feat: support conv dynamo converter

fix a squeeze bug

minor fix and issue #2185

Author: zewenli98

py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py

@@ -812,3 +812,34 @@ def aten_ops_isinf(
         name,
         args[0],
     )
+
+
+@dynamo_tensorrt_converter(torch.ops.aten.convolution.default)
+def aten_ops_convolution(
+    network: TRTNetwork,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> Union[TRTTensor, Sequence[TRTTensor]]:
+    # we do not handle transposed.
+    if args[6] is True:
+        raise RuntimeError(f"Target {target} does not support `transposed=True` ")
+    # we do not handle output_padding.
+    if args[7] not in ([0], [0, 0], [0, 0, 0]):
+        raise RuntimeError(f"Target {target} has non-0 output_padding")
+
+    return impl.conv.convNd(
+        network,
+        target,
+        source_ir=SourceIR.ATEN,
+        name=name,
+        is_conv1d=len(args[3]) == 1,
+        input=args[0],
+        weight=args[1],
+        bias=args[2],
+        stride=args[3],
+        padding=args[4],
+        dilation=args[5],
+        groups=args[8],
+    )

py/torch_tensorrt/dynamo/conversion/converter_utils.py

@@ -1,6 +1,6 @@
 import logging
 import re
-from typing import List, Optional
+from typing import Any, List, Optional, Tuple
 
 import tensorrt as trt
 import torch
@@ -157,3 +157,27 @@ def broadcastable(
         if not (a_shape[i] == b_shape[i] or a_shape[i] == 1 or b_shape[i] == 1):
             return False
     return True
+
+
+def extend_attr_to_tuple(
+    val: Any,
+    num_elem: int,
+) -> Tuple[Any, ...]:
+    """
+    If `val` is not a tuple or a list, then we make a tuple of size `num_elem` by
+    replicating `val` `num_elem` times.
+
+    Args:
+        val (Any): Value that we want to process.
+
+    Returns:
+        A tuple.
+    """
+    if not isinstance(val, (tuple, list)):
+        val = (val,) * num_elem
+    elif len(val) == 1:
+        val = (val[0],) * num_elem
+
+    if isinstance(val, list):
+        val = tuple(val)
+    return val

py/torch_tensorrt/dynamo/conversion/impl/__init__.py

@@ -4,6 +4,7 @@
     activation,
     cast,
     condition,
+    conv,
     elementwise,
     embedding,
     matmul,

py/torch_tensorrt/dynamo/conversion/impl/conv.py

@@ -0,0 +1,129 @@
+from typing import Optional, Sequence, Union
+
+import numpy as np
+
+# @manual=//deeplearning/trt/python:py_tensorrt
+import tensorrt as trt
+import torch
+from torch.fx.node import Target
+from torch_tensorrt.dynamo.conversion import aten_ops_converters
+from torch_tensorrt.dynamo.conversion.converter_utils import extend_attr_to_tuple
+from torch_tensorrt.fx.converters.converter_utils import (
+    SourceIR,
+    get_dyn_range,
+    get_trt_tensor,
+    has_dynamic_shape,
+    mark_as_int8_layer,
+    set_layer_name,
+    to_numpy,
+)
+from torch_tensorrt.fx.types import TRTNetwork, TRTTensor
+
+
+def convNd(
+    network: TRTNetwork,
+    target: Union[Target, str],
+    source_ir: Optional[SourceIR],
+    name: str,
+    is_conv1d: bool,
+    input: TRTTensor,
+    weight: Union[TRTTensor, torch.Tensor],
+    bias: Optional[Union[TRTTensor, torch.Tensor]],
+    stride: Optional[Union[int, Sequence[int]]],
+    padding: Optional[Union[int, Sequence[int]]],
+    dilation: Optional[Union[int, Sequence[int]]],
+    groups: Optional[int],
+    scale: Optional[Union[torch.Tensor, float]] = None,
+    zero_point: Optional[Union[torch.Tensor, float]] = None,
+) -> TRTTensor:
+    if has_dynamic_shape(input.shape):
+        assert input.shape[1] != -1, "Channel dim can't be dynamic for convolution."
+
+    if is_conv1d:
+        # Apply an unsqueeze operation to transform the conv1d problem into conv2d
+        input = aten_ops_converters.aten_ops_unsqueeze(
+            network, target, (input, -1), {}, name + "_unsqueeze"
+        )
+
+    # Process bias terms
+    if isinstance(bias, (torch.Tensor, np.ndarray)):
+        # Transform the bias constant into a Numpy array
+        bias = to_numpy(bias)
+
+    elif isinstance(bias, TRTTensor):
+        bias = get_trt_tensor(network, bias, f"{name}_bias")
+
+    elif bias is not None:
+        raise RuntimeError(
+            f"Convolution {name} has bias of type {type(bias)}, Expected Torch Tensor or TRT Tensor"
+        )
+
+    # Process weight terms
+    if network.has_explicit_precision or isinstance(weight, TRTTensor):
+        weight = get_trt_tensor(network, weight, f"{name}_weight")
+        # Append new dimension (unsqueeze) if the convolution is 1d
+        if is_conv1d:
+            weight = aten_ops_converters.aten_ops_unsqueeze(
+                network, target, (weight, -1), {}, name + "_unsqueeze_weight"
+            )
+
+    elif isinstance(weight, (torch.Tensor, np.ndarray)):
+        # Transform the weight constant into a Numpy array
+        weight = to_numpy(weight)
+
+        # Append new dimension (unsqueeze) if the convolution is 1d
+        if is_conv1d:
+            weight = np.expand_dims(weight, -1)
+
+    else:
+        raise RuntimeError(
+            f"Convolution {name} has weight of type {type(weight)}, Expect Optional[Tensor]"
+        )
+
+    # add conv layer
+    conv_layer = network.add_convolution_nd(
+        input=input,
+        num_output_maps=weight.shape[0],
+        kernel_shape=weight.shape[2:],
+        kernel=trt.Weights() if isinstance(weight, TRTTensor) else weight,
+        bias=trt.Weights() if isinstance(bias, TRTTensor) else bias,
+    )
+
+    # If the weight is a TRTTensor, set it as an input of the layer
+    if isinstance(weight, TRTTensor):
+        conv_layer.set_input(1, weight)
+
+    # If the bias is a TRTTensor, set it as an input of the layer
+    if isinstance(bias, TRTTensor):
+        conv_layer.set_input(2, bias)
+
+    # Expand parameters manually for Conv1D computations
+    if is_conv1d:
+        padding = tuple(padding) + (0,)
+        stride = extend_attr_to_tuple(stride, 2)
+        dilation = extend_attr_to_tuple(dilation, 2)
+
+    set_layer_name(conv_layer, target, name, source_ir)
+
+    # Set relevant attributes of convolution layer
+    conv_layer.padding_nd = padding
+    conv_layer.stride_nd = stride
+    conv_layer.dilation_nd = dilation
+
+    if groups is not None:
+        conv_layer.num_groups = groups
+
+    # Handle quantization cases
+    if scale is not None and zero_point is not None:
+        # Assume the dtype of activation is torch.quint8
+        mark_as_int8_layer(conv_layer, get_dyn_range(scale, zero_point, torch.quint8))
+
+    result = conv_layer.get_output(0)
+
+    if is_conv1d:
+        # Apply a squeeze operation to transform the conv2d problem back into conv1d
+        result = aten_ops_converters.aten_ops_squeeze(
+            network, target, (result, -1), {}, name + "_squeeze"
+        )
+
+    return result

py/torch_tensorrt/dynamo/conversion/impl/squeeze.py

@@ -18,11 +18,6 @@ def squeeze(
     input: TRTTensor,
     dim: Optional[Any] = None,
 ) -> TRTTensor:
-    if not isinstance(input, TRTTensor):
-        raise RuntimeError(
-            f"squeeze received input {input} that is not part "
-            "of the TensorRT region!"
-        )
     dims = []
     if dim is not None:
         if isinstance(dim, int):
@@ -35,6 +30,7 @@ def squeeze(
     # dim, which is a very rare case. For now we just claim not supporting dim=None.
     assert not (len(dims) == 0), "We don't support dim=None right now for squeeze."
 
+    new_dims = []
     for dim in dims:
         dim = get_positive_dim(
             dim,
@@ -48,13 +44,14 @@ def squeeze(
         assert (
             len(get_dynamic_dims(input.shape)) <= 1
         ), "Currently more than one dynamic dim for input to squeeze is not supported."
+        new_dims.append(dim)
 
     output_shape = []
     for i, s in enumerate(input.shape):
-        if (i in dims) and s == 1:
+        if (i in new_dims) and s == 1:
             continue
         output_shape.append(s)
     layer = network.add_shuffle(input)
     layer.reshape_dims = tuple(output_shape)
-    set_layer_name(layer, target, name)
+    set_layer_name(layer, target, name, source_ir)
     return layer.get_output(0)