[quant] Add dequantize.tensors

aten/src/ATen/native/native_functions.yaml

@@ -3629,12 +3629,17 @@
   dispatch:
     CPU: quantize_per_channel_cpu
 
-- func: dequantize(Tensor self) -> Tensor
+- func: dequantize.self(Tensor self) -> Tensor
   use_c10_dispatcher: full
   variants: function, method
   dispatch:
     QuantizedCPU: dequantize_quant
 
+- func: dequantize.tensors(Tensor[] tensors) -> Tensor[]
+  variants: function
+  dispatch:
+    QuantizedCPU: dequantize_tensors_quant
+
 - func: q_scale(Tensor self) -> float
   use_c10_dispatcher: full
   variants: function, method

aten/src/ATen/native/quantized/QTensor.cpp

@@ -32,6 +32,14 @@ Tensor dequantize_quant(const Tensor& self) {
   return get_qtensorimpl(self)->quantizer()->dequantize(self);
 }
 
+std::vector<Tensor> dequantize_tensors_quant(TensorList tensors) {
+  std::vector<Tensor> dequantized_tensors;
+  for (auto i = 0; i < tensors.size(); ++i) {
+    dequantized_tensors.push_back(tensors[i].dequantize());
+  }
+  return dequantized_tensors;
+}
+
 double q_scale_quant(const Tensor& self) {
   auto quantizer = get_qtensorimpl(self)->quantizer();
   TORCH_CHECK(quantizer->qscheme() == kPerTensorAffine);

docs/source/torch.rst

@@ -55,6 +55,7 @@ Creation Ops
 .. autofunction:: full_like
 .. autofunction:: quantize_per_tensor
 .. autofunction:: quantize_per_channel
+.. autofunction:: dequantize
 
 Indexing, Slicing, Joining, Mutating Ops
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

test/backward_compatibility/check_backward_compatibility.py

@@ -112,6 +112,7 @@
     ('aten::_linear_prepack', datetime.date(2020, 4, 1)),
     ('aten::_conv2d_packed', datetime.date(2020, 4, 1)),
     ('aten::_conv2d_prepack', datetime.date(2020, 4, 1)),
+    ('aten::dequantize', datetime.date(2020, 4, 1)),
     ('aten::confirmed_by_owner', datetime.date(2020, 3, 17)),
     ('aten::owner', datetime.date(2020, 3, 27)),
     ('aten::owner_name', datetime.date(2020, 3, 27)),

tools/pyi/gen_pyi.py

@@ -92,6 +92,7 @@
     'div_',
     'div_out',
     'floor_divide', 'floor_divide_', 'floor_divide_out',
+    'dequantize',
 ]
 
 

torch/__init__.pyi.in

@@ -158,3 +158,12 @@ def compiled_with_cxx11_abi() -> _bool: ...
 # (similar to `unique`, `lu`, etc.); as such, it is not
 # possible to type correctly
 def nonzero(input: Tensor, *, out: Optional[Tensor]=None, as_tuple: Optional[_bool]=None): ...
+
+# we can't auto generate hints for torch.dequantize because it will generate
+# `dequantize(*tensors) -> Union[Tuple[Tensor, ...], List[Tensor]]: ...`
+# which overlaps with
+# `dequantize(self: Tensor) -> Tensor: ...
+@overload
+def dequantize(self: Tensor) -> Tensor: ...
+@overload
+def dequantize(tensors: Union[Tuple[Tensor, ...], List[Tensor]]) -> Union[Tuple[Tensor, ...], List[Tensor]]: ...

torch/_torch_docs.py

@@ -1513,6 +1513,23 @@ def merge_dicts(*dicts):
             -1.8209, -2.9780, -3.4022])
 """.format(**reduceops_common_args))
 
+add_docstr(torch.dequantize,
+           r"""
+.. function:: dequantize(tensor) -> Tensor
+
+Given a quantized Tensor, dequantize it and return an fp32 Tensor
+
+Args:
+    tensor (Tensor): A quantized Tensor
+
+.. function:: dequantize(tensors) -> sequence of Tensors
+
+Given a list of quantized Tensors, dequantize them and return a list of fp32 Tensors
+
+Args:
+     tensors (sequence of Tensors): A list of quantized Tensors
+""")
+
 add_docstr(torch.diag,
            r"""
 diag(input, diagonal=0, out=None) -> Tensor