[Inductor][float8] Register qconv-binary fusion pass for float8

Author: jiayisunx

test/quantization/pt2e/test_x86inductor_fusion.py

@@ -770,7 +770,7 @@ def test_qconv2d_silu_fp8_mixed_bf16_cpu(self):
         )
 
     def _qconv2d_add_test_helper(
-        self, device="cpu", use_relu=False, int8_mixed_bf16=False
+        self, device="cpu", use_relu=False, mixed_bf16=False, is_fp8=False
     ):
         r"""
         This testcase will quantize a Conv2d->Add pattern as:
@@ -844,11 +844,12 @@ def matcher_check_fn():
                 (v,),
                 matcher_check_fn,
                 check_quantization=True,
-                check_autocast=torch.bfloat16 if int8_mixed_bf16 else torch.float32,
+                check_autocast=torch.bfloat16 if mixed_bf16 else torch.float32,
+                is_fp8=is_fp8,
             )
 
     def _qconv2d_add_test_helper2(
-        self, device="cpu", use_relu=False, int8_mixed_bf16=False
+        self, device="cpu", use_relu=False, mixed_bf16=False, is_fp8=False
     ):
         r"""
         This testcase will quantize two Conv2d->Add patterns as:
@@ -907,8 +908,11 @@ def forward(self, x, x2, x3):
                     res = self.relu2(res)
                 return res
 
+        add_fn_list = quantization_add_fn_list
+        if not is_fp8:
+            add_fn_list = add_fn_list + quantization_inplace_add_fn_list
         for add_fn, swap_inputs in itertools.product(
-            quantization_add_fn_list + quantization_inplace_add_fn_list, [False, True]
+            add_fn_list, [False, True]
         ):
             mod = M(add_fn, use_relu, swap_inputs).eval().to(device=device)
             x = torch.randn(
@@ -941,7 +945,8 @@ def matcher_check_fn():
                 (x, x2, x3),
                 matcher_check_fn,
                 check_quantization=True,
-                check_autocast=torch.bfloat16 if int8_mixed_bf16 else torch.float32,
+                check_autocast=torch.bfloat16 if mixed_bf16 else torch.float32,
+                is_fp8=is_fp8,
             )
 
     @skipIfNoDynamoSupport
@@ -950,25 +955,55 @@ def test_qconv2d_add_cpu(self):
         self._qconv2d_add_test_helper()
         self._qconv2d_add_test_helper2()
 
+    @skipIfNoDynamoSupport
+    @skipIfNoONEDNN
+    @skipIfNoFloat8Support
+    def test_qconv2d_add_fp8_cpu(self):
+        self._qconv2d_add_test_helper(is_fp8=True)
+        self._qconv2d_add_test_helper2(is_fp8=True)
+
     @skipIfNoDynamoSupport
     @skipIfNoONEDNNBF16
     @skipIfNoONEDNN
     def test_qconv2d_add_int8_mixed_bf16(self):
-        self._qconv2d_add_test_helper(int8_mixed_bf16=True)
-        self._qconv2d_add_test_helper2(int8_mixed_bf16=True)
+        self._qconv2d_add_test_helper(mixed_bf16=True)
+        self._qconv2d_add_test_helper2(mixed_bf16=True)
+
+    @skipIfNoDynamoSupport
+    @skipIfNoONEDNNBF16
+    @skipIfNoONEDNN
+    @skipIfNoFloat8Support
+    def test_qconv2d_add_fp8_mixed_bf16(self):
+        self._qconv2d_add_test_helper(mixed_bf16=True, is_fp8=True)
+        self._qconv2d_add_test_helper2(mixed_bf16=True, is_fp8=True)
 
     @skipIfNoDynamoSupport
     @skipIfNoONEDNN
     def test_qconv2d_add_relu_cpu(self):
         self._qconv2d_add_test_helper(use_relu=True)
         self._qconv2d_add_test_helper2(use_relu=True)
 
+    @skipIfNoDynamoSupport
+    @skipIfNoONEDNN
+    @skipIfNoFloat8Support
+    def test_qconv2d_add_relu_fp8_cpu(self):
+        self._qconv2d_add_test_helper(use_relu=True, is_fp8=True)
+        self._qconv2d_add_test_helper2(use_relu=True, is_fp8=True)
+
     @skipIfNoDynamoSupport
     @skipIfNoONEDNNBF16
     @skipIfNoONEDNN
     def test_qconv2d_add_relu_int8_mixed_bf16(self):
-        self._qconv2d_add_test_helper(use_relu=True, int8_mixed_bf16=True)
-        self._qconv2d_add_test_helper2(use_relu=True, int8_mixed_bf16=True)
+        self._qconv2d_add_test_helper(use_relu=True, mixed_bf16=True)
+        self._qconv2d_add_test_helper2(use_relu=True, mixed_bf16=True)
+
+    @skipIfNoDynamoSupport
+    @skipIfNoONEDNNBF16
+    @skipIfNoONEDNN
+    @skipIfNoFloat8Support
+    def test_qconv2d_add_relu_fp8_mixed_bf16(self):
+        self._qconv2d_add_test_helper(use_relu=True, mixed_bf16=True, is_fp8=True)
+        self._qconv2d_add_test_helper2(use_relu=True, mixed_bf16=True, is_fp8=True)
 
     @skipIfNoDynamoSupport
     @skipIfNoONEDNN

torchao/quantization/pt2e/inductor_passes/x86.py

@@ -726,6 +726,7 @@ def fn(match):
             return False
         binary_node_inputs = next(iter(compute_node.users)).args
         assert len(binary_node_inputs) == 2, "Expects binary node with 2 inputs"
+        is_fp8 = match.kwargs["x"].meta["val"].dtype is torch.float8_e4m3fn
         if output_dtype in [torch.float32, torch.bfloat16]:
             extra_input_of_binary_node = None
             for arg in binary_node_inputs:
@@ -734,7 +735,7 @@ def fn(match):
                     break
             assert extra_input_of_binary_node is not None
             # Extra input of binary node comes from dequant pattern
-            if extra_input_from_dequant and (
+            if not is_fp8 and extra_input_from_dequant and (
                 (not isinstance(extra_input_of_binary_node, torch.fx.Node))
                 or (
                     extra_input_of_binary_node.target
@@ -3228,37 +3229,44 @@ def _register_qconv_unary_fusion():
 
 
 def _register_qconv_binary_fusion():
-    for int8_mixed_bf16_with_inplace_add in [False, True]:
+    for int8_mixed_bf16_with_inplace_add, x_scale_zp_are_tensors in itertools.product([False, True], [False, True]):
+        qconv_binary_op = (
+            torch.ops.onednn.qconv2d_pointwise.binary_tensor
+            if x_scale_zp_are_tensors
+            else torch.ops.onednn.qconv2d_pointwise.binary
+        )
         # Priority 1 to match: QConv2d Binary or Binary-Unary pattern with int8 output
         swap_binary_inputs_list = [False, True]
         binary_replace_patterns = {}
-        for swap_inputs in swap_binary_inputs_list:
+        for swap_inputs, is_fp8 in itertools.product(swap_binary_inputs_list, [False, True]):
             binary_replace_patterns.update(
                 {
                     PostOpAttr(
                         "sum", 1.0, "none", [], ""
                     ): generate_pattern_with_output_quant(
                         generate_pattern_with_binary(
                             aten.add.Tensor,
-                            get_qconv_pt2e_pattern(users=1),
+                            get_qconv_pt2e_pattern(x_scale_zp_are_tensors, 1),
                             dequantize_accum_pattern,
                             int8_mixed_bf16_with_inplace_add,
                             swap_inputs=swap_inputs,
                         ),
+                        is_fp8=is_fp8,
                     ),
                     PostOpAttr(
                         "sum", 1.0, "relu", [], ""
                     ): generate_pattern_with_output_quant(
                         generate_pattern_with_unary(
                             generate_pattern_with_binary(
                                 aten.add.Tensor,
-                                get_qconv_pt2e_pattern(users=1),
+                                get_qconv_pt2e_pattern(x_scale_zp_are_tensors, 1),
                                 dequantize_accum_pattern,
                                 int8_mixed_bf16_with_inplace_add,
                                 swap_inputs=swap_inputs,
                             ),
                             aten.relu.default,
                         ),
+                        is_fp8=is_fp8,
                     ),
                 }
             )
@@ -3267,7 +3275,7 @@ def _register_qconv_binary_fusion():
             _register_qconv_post_op_fusion_pass(
                 patterns,
                 3,  # pass_number
-                torch.ops.onednn.qconv2d_pointwise.binary,  # computation_op
+                qconv_binary_op,  # computation_op
                 binary_unary_attr,  # binary_unary_attr
             )
 
@@ -3279,7 +3287,7 @@ def _register_qconv_binary_fusion():
                     PostOpAttr("sum", 1.0, "relu", [], ""): generate_pattern_with_unary(
                         generate_pattern_with_binary(
                             aten.add.Tensor,
-                            get_qconv_pt2e_pattern(users=1),
+                            get_qconv_pt2e_pattern(x_scale_zp_are_tensors, 1),
                             KeywordArg("accum_after_dequant"),
                             int8_mixed_bf16_with_inplace_add,
                             swap_inputs=swap_inputs,
@@ -3297,14 +3305,14 @@ def _register_qconv_binary_fusion():
                 _register_qconv_post_op_fusion_pass(
                     patterns,
                     3,  # pass_number
-                    torch.ops.onednn.qconv2d_pointwise.binary,  # computation_op
+                    qconv_binary_op,  # computation_op
                     binary_unary_attr,  # binary_unary_attr
                 )
             else:
                 _register_qconv_post_op_fusion_pass(
                     patterns,
                     4,  # pass_number
-                    torch.ops.onednn.qconv2d_pointwise.binary,  # computation_op
+                    qconv_binary_op,  # computation_op
                     binary_unary_attr,  # binary_unary_attr
                 )
 
@@ -3317,7 +3325,7 @@ def _register_qconv_binary_fusion():
                         "sum", 1.0, "none", [], ""
                     ): generate_pattern_with_binary(
                         aten.add.Tensor,
-                        get_qconv_pt2e_pattern(users=1),
+                        get_qconv_pt2e_pattern(x_scale_zp_are_tensors, 1),
                         KeywordArg("accum_after_dequant"),
                         int8_mixed_bf16_with_inplace_add,
                         swap_inputs=swap_inputs,
@@ -3332,7 +3340,7 @@ def _register_qconv_binary_fusion():
             _register_qconv_post_op_fusion_pass(
                 patterns,
                 4 if int8_mixed_bf16_with_inplace_add else 5,  # pass_number
-                torch.ops.onednn.qconv2d_pointwise.binary,  # computation_op
+                qconv_binary_op,  # computation_op
                 binary_unary_attr,  # binary_unary_attr
             )