[FX] Dynamic Shapes Support (#3225)

### Changes

Modify NNCF Graph Builder for FX backend to correctly get and insert the
dynamic shapes into NNCFGraph

### Reason for changes

To support quantization of Torch FX models exported with dynamic shapes

### Tests

test is added to `tests/torch/fx/test_models.py` in
test_quantized_models(). Currently only the synthetic transformer is
tested because torch.export.dynamic_shapes.Dim.DYNAMIC is not supported
in pytorch but is supported in upcoming releases.
https://pytorch.org/tutorials/intermediate/torch_export_tutorial.html#constraints-dynamic-shapes

`test_dynamic_edge()` is also added in `tests/torch/fx/test_models.py`
to check that the tensor shape in NNCF Graph edge has values only of
type int or str and not SymInt.

---------

Co-authored-by: Alexander Dokuchaev <alexander.dokuchaev@intel.com>

Author: anzr299

nncf/experimental/torch/fx/nncf_graph_builder.py

@@ -196,7 +196,9 @@ def get_edge_params(
             else:
                 tensor = source_node.meta["val"]
             if isinstance(tensor, torch.Tensor):
-                tensor_shape = tuple(tensor.shape)
+                tensor_shape = tuple(-1 if isinstance(i, torch.SymInt) else i for i in tensor.shape)
+            elif isinstance(tensor, torch.SymInt):
+                tensor_shape = (-1,)
 
         if tensor_shape is None:
             # TODO(dlyakhov): Refactor algorithms to always have knowns edges shapes.

tests/torch/data/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/mobilenet_v3_small.dot

@@ -0,0 +1,49 @@
+strict digraph  {
+"0 wte_weight" [id=0, type=get_attr];
+"1 linear_bias" [id=1, type=get_attr];
+"2 lm_head_bias" [id=2, type=get_attr];
+"3 input_ids" [id=3, type=input];
+"4 embedding" [id=4, type=embedding];
+"5 embedding_0_0_nncf_smooth_quant_0" [id=5, type=call_module];
+"6 quantize_per_tensor_default" [id=6, type=quantize_per_tensor];
+"7 dequantize_per_tensor_default" [id=7, type=dequantize_per_tensor];
+"8 scale_updated_constant0" [id=8, type=get_attr];
+"9 compressed_weight_updated_constant0" [id=9, type=get_attr];
+"10 mul_tensor" [id=10, type=mul];
+"11 zero_point_updated_constant0" [id=11, type=get_attr];
+"12 sub_tensor" [id=12, type=sub];
+"13 linear" [id=13, type=linear];
+"14 linear_0_0_nncf_smooth_quant_0" [id=14, type=call_module];
+"15 quantize_per_tensor_default_1" [id=15, type=quantize_per_tensor];
+"16 dequantize_per_tensor_default_1" [id=16, type=dequantize_per_tensor];
+"17 scale_updated_constant1" [id=17, type=get_attr];
+"18 compressed_weight_updated_constant1" [id=18, type=get_attr];
+"19 mul_tensor_1" [id=19, type=mul];
+"20 zero_point_updated_constant1" [id=20, type=get_attr];
+"21 sub_tensor_1" [id=21, type=sub];
+"22 linear_1" [id=22, type=linear];
+"23 output" [id=23, type=output];
+"0 wte_weight" -> "4 embedding"  [label="(10, 5)", style=solid];
+"1 linear_bias" -> "13 linear"  [label="(5,)", style=solid];
+"2 lm_head_bias" -> "22 linear_1"  [label="(10,)", style=solid];
+"3 input_ids" -> "4 embedding"  [label="(-1,)", style=solid];
+"4 embedding" -> "5 embedding_0_0_nncf_smooth_quant_0"  [label="(-1, 5)", style=solid];
+"5 embedding_0_0_nncf_smooth_quant_0" -> "6 quantize_per_tensor_default"  [label="(-1, 5)", style=solid];
+"6 quantize_per_tensor_default" -> "7 dequantize_per_tensor_default"  [label="(-1, 5)", style=solid];
+"7 dequantize_per_tensor_default" -> "13 linear"  [label="(-1, 5)", style=solid];
+"8 scale_updated_constant0" -> "10 mul_tensor"  [label="(5, 1)", style=solid];
+"9 compressed_weight_updated_constant0" -> "10 mul_tensor"  [label="(5, 5)", style=solid];
+"10 mul_tensor" -> "12 sub_tensor"  [label="(5, 5)", style=solid];
+"11 zero_point_updated_constant0" -> "12 sub_tensor"  [label="(5, 1)", style=solid];
+"12 sub_tensor" -> "13 linear"  [label="(5, 5)", style=solid];
+"13 linear" -> "14 linear_0_0_nncf_smooth_quant_0"  [label="(-1, 5)", style=solid];
+"14 linear_0_0_nncf_smooth_quant_0" -> "15 quantize_per_tensor_default_1"  [label="(-1, 5)", style=solid];
+"15 quantize_per_tensor_default_1" -> "16 dequantize_per_tensor_default_1"  [label="(-1, 5)", style=solid];
+"16 dequantize_per_tensor_default_1" -> "22 linear_1"  [label="(-1, 5)", style=solid];
+"17 scale_updated_constant1" -> "19 mul_tensor_1"  [label="(10, 1)", style=solid];
+"18 compressed_weight_updated_constant1" -> "19 mul_tensor_1"  [label="(10, 5)", style=solid];
+"19 mul_tensor_1" -> "21 sub_tensor_1"  [label="(10, 5)", style=solid];
+"20 zero_point_updated_constant1" -> "21 sub_tensor_1"  [label="(10, 1)", style=solid];
+"21 sub_tensor_1" -> "22 linear_1"  [label="(10, 5)", style=solid];
+"22 linear_1" -> "23 output"  [label="(-1, 10)", style=solid];
+}

tests/torch/data/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/resnet18.dot

@@ -0,0 +1,63 @@
+strict digraph  {
+"0 x" [id=0, type=input];
+"1 x_0_0_nncf_smooth_quant_0" [id=1, type=call_module];
+"2 quantize_per_tensor_default" [id=2, type=quantize_per_tensor];
+"3 dequantize_per_tensor_default" [id=3, type=dequantize_per_tensor];
+"4 scale_updated_constant0" [id=4, type=get_attr];
+"5 compressed_weight_updated_constant0" [id=5, type=get_attr];
+"6 mul_tensor" [id=6, type=mul];
+"7 zero_point_updated_constant0" [id=7, type=get_attr];
+"8 sub_tensor" [id=8, type=sub];
+"9 linear" [id=9, type=linear];
+"10 quantize_per_tensor_default_1" [id=10, type=quantize_per_tensor];
+"11 dequantize_per_tensor_default_1" [id=11, type=dequantize_per_tensor];
+"12 slice_1" [id=12, type=slice];
+"13 slice_2" [id=13, type=slice];
+"14 slice_3" [id=14, type=slice];
+"15 quantize_per_tensor_default_2" [id=15, type=quantize_per_tensor];
+"16 dequantize_per_tensor_default_2" [id=16, type=dequantize_per_tensor];
+"17 slice_4" [id=17, type=slice];
+"18 slice_5" [id=18, type=slice];
+"19 slice_6" [id=19, type=slice];
+"20 slice_7" [id=20, type=slice];
+"21 slice_8" [id=21, type=slice];
+"22 slice_9" [id=22, type=slice];
+"23 transpose" [id=23, type=transpose];
+"24 matmul" [id=24, type=matmul];
+"25 div_" [id=25, type=div_];
+"26 softmax" [id=26, type=softmax];
+"27 transpose_1" [id=27, type=transpose];
+"28 matmul_1" [id=28, type=matmul];
+"29 output" [id=29, type=output];
+"0 x" -> "1 x_0_0_nncf_smooth_quant_0"  [label="(1, -1, -1)", style=solid];
+"1 x_0_0_nncf_smooth_quant_0" -> "2 quantize_per_tensor_default"  [label="(1, -1, 4)", style=solid];
+"2 quantize_per_tensor_default" -> "3 dequantize_per_tensor_default"  [label="(1, -1, 4)", style=solid];
+"3 dequantize_per_tensor_default" -> "9 linear"  [label="(1, -1, 4)", style=solid];
+"4 scale_updated_constant0" -> "6 mul_tensor"  [label="(12, 1)", style=solid];
+"5 compressed_weight_updated_constant0" -> "6 mul_tensor"  [label="(12, 4)", style=solid];
+"6 mul_tensor" -> "8 sub_tensor"  [label="(12, 4)", style=solid];
+"7 zero_point_updated_constant0" -> "8 sub_tensor"  [label="(12, 1)", style=solid];
+"8 sub_tensor" -> "9 linear"  [label="(12, 4)", style=solid];
+"9 linear" -> "10 quantize_per_tensor_default_1"  [label="(1, -1, 12)", style=solid];
+"9 linear" -> "15 quantize_per_tensor_default_2"  [label="(1, -1, 12)", style=solid];
+"9 linear" -> "20 slice_7"  [label="(1, -1, 12)", style=solid];
+"10 quantize_per_tensor_default_1" -> "11 dequantize_per_tensor_default_1"  [label="(1, -1, 12)", style=solid];
+"11 dequantize_per_tensor_default_1" -> "12 slice_1"  [label="(1, -1, 12)", style=solid];
+"12 slice_1" -> "13 slice_2"  [label="(1, -1, 12)", style=solid];
+"13 slice_2" -> "14 slice_3"  [label="(1, -1, 12)", style=solid];
+"14 slice_3" -> "24 matmul"  [label="(1, -1, 4)", style=solid];
+"15 quantize_per_tensor_default_2" -> "16 dequantize_per_tensor_default_2"  [label="(1, -1, 12)", style=solid];
+"16 dequantize_per_tensor_default_2" -> "17 slice_4"  [label="(1, -1, 12)", style=solid];
+"17 slice_4" -> "18 slice_5"  [label="(1, -1, 12)", style=solid];
+"18 slice_5" -> "19 slice_6"  [label="(1, -1, 12)", style=solid];
+"19 slice_6" -> "23 transpose"  [label="(1, -1, 4)", style=solid];
+"20 slice_7" -> "21 slice_8"  [label="(1, -1, 12)", style=solid];
+"21 slice_8" -> "22 slice_9"  [label="(1, -1, 12)", style=solid];
+"22 slice_9" -> "28 matmul_1"  [label="(1, -1, 4)", style=solid];
+"23 transpose" -> "24 matmul"  [label="(1, 4, -1)", style=solid];
+"24 matmul" -> "25 div_"  [label="(1, -1, -1)", style=solid];
+"25 div_" -> "26 softmax"  [label="(1, -1, -1)", style=solid];
+"26 softmax" -> "27 transpose_1"  [label="(1, -1, -1)", style=solid];
+"27 transpose_1" -> "28 matmul_1"  [label="(1, -1, -1)", style=solid];
+"28 matmul_1" -> "29 output"  [label="(1, -1, 4)", style=solid];
+}