bump PyTorch version to 1.11 (#10794)

* bump PyTorch version to 1.11

* disable some caffe2 ci

* Fix sub conversion in PyTorch frontend

* use fuse_modules_qat if available, fallback to fuse_modules for older PyTorch

* Re-Run CI

Author: t-vi

docker/Dockerfile.ci_gpu

@@ -77,9 +77,6 @@ RUN bash /install/ubuntu_install_onnx.sh
 COPY install/ubuntu_install_tflite.sh /install/ubuntu_install_tflite.sh
 RUN bash /install/ubuntu_install_tflite.sh
 
-COPY install/ubuntu_install_caffe2.sh /install/ubuntu_install_caffe2.sh
-RUN bash /install/ubuntu_install_caffe2.sh
-
 COPY install/ubuntu_install_dgl.sh /install/ubuntu_install_dgl.sh
 RUN bash /install/ubuntu_install_dgl.sh
 

docker/install/ubuntu_install_onnx.sh

@@ -36,5 +36,5 @@ pip3 install \
 pip3 install future
 
 pip3 install \
-    torch==1.10.1 \
-    torchvision==0.11.2
+    torch==1.11.0 \
+    torchvision==0.12.0

gallery/how_to/compile_models/from_caffe2.py

@@ -2039,6 +2039,14 @@ def stack(self, inputs, input_types):
             assert isinstance(ty, tvm.ir.TypeCall) and ty.func == list_ty, msg
             return self.tensor_array_stack(inputs, input_types)
 
+    def sub(self, inputs, input_types):
+        if len(inputs) == 3:
+            data0, data1, alpha = self.pytorch_promote_types(inputs, input_types)
+            return get_relay_op("subtract")(data0, alpha * data1)
+        else:
+            data0, data1= self.pytorch_promote_types(inputs, input_types)
+            return get_relay_op("subtract")(data0, data1)
+
     def rsub(self, inputs, input_types):
         data0, data1, alpha = self.pytorch_promote_types(inputs, input_types)
 
@@ -2859,7 +2867,10 @@ def all_any_common(self, op, inputs, input_types):
             inp = inputs[0]
         return op(inp, axis=dim, keepdims=keepdim)
 
-    def searchsorted_common(self, sorted_sequence, values, out_int32, right):
+    def searchsorted_common(
+        self, sorted_sequence, values, out_int32, right, side=None, out=None, sorter=None
+    ):
+        assert side is None and out is None and sorter is None, "unsupported parameters"
         dtype = "int32" if out_int32 else "int64"
         values_shape = _infer_shape(values)
 
@@ -2959,7 +2970,7 @@ def create_convert_map(self):
             "aten::pixel_shuffle": self.pixel_shuffle,
             "aten::device": self.none,
             "prim::device": self.none,
-            "aten::sub": self.make_elemwise("subtract"),
+            "aten::sub": self.sub,
             "aten::max": self.max,
             "aten::min": self.min,
             "aten::mul": self.make_elemwise("multiply"),

python/tvm/relay/frontend/pytorch.py

@@ -743,9 +743,10 @@ def __init__(self, inputsize=(128, 128)):
             self.backbone = Backbone()
 
         def fuse_model(self):
+            fuse_modules_qat = getattr(torch.ao.quantization, "fuse_modules_qat", fuse_modules)
             for idx, m in enumerate(self.modules()):
                 if type(m) == ConvBnRelu:
-                    torch.quantization.fuse_modules(m, ["conv", "bn", "relu"], inplace=True)
+                    fuse_modules_qat(m, ["conv", "bn", "relu"], inplace=True)
 
         def forward(self, input):
             input = self.quant(input)

tests/python/frontend/pytorch/qnn_test.py

@@ -56,9 +56,6 @@ function shard2 {
         i=$((i+1))
     done
 
-    echo "Running relay caffe2 frontend test..."
-    run_pytest cython python-frontend-caffe2 tests/python/frontend/caffe2
-
     echo "Running relay DarkNet frontend test..."
     run_pytest cython python-frontend-darknet tests/python/frontend/darknet