Merge pull request #475 from Xilinx/bump_to_42ba541c

[AutoBump] Merge with fixes of 42ba541 (Oct 22) (91)

Author: mgehre-amd

projects/pt1/e2e_testing/xfail_sets.py

@@ -420,15 +420,7 @@
     "ContainsIntList_False",
     "ContainsIntList_True",
     "Conv1dNoPaddingGroupModule_basic",
-    "Conv2dQInt8Module_basic",
-    "Conv2dQInt8Module_depthwise",
-    "Conv2dQInt8Module_grouped",
-    "Conv2dQInt8Module_not_depthwise",
-    "Conv2dQInt8PerChannelModule_basic",
-    "Conv2dQInt8PerChannelModule_depthwise",
-    "Conv2dQInt8PerChannelModule_grouped",
     "ConvTbcModule_basic",
-    "ConvTranspose2DQInt8_basic",
     "ConvolutionBackwardModule2DPadded_basic",
     "ConvolutionBackwardModule2DStrided_basic",
     "ConvolutionBackwardModule2D_basic",

projects/pt1/python/torch_mlir_e2e_test/test_suite/conv.py

@@ -1300,23 +1300,28 @@ def ConvTbcModule_basic(module, tu: TestUtils):
     module.forward(tu.rand(9, 4, 5), tu.rand(3, 5, 6), tu.rand(6))
 
 
+# For DQ-Q fake quantization ops
+import torch.ao.quantization.fx._decomposed
+
+
 class Conv2dQInt8ModuleBase(torch.nn.Module):
     def __init__(self, groups=1):
         self.groups = groups
         super().__init__()
 
-    def _forward(self, inputVec, weight, bias):
-        inputVec = torch._make_per_tensor_quantized_tensor(inputVec, 0.01, 7)
-        inputVec = torch.dequantize(inputVec)
-
-        weight = torch._make_per_tensor_quantized_tensor(weight, 0.01, 3)
-        weight = torch.dequantize(weight)
-
-        bias = torch.quantize_per_tensor(bias, 0.0001, 0, torch.qint32)
-        bias = torch.dequantize(bias)
+    def _forward(self, input, weight, bias):
+        input = torch.ops.quantized_decomposed.dequantize_per_tensor.default(
+            input, 0.01, 7, -128, 127, torch.int8
+        )
+        weight = torch.ops.quantized_decomposed.dequantize_per_tensor.default(
+            weight, 0.01, 3, -128, 127, torch.int8
+        )
+        bias = torch.ops.quantized_decomposed.dequantize_per_tensor.default(
+            bias, 1, 0, -1000, 1000, torch.int32
+        )
 
-        return torch.ops.aten.conv2d(
-            inputVec,
+        conv = torch.ops.aten.conv2d(
+            input,
             weight,
             bias=bias,
             stride=[1, 1],
@@ -1325,6 +1330,11 @@ def _forward(self, inputVec, weight, bias):
             groups=self.groups,
         )
 
+        # Use int32 to avoid overflows
+        return torch.ops.quantized_decomposed.quantize_per_tensor.default(
+            conv, 1, 0, -(2**31), 2**31 - 1, torch.int32
+        )
+
 
 class Conv2dQInt8ModuleDyn(Conv2dQInt8ModuleBase):
     @export
@@ -1333,7 +1343,7 @@ class Conv2dQInt8ModuleDyn(Conv2dQInt8ModuleBase):
             None,
             ([-1, -1, -1, -1], torch.int8, True),
             ([-1, -1, -1, -1], torch.int8, True),
-            ([-1], torch.float, True),
+            ([-1], torch.int32, True),
         ]
     )
     def forward(self, inputVec, weight, bias):
@@ -1347,7 +1357,7 @@ class Conv2dQInt8ModuleStatic(Conv2dQInt8ModuleBase):
             None,
             ([2, 3, 12, 12], torch.int8, True),
             ([3, 1, 5, 3], torch.int8, True),
-            ([3], torch.float, True),
+            ([3], torch.int32, True),
         ]
     )
     def forward(self, inputVec, weight, bias):
@@ -1361,7 +1371,7 @@ class Conv2dQInt8ModuleStatic_MoreOutChannels(Conv2dQInt8ModuleBase):
             None,
             ([2, 3, 12, 12], torch.int8, True),
             ([6, 1, 5, 3], torch.int8, True),
-            ([6], torch.float, True),
+            ([6], torch.int32, True),
         ]
     )
     def forward(self, inputVec, weight, bias):
@@ -1372,23 +1382,23 @@ def forward(self, inputVec, weight, bias):
 def Conv2dQInt8Module_basic(module, tu: TestUtils):
     inputVec = tu.randint(2, 4, 7, 8, low=-128, high=127).to(torch.int8)
     weight = tu.randint(3, 4, 3, 2, low=-128, high=127).to(torch.int8)
-    bias = torch.rand(3)
+    bias = tu.randint(3, low=-1000, high=1000).to(torch.int32)
     module.forward(inputVec, weight, bias)
 
 
 @register_test_case(module_factory=lambda: Conv2dQInt8ModuleDyn(groups=2))
 def Conv2dQInt8Module_grouped(module, tu: TestUtils):
     inputVec = tu.randint(2, 8, 7, 8, low=-128, high=127).to(torch.int8)
     weight = tu.randint(6, 4, 3, 2, low=-128, high=127).to(torch.int8)
-    bias = torch.rand(6)
+    bias = tu.randint(6, low=-1000, high=1000).to(torch.int32)
     module.forward(inputVec, weight, bias)
 
 
 @register_test_case(module_factory=lambda: Conv2dQInt8ModuleStatic(groups=3))
 def Conv2dQInt8Module_depthwise(module, tu: TestUtils):
     inputVec = tu.randint(2, 3, 12, 12, low=-128, high=127).to(torch.int8)
     weight = tu.randint(3, 1, 5, 3, low=-128, high=127).to(torch.int8)
-    bias = torch.rand(3)
+    bias = tu.randint(3, low=-1000, high=1000).to(torch.int32)
     module.forward(inputVec, weight, bias)
 
 
@@ -1398,7 +1408,7 @@ def Conv2dQInt8Module_depthwise(module, tu: TestUtils):
 def Conv2dQInt8Module_not_depthwise(module, tu: TestUtils):
     inputVec = tu.randint(2, 3, 12, 12, low=-128, high=127).to(torch.int8)
     weight = tu.randint(6, 1, 5, 3, low=-128, high=127).to(torch.int8)
-    bias = torch.rand(6)
+    bias = tu.randint(6, low=-1000, high=1000).to(torch.int32)
     module.forward(inputVec, weight, bias)
 
 
@@ -1417,24 +1427,29 @@ def __init__(self):
         ]
     )
     def forward(self, input, weight, bias):
-        qinput = torch._make_per_tensor_quantized_tensor(input, 0.01, -25)
-        qinput = torch.dequantize(qinput)
-        qweight = torch._make_per_tensor_quantized_tensor(weight, 0.01, 50)
-        qweight = torch.dequantize(qweight)
-        qbias = torch.quantize_per_tensor(bias, 0.0001, 0, torch.qint32)
-        qbias = torch.dequantize(qbias)
-        qz = torch.ops.aten.convolution(
-            qinput,
-            qweight,
-            bias=qbias,
+        input = torch.ops.quantized_decomposed.dequantize_per_tensor.default(
+            input, 0.01, -25, -128, 127, torch.int8
+        )
+        weight = torch.ops.quantized_decomposed.dequantize_per_tensor.default(
+            weight, 0.01, 50, -128, 127, torch.int8
+        )
+
+        res = torch.ops.aten.convolution(
+            input,
+            weight,
+            bias=bias,
             stride=[2, 1],
             padding=[1, 1],
             dilation=[1, 1],
             transposed=True,
             output_padding=[0, 0],
             groups=1,
         )
-        return qz
+
+        # Use int32 to avoid overflows
+        return torch.ops.quantized_decomposed.quantize_per_tensor.default(
+            res, 1, 0, -(2**31), 2**31 - 1, torch.int32
+        )
 
 
 @register_test_case(module_factory=lambda: ConvTranspose2DQInt8Module())
@@ -1459,18 +1474,14 @@ def __init__(self, groups=1):
         super().__init__()
 
     def _forward(self, inputVec, weight, scales, zeropoints, bias):
-        inputVec = torch._make_per_tensor_quantized_tensor(inputVec, 0.01, 7)
-        inputVec = torch.dequantize(inputVec)
-
-        weight = torch._make_per_channel_quantized_tensor(
-            weight, scales, zeropoints, axis=0
+        inputVec = torch.ops.quantized_decomposed.dequantize_per_tensor.default(
+            inputVec, 0.01, 7, -128, 127, torch.int8
+        )
+        weight = torch.ops.quantized_decomposed.dequantize_per_channel.default(
+            weight, scales, zeropoints, 0, -128, 127, torch.int8
         )
-        weight = torch.dequantize(weight)
-
-        bias = torch.quantize_per_tensor(bias, 0.0001, 0, torch.qint32)
-        bias = torch.dequantize(bias)
 
-        return torch.ops.aten.conv2d(
+        conv = torch.ops.aten.conv2d(
             inputVec,
             weight,
             bias=bias,
@@ -1480,6 +1491,11 @@ def _forward(self, inputVec, weight, scales, zeropoints, bias):
             groups=self.groups,
         )
 
+        # Use int32 to avoid overflows
+        return torch.ops.quantized_decomposed.quantize_per_tensor.default(
+            conv, 1, 0, -(2**31), 2**31 - 1, torch.int32
+        )
+
 
 class Conv2dQInt8PerChannelModuleDyn(Conv2dQInt8PerChannelModuleBase):
     @export

python/torch_mlir/fx.py

@@ -41,7 +41,7 @@ def _module_lowering(
     option_string = "{extra-library=" + extra_library_file_name + "}"
     run_pipeline_with_repro_report(
         torch_mod,
-        f"builtin.module(torchdynamo-export-to-torch-backend-pipeline{option_string})",
+        f"builtin.module(func.func(torch-match-quantized-custom-ops), torchdynamo-export-to-torch-backend-pipeline{option_string})",
         "Lowering TorchFX IR -> Torch Backend IR",
         enable_ir_printing=verbose,
     )