhs_dtype)
 
     # When lhs is scalar, and rhs has shape [1,], then currently the assert
     # will fail because lhs shape has fewer dimensions than rhs shape.  This
@@ -482,8 +487,8 @@ def add_binary_elementwise_layer(
         if isinstance(rhs_val, torch.Tensor):
             rhs_val = squeeze_left(rhs_val)
 
-    lhs_val = get_trt_tensor(network, lhs_val, f"{name}_lhs", dtype)
-    rhs_val = get_trt_tensor(network, rhs_val, f"{name}_rhs", dtype)
+    lhs_val = get_trt_tensor(network, lhs_val, f"{name}_lhs", lhs_dtype)
+    rhs_val = get_trt_tensor(network, rhs_val, f"{name}_rhs", rhs_dtype)
 
     # Check the limitation in the doc string.
     if network.has_implicit_batch_dimension:

py/torch_tensorrt/fx/fx2trt.py

@@ -80,7 +80,7 @@ def __init__(
         ] = dict()
 
     def validate_input_specs(self):
-        for shape, dtpe, _, shape_ranges, has_batch_dim in self.input_specs:
+        for shape, _, _, shape_ranges, has_batch_dim in self.input_specs:
             if not self.network.has_implicit_batch_dimension:
                 assert (
                     has_batch_dim

py/torch_tensorrt/fx/passes/graph_opts.py

@@ -0,0 +1,74 @@
+from collections.abc import Sequence
+
+import torch
+import torch.fx
+
+
+def common_subexpression_elimination(graph_module: torch.fx.GraphModule) -> bool:
+    """
+    Optimize quantization by removing repeated subexpressions.
+
+    Args:
+        graph_module(torch.fx.GraphModule): target module to be optimized
+
+    Returns:
+        Graph changed or not.
+    """
+
+    def seq_hashable(seq):
+        if seq is None:
+            return None
+
+        items = []
+        for old in seq:
+            if isinstance(old, Sequence) and not isinstance(old, str):
+                new = seq_hashable(old)
+            elif isinstance(old, dict):
+                new = dict_hashable(old)
+            elif isinstance(old, slice):
+                new = old.__reduce__()
+            else:
+                new = old
+
+            items.append(new)
+
+        return tuple(items)
+
+    def dict_hashable(d):
+        if d is None:
+            return None
+
+        items = []
+        for k, old_v in d.items():
+            if isinstance(old_v, Sequence):
+                new_v = seq_hashable(old_v)
+            elif isinstance(old_v, dict):
+                new_v = dict_hashable(old_v)
+            elif isinstance(old_v, slice):
+                new_v = old_v.__reduce__()
+            else:
+                new_v = old_v
+
+            items.append((k, new_v))
+        return tuple(sorted(items))
+
+    changed = False
+    env = {}
+    for n in graph_module.graph.nodes:
+        # do not CSE away impure ops
+        if n.op not in {"call_function", "call_method"} or n.is_impure():
+            continue
+
+        # hash target, args, kwargs
+        hash_val = (n.target, seq_hashable(n.args), dict_hashable(n.kwargs))
+
+        # check if a node has a substitute and can be eliminated
+        if hash_val in env:
+            n.replace_all_uses_with(env[hash_val])
+            graph_module.graph.erase_node(n)
+            changed = True
+            continue
+
+        env[hash_val] = n
+
+    return changed

py/torch_tensorrt/fx/passes/lower_pass_manager_builder.py

@@ -1,5 +1,5 @@
 from functools import partial, wraps
-from typing import Any, Callable, NamedTuple, Sequence
+from typing import Any, Callable, Sequence
 
 import torch
 from torch import nn
@@ -10,6 +10,7 @@
 from ..lower_setting import LowerSetting
 from ..observer import Observer
 from ..passes.remove_duplicate_output_args import remove_duplicate_output_args
+from .graph_opts import common_subexpression_elimination
 
 from .lower_basic_pass import run_const_fold
 
@@ -94,6 +95,8 @@ def graph_optimization_pass(self) -> PassManager:
             passes.append(wrapper(p, self._input))
         for p in self.lower_setting.lower_basic_fuse_pass.passes:
             passes.append(wrapper(p, self._input))
+
+        passes.append(inplace_wrapper(common_subexpression_elimination))
         passes.append(
             inplace_wrapper(lambda m: FUSE_PASSES_POST_OBSERVER.observe(m, self._input))
         )

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

@@ -92,6 +92,8 @@ def forward(self, x):
             TestModule(), input_specs, expected_ops={acc_ops.adaptive_avg_pool3d}
         )
 
+    #  Testing with shape(-1, -1, -1, -1) results into error: "AdaptiveAvgPool2d and AdaptiveAvgPool3d currently doesn't support dynamic shapes for last two dims."
+
 
 if __name__ == "__main__":
     run_tests()

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

@@ -5,6 +5,8 @@
 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 InputTensorSpec
+
 
 class TestAnyConverters(AccTestCase):
     @parameterized.expand(
@@ -64,6 +66,26 @@ def forward(self, x):
             test_implicit_batch_dim=False,
         )
 
+    # Testing with shape (-1, -1, -1, -1) results into error: torch.zeros(tuple([*input_t.shape])). Trying to create tensor with negative dimension -1: [-1, -1, -1, -1]
+    """
+    def test_ops_with_dynamic_shape_four_dimensions(self):
+        class TestModule(nn.Module):
+            def forward(self, x):
+                return torch.any(x)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 256, 256), (3, 3, 256, 256), (5, 5, 256, 256))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            TestModule(), input_specs, expected_ops={acc_ops.any}
+        )
+    """
+
 
 if __name__ == "__main__":
     run_tests()

py/torch_tensorrt/fx/test/converters/acc_op/test_as_strided.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 +30,26 @@ def forward(self, x):
             test_implicit_batch_dim=False,
         )
 
+    # Testing with shape (-1, -1, -1, -1) results into error: RuntimeError: setStorage: sizes [2, 3], strides [1, 2], storage offset 0, and itemsize 8 requiring a storage size of 48 are out of bounds for storage of size 1
+    """
+    def test_as_strided_with_dynamic_shape_four_dimensions(self):
+        class Stride(nn.Module):
+            def forward(self, x):
+                return torch.as_strided(torch.tensor([5, 5]), (2, 3), (1, 2), 0)
+
+        input_specs = [
+            InputTensorSpec(
+                shape=(-1, -1, -1, -1),
+                dtype=torch.float32,
+                shape_ranges=[((1, 1, 1, 1), (3, 3, 3, 3), (5, 5, 5, 5))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
+            Stride(), input_specs, expected_ops={acc_ops.as_strided}
+        )
+    """
+
 
 if __name__ == "__main__":
     run_tests()