[BE] Do not use unicode quotes (#99446)

They are mostly used in commented code examples, but even Python-3.12
does not recognize `“foobar”` as valid string literal

I.e. just `s/[“”]/"/`

Pull Request resolved: #99446
Approved by: https://github.com/huydhn, https://github.com/ezyang

torch/_dynamo/variables/builder.py

@@ -1136,18 +1136,18 @@ def wrap_to_fake_tensor_and_record(
         curr_sizes = None
         if name not in tx.output.frame_state:
             # If there is no entry for this source, add the tensor to frame state with its current static size.
-            # E.g., {} -> {“x”: [2, 4]}
+            # E.g., {} -> {"x": [2, 4]}
             curr_sizes = list(e.size())
         else:
             curr_sizes = tx.output.frame_state[name]
             if curr_sizes is not None:
                 if e.ndim != len(curr_sizes):
                     # If there is already an entry, and the dim mismatches, replace the frame state entry with None.
-                    # E.g. {“x”: [2, 3, 4]} -> {“x”: None}
+                    # E.g. {"x": [2, 3, 4]} -> {"x": None}
                     curr_sizes = None
                 else:
                     # If there is already an entry, and the dim matches, for every size in the frame state which
-                    # disagrees with the current static size, replace it with None. E.g., {“x”: [2, 3]} -> {“x”: [2, None]}
+                    # disagrees with the current static size, replace it with None. E.g., {"x": [2, 3]} -> {"x": [2, None]}
                     for i, dim in enumerate(curr_sizes):
                         if e.size()[i] != dim:
                             curr_sizes[i] = None

torch/_functorch/autograd_function.py

@@ -500,7 +500,7 @@ def get_tangents_in_dims(input_dims, tangents):
 # def backward_no_context(gy):
 #     return gy.expand([B, 4])
 #
-# gx = vmap(backward_no_context, dims)(gy: “Tensor[B]”)
+# gx = vmap(backward_no_context, dims)(gy: "Tensor[B]")
 #
 # This gives us the wrong result (gx has shape [B, B, 4], but it should
 # have shape [4]). Performing vmap over setup_context means the shape

torch/ao/quantization/fx/README.md

@@ -202,10 +202,10 @@ The overall logic to insert QDQStub1 and QDQStub2 inplace is the following:
 # node_name_to_target_dtype_info =
 # {
 #     # this is placeholder node in FX Graph
-#     “input” : {“input_activation”: torch.float32, “output_activation”: torch.float32},
-#     “qat_linear_relu”: {“input_activation”: torch.quint8, “output_activation”: torch.quint8, “weight”: ...}
+#     "input" : {"input_activation": torch.float32, "output_activation": torch.float32},
+#     "qat_linear_relu": {"input_activation": torch.quint8, "output_activation": torch.quint8, "weight": ...}
 #     # this is the return node in FX Graph
-#     “output”: {“input_activation”: torch.float32, “output_activation”: torch.float32}
+#     "output": {"input_activation": torch.float32, "output_activation": torch.float32}
 # }
 ```
 Note: this map is generated before we insert qdqstub to graph1, and will not change in the process.
@@ -259,7 +259,7 @@ Let’s say the output of `qat_linear_relu` Node is configured as float32, both
 }
 ```
 
-What we’ll do here is when we are trying to insert output QDQStub for `qat_linear_relu`, we look at the target output dtype for this node (node_name_to_target_dtype_info[“qat_linear_relu”][“output_activation”], and find that it is float, which is not a quantized dtype, so
+What we’ll do here is when we are trying to insert output QDQStub for `qat_linear_relu`, we look at the target output dtype for this node (node_name_to_target_dtype_info["qat_linear_relu"]["output_activation"], and find that it is float, which is not a quantized dtype, so
 will do nothing here.
 Note that this does not prevent other operators following `qat_linear_relu` to insert a QDQStub at the output of `qat_linear_relu`, since we are dealing with an `edge` of the graph here, and an `edge` is connected to two nodes, which means
 the output of `qat_linear_relu` will also be the input of a node following `qat_linear_relu`.