[inductor] fix linear_add_bias path (#127597)

Previous the `linear_add_bias` path do not work.
This PR is to fix it and add more ut with it.

**TestPlan**
```
python test/inductor/test_mkldnn_pattern_matcher.py -k test_linear_add_bias
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/127597
Approved by: https://github.com/jgong5, https://github.com/jansel

test/inductor/test_mkldnn_pattern_matcher.py

@@ -396,6 +396,39 @@ class TestPatternMatcher(TestPatternMatcherBase):
             matcher_nodes = 1
             self._test_common(mod, (v,), matcher_count, matcher_nodes)
 
+    def test_linear_add_bias(self):
+        class M(torch.nn.Module):
+            def __init__(self, dtype, unary_fn):
+                super().__init__()
+                self.linear = torch.nn.Linear(10, 64, bias=False)
+                self.bias = torch.randn(64).to(dtype=dtype)
+                self.unary_fn = unary_fn
+
+            def forward(self, x):
+                x = self.linear(x) + self.bias
+                return self.unary_fn(x)
+
+        dtypes = []
+        if torch.ops.mkldnn._is_mkldnn_bf16_supported():
+            dtypes.append(torch.bfloat16)
+        if torch.ops.mkldnn._is_mkldnn_fp16_supported():
+            dtypes.append(torch.float16)
+        options = itertools.product(unary_list, dtypes)
+        for unary_fn, dtype in options:
+            metrics.reset()
+            mod = M(dtype, unary_fn).eval()
+            v = torch.randn(2, 10)
+            matcher_count = 3
+            # Add 1 for weight packing pass, add 2 for bias folding pass.
+            matcher_nodes = unary_list[unary_fn] + 3
+            if self._check_unary_is_decomposed(unary_fn):
+                # Has extra dtype conversion nodes for autocast.
+                matcher_nodes += 2
+            self._test_common(
+                mod, (v,), matcher_count, matcher_nodes, check_autocast=dtype
+            )
+            self.assertEqual(metrics.generated_kernel_count, 1)
+
     @skipIfNoDynamoSupport
     @skipIfNoONEDNN
     @skipIfRocm

torch/_inductor/fx_passes/mkldnn_fusion.py

@@ -788,14 +788,22 @@ if torch._C._has_mkldnn:
         def is_linear_add_bias(match):
             add_node = match.output_node()
             linear_node = add_node.args[0]
-            weight_meta = linear_node.args[1].meta.get("val")
+            packed_weight_node = linear_node.args[1]
+            assert packed_weight_node.name == "_reorder_linear_weight"
+            transpose_weight_node = packed_weight_node.args[0]
+            assert transpose_weight_node.name == "permute_default"
+            weight_meta = transpose_weight_node.args[0].meta.get("val")
+            bias_node = add_node.args[1]
+            if isinstance(bias_node, int):
+                # we only folding bias if it is a constant
+                return False
             bias_meta = add_node.args[1].meta.get("val")
             if weight_meta is None or bias_meta is None:
                 return False
             return (
                 linear_node.args[2] is None
                 and bias_meta.dim() == 1
-                and bias_meta.size(0) == weight_meta.size(0)
+                and bias_meta.size(0) == weight_meta.size(1)
             )
 
         # convert linear+bias to a single linear for applying fusion path.