Update

[ghstack-poisoned]

test/dynamo/test_activation_checkpointing.py

@@ -1852,14 +1852,13 @@ class GraphModule(torch.nn.Module):
 
         wrap_body_0 = self.wrap_body_0
         tag_activation_checkpoint = torch.ops.higher_order.tag_activation_checkpoint(wrap_body_0, l_x_, use_reentrant = True);  wrap_body_0 = l_x_ = None
-        getitem: "f32[4, 4]" = tag_activation_checkpoint[0]
-        getitem_1: "f32[4, 4]" = tag_activation_checkpoint[1];  tag_activation_checkpoint = None
-        return (getitem, getitem_1)
+        getitem: "f32[4, 4]" = tag_activation_checkpoint[0];  tag_activation_checkpoint = None
+        return (getitem,)
 
     class wrap_body_0(torch.nn.Module):
         def forward(self, l_x_: "f32[4, 4]"):
             y: "f32[4, 4]" = torch.sin(l_x_);  l_x_ = None
-            return (y, y)
+            return (y,)
 """,
         )
 
@@ -1969,9 +1968,9 @@ class GraphModule(torch.nn.Module):
             out: "f32[4, 4]" = l_x_.sin()
 
             sin_1: "f32[4, 4]" = torch.sin(o)
-            child: "f32[4, 4]" = torch.cos(sin_1)
-            child_1: "f32[4, 4]" = torch.sin(l_x_);  l_x_ = None
-            return (child, child_1, matmul, o, out, sin_1)
+            cos: "f32[4, 4]" = torch.cos(sin_1)
+            sin_2: "f32[4, 4]" = torch.sin(l_x_);  l_x_ = None
+            return (cos, sin_2, matmul, o, out, sin_1)
 """,
         )
 

test/dynamo/test_base_hop.py

@@ -222,13 +222,13 @@ class GraphModule(torch.nn.Module):
 
             matmul: "f32[3, 3]" = l_x_ @ l_y_
             sin: "f32[3, 3]" = matmul.sin();  matmul = None
-            child: "f32[3, 3]" = sin.cos();  sin = None
+            cos: "f32[3, 3]" = sin.cos();  sin = None
 
-            child_1: "f32[3, 3]" = l_x_ + l_y_
-            child_2: "f32[3, 3]" = l_x_ - l_y_
+            add: "f32[3, 3]" = l_x_ + l_y_
+            sub: "f32[3, 3]" = l_x_ - l_y_
 
-            child_3: "f32[3, 3]" = l_x_ @ l_y_;  l_x_ = l_y_ = None
-            return (child, child_1, child_2, child_3)
+            matmul_1: "f32[3, 3]" = l_x_ @ l_y_;  l_x_ = l_y_ = None
+            return (cos, add, sub, matmul_1)
 """,  # noqa: B950
         )
         self.assertExpectedInline(

test/dynamo/test_higher_order_ops.py

@@ -249,7 +249,7 @@ class HigherOrderOpTests(torch._dynamo.test_case.TestCase):
 
         # when testing with dynamic shape, symbols are lifted as input
         arg_count = ifdynstaticdefault(2, 3)
-        self._test_wrap_simple(fn, default_args_generator((x,)), arg_count)
+        self._test_wrap_simple(fn, default_args_generator((x,)), arg_count, 1)
 
     def test_return_captured_vars(self):
         freevar1 = torch.randn(3)
@@ -267,7 +267,7 @@ class HigherOrderOpTests(torch._dynamo.test_case.TestCase):
         # be the input.
         # when testing with dynamic shape, a symbol is lifted as input
         arg_count = ifdynstaticdefault(3, 4)
-        self._test_wrap_simple(fn, default_args_generator((x,)), arg_count, 4)
+        self._test_wrap_simple(fn, default_args_generator((x,)), arg_count, 1)
 
     def test_return_captured_var_used_multiple_times(self):
         freevar = torch.randn(3)
@@ -282,7 +282,7 @@ class HigherOrderOpTests(torch._dynamo.test_case.TestCase):
         x = torch.randn(3)
         # when testing with dynamic shape, a symbol is lifted as input
         arg_count = ifdynstaticdefault(3, 4)
-        self._test_wrap_simple(fn, default_args_generator((x,)), arg_count, 3)
+        self._test_wrap_simple(fn, default_args_generator((x,)), arg_count, 2)
 
     def test_capture_untracked_global(self):
         def f(x):
@@ -762,15 +762,15 @@ class GraphModule(torch.nn.Module):
         def forward(self, s77: "Sym(s77)", l_x_: "f32[s77]", u0: "Sym(u0)", c: "i64[u0, 1]"):
             wrap_body_0 = self.wrap_body_0
             wrap = torch.ops.higher_order.wrap(wrap_body_0, s77, l_x_, u0, c);  wrap_body_0 = s77 = l_x_ = u0 = c = None
-            child: "f32[s77]" = wrap[0]
-            child_1: "f32[u0, 1]" = wrap[1];  wrap = None
-            return (child, child_1)
+            getitem: "f32[s77]" = wrap[0]
+            getitem_1: "f32[u0, 1]" = wrap[1];  wrap = None
+            return (getitem, getitem_1)
 
         class wrap_body_0(torch.nn.Module):
             def forward(self, s77: "Sym(s77)", l_x_: "f32[s77]", u0: "Sym(u0)", c: "i64[u0, 1]"):
-                child: "f32[s77]" = l_x_.sin();  l_x_ = None
-                child_1: "f32[u0, 1]" = c.sin();  c = None
-                return (child, child_1)
+                sin: "f32[s77]" = l_x_.sin();  l_x_ = None
+                sin_1: "f32[u0, 1]" = c.sin();  c = None
+                return (sin, sin_1)
 """,
             )
         else:
@@ -801,15 +801,15 @@ class GraphModule(torch.nn.Module):
         def forward(self, l_x_: "f32[3]", u0: "Sym(u0)", c: "i64[u0, 1]"):
             wrap_body_0 = self.wrap_body_0
             wrap = torch.ops.higher_order.wrap(wrap_body_0, l_x_, u0, c);  wrap_body_0 = l_x_ = u0 = c = None
-            child: "f32[3]" = wrap[0]
-            child_1: "f32[u0, 1]" = wrap[1];  wrap = None
-            return (child, child_1)
+            getitem: "f32[3]" = wrap[0]
+            getitem_1: "f32[u0, 1]" = wrap[1];  wrap = None
+            return (getitem, getitem_1)
 
         class wrap_body_0(torch.nn.Module):
             def forward(self, l_x_: "f32[3]", u0: "Sym(u0)", c: "i64[u0, 1]"):
-                child: "f32[3]" = l_x_.sin();  l_x_ = None
-                child_1: "f32[u0, 1]" = c.sin();  c = None
-                return (child, child_1)
+                sin: "f32[3]" = l_x_.sin();  l_x_ = None
+                sin_1: "f32[u0, 1]" = c.sin();  c = None
+                return (sin, sin_1)
 """,
             )
 
@@ -922,16 +922,16 @@ class GraphModule(torch.nn.Module):
         def forward(self, l_x_: "f32[3]", size: "Sym(u0)", c: "i64[u0, 1]"):
             wrap_body_0 = self.wrap_body_0
             wrap = torch.ops.higher_order.wrap(wrap_body_0, l_x_, size, c);  wrap_body_0 = l_x_ = size = c = None
-            child: "f32[3]" = wrap[0]
-            child_1: "f32[u0, 1]" = wrap[1];  wrap = None
-            return (child, child_1)
+            getitem: "f32[3]" = wrap[0]
+            getitem_1: "f32[u0, 1]" = wrap[1];  wrap = None
+            return (getitem, getitem_1)
 
         class wrap_body_0(torch.nn.Module):
             def forward(self, l_x_: "f32[3]", size: "Sym(u0)", c: "i64[u0, 1]"):
                 sin: "f32[3]" = l_x_.sin();  l_x_ = None
-                child: "f32[3]" = sin + size;  sin = size = None
-                child_1: "f32[u0, 1]" = c.sin();  c = None
-                return (child, child_1)
+                add: "f32[3]" = sin + size;  sin = size = None
+                sin_1: "f32[u0, 1]" = c.sin();  c = None
+                return (add, sin_1)
 """,
             )
 
@@ -2458,10 +2458,10 @@ class GraphModule(torch.nn.Module):
 
     class wrap_body_0(torch.nn.Module):
         def forward(self, l_arg1_0_: "f32[3]", l_arg2_0_: "f32[3]"):
-            child: "f32[3]" = l_arg1_0_ + 1;  l_arg1_0_ = None
+            add: "f32[3]" = l_arg1_0_ + 1;  l_arg1_0_ = None
 
-            child_1: "f32[3]" = l_arg2_0_ + 1;  l_arg2_0_ = None
-            return (child, child_1)
+            add_1: "f32[3]" = l_arg2_0_ + 1;  l_arg2_0_ = None
+            return (add, add_1)
 """,
         )
 
@@ -2655,9 +2655,9 @@ class GraphModule(torch.nn.Module):
 
     class wrap_body_0(torch.nn.Module):
         def forward(self, l_x_: "f32[2, 3]"):
-            child: "f32[2, 3]" = l_x_.sin()
-            child_1: "f32[2, 3]" = l_x_.cos();  l_x_ = None
-            return (child, child_1)
+            sin: "f32[2, 3]" = l_x_.sin()
+            cos: "f32[2, 3]" = l_x_.cos();  l_x_ = None
+            return (sin, cos)
 """,
         )
 
@@ -2687,13 +2687,13 @@ class GraphModule(torch.nn.Module):
 
         wrap_body_0 = self.wrap_body_0
         wrap = torch.ops.higher_order.wrap(wrap_body_0, l_x_);  wrap_body_0 = l_x_ = None
-        value: "f32[3]" = wrap[0];  wrap = None
-        return (value,)
+        getitem: "f32[3]" = wrap[0];  wrap = None
+        return (getitem,)
 
     class wrap_body_0(torch.nn.Module):
         def forward(self, l_x_: "f32[3]"):
-            child: "f32[3]" = -l_x_;  l_x_ = None
-            return (child,)
+            neg: "f32[3]" = -l_x_;  l_x_ = None
+            return (neg,)
 """,
         )
 

test/higher_order_ops/test_invoke_subgraph.py

@@ -899,14 +899,14 @@ class GraphModule(torch.nn.Module):
     class subgraph_0(torch.nn.Module):
         def forward(self, l_x_: "f32[8]", l_y_: "f32[8]"):
             mul: "f32[8]" = torch.mul(l_x_, l_y_);  l_x_ = l_y_ = None
-            child: "f32[8]" = mul * 2;  mul = None
-            return (child,)
+            mul_1: "f32[8]" = mul * 2;  mul = None
+            return (mul_1,)
 
     class subgraph_1(torch.nn.Module):
         def forward(self, a: "f32[8]", l_y_: "f32[8]"):
             mul: "f32[8]" = torch.mul(a, l_y_);  a = l_y_ = None
-            child: "f32[8]" = mul * 3;  mul = None
-            return (child,)
+            mul_1: "f32[8]" = mul * 3;  mul = None
+            return (mul_1,)
 """,
             )
 
@@ -983,20 +983,20 @@ class GraphModule(torch.nn.Module):
 
         subgraph_0 = self.subgraph_0
         invoke_subgraph = torch.ops.higher_order.invoke_subgraph(subgraph_0, 'subgraph_0', l_x_, l_y_);  subgraph_0 = l_x_ = None
-        x: "f32[8]" = invoke_subgraph[0];  invoke_subgraph = None
+        getitem: "f32[8]" = invoke_subgraph[0];  invoke_subgraph = None
         subgraph_1 = self.subgraph_0
-        invoke_subgraph_1 = torch.ops.higher_order.invoke_subgraph(subgraph_1, 'subgraph_0', x, l_y_);  subgraph_1 = x = None
-        x_1: "f32[8]" = invoke_subgraph_1[0];  invoke_subgraph_1 = None
+        invoke_subgraph_1 = torch.ops.higher_order.invoke_subgraph(subgraph_1, 'subgraph_0', getitem, l_y_);  subgraph_1 = getitem = None
+        getitem_1: "f32[8]" = invoke_subgraph_1[0];  invoke_subgraph_1 = None
         subgraph_2 = self.subgraph_0
-        invoke_subgraph_2 = torch.ops.higher_order.invoke_subgraph(subgraph_2, 'subgraph_0', x_1, l_y_);  subgraph_2 = x_1 = None
-        x_2: "f32[8]" = invoke_subgraph_2[0];  invoke_subgraph_2 = None
+        invoke_subgraph_2 = torch.ops.higher_order.invoke_subgraph(subgraph_2, 'subgraph_0', getitem_1, l_y_);  subgraph_2 = getitem_1 = None
+        getitem_2: "f32[8]" = invoke_subgraph_2[0];  invoke_subgraph_2 = None
         subgraph_3 = self.subgraph_0
-        invoke_subgraph_3 = torch.ops.higher_order.invoke_subgraph(subgraph_3, 'subgraph_0', x_2, l_y_);  subgraph_3 = x_2 = None
-        x_3: "f32[8]" = invoke_subgraph_3[0];  invoke_subgraph_3 = None
+        invoke_subgraph_3 = torch.ops.higher_order.invoke_subgraph(subgraph_3, 'subgraph_0', getitem_2, l_y_);  subgraph_3 = getitem_2 = None
+        getitem_3: "f32[8]" = invoke_subgraph_3[0];  invoke_subgraph_3 = None
         subgraph_4 = self.subgraph_0
-        invoke_subgraph_4 = torch.ops.higher_order.invoke_subgraph(subgraph_4, 'subgraph_0', x_3, l_y_);  subgraph_4 = x_3 = l_y_ = None
-        x_4: "f32[8]" = invoke_subgraph_4[0];  invoke_subgraph_4 = None
-        return (x_4,)
+        invoke_subgraph_4 = torch.ops.higher_order.invoke_subgraph(subgraph_4, 'subgraph_0', getitem_3, l_y_);  subgraph_4 = getitem_3 = l_y_ = None
+        getitem_4: "f32[8]" = invoke_subgraph_4[0];  invoke_subgraph_4 = None
+        return (getitem_4,)
 
     class subgraph_0(torch.nn.Module):
         def forward(self, l_x_: "f32[8]", l_y_: "f32[8]"):
@@ -1495,9 +1495,9 @@ class GraphModule(torch.nn.Module):
 
     class subgraph_0(torch.nn.Module):
         def forward(self, l_x_: "f32[8, 8]"):
-            child: "f32[8, 8]" = l_x_ * 2
-            child_1: "f32[8, 8]" = l_x_ * 3;  l_x_ = None
-            return (child, child_1)
+            mul: "f32[8, 8]" = l_x_ * 2
+            mul_1: "f32[8, 8]" = l_x_ * 3;  l_x_ = None
+            return (mul, mul_1)
 """,
             )
 

test/higher_order_ops/test_local_map.py

@@ -286,47 +286,31 @@ class GraphModule(torch.nn.Module):
         l_self_modules_wo_parameters_weight_ = L_self_modules_wo_parameters_weight_
         l_self_modules_w1_parameters_weight_ = L_self_modules_w1_parameters_weight_
         l_self_modules_w2_parameters_weight_ = L_self_modules_w2_parameters_weight_
-
         q: "f32[8, 16, 96]" = torch._C._nn.linear(l_x_, l_self_modules_wq_parameters_weight_, None);  l_self_modules_wq_parameters_weight_ = None
-
         k: "f32[8, 16, 96]" = torch._C._nn.linear(l_x_, l_self_modules_wk_parameters_weight_, None);  l_self_modules_wk_parameters_weight_ = None
-
         v: "f32[8, 16, 96]" = torch._C._nn.linear(l_x_, l_self_modules_wv_parameters_weight_, None);  l_self_modules_wv_parameters_weight_ = None
-
         unflatten: "f32[8, 16, 16, 6]" = q.unflatten(-1, (16, -1));  q = None
         q_1: "f32[8, 16, 16, 6]" = unflatten.permute(0, 2, 1, 3);  unflatten = None
-
         unflatten_1: "f32[8, 16, 16, 6]" = k.unflatten(-1, (16, -1));  k = None
         k_1: "f32[8, 16, 16, 6]" = unflatten_1.permute(0, 2, 1, 3);  unflatten_1 = None
-
         unflatten_2: "f32[8, 16, 16, 6]" = v.unflatten(-1, (16, -1));  v = None
         v_1: "f32[8, 16, 16, 6]" = unflatten_2.permute(0, 2, 1, 3);  unflatten_2 = None
-
         subgraph_0 = self.subgraph_0
         local_map_hop = torch.ops.higher_order.local_map_hop(subgraph_0, q_1, k_1, v_1);  subgraph_0 = q_1 = k_1 = v_1 = None
-        o: "f32[8, 16, 16, 6]" = local_map_hop[0];  local_map_hop = None
-
-        permute_3: "f32[8, 16, 16, 6]" = o.permute(0, 2, 1, 3);  o = None
-        o_1: "f32[8, 16, 96]" = permute_3.flatten(-2);  permute_3 = None
-
-        o_2: "f32[8, 16, 96]" = torch._C._nn.linear(o_1, l_self_modules_wo_parameters_weight_, None);  o_1 = l_self_modules_wo_parameters_weight_ = None
-
-        o0: "f32[8, 16, 96]" = o_2 + l_x_;  o_2 = l_x_ = None
-
-        o_3: "f32[8, 16, 384]" = torch._C._nn.linear(o0, l_self_modules_w1_parameters_weight_, None);  l_self_modules_w1_parameters_weight_ = None
-
-        o_4: "f32[8, 16, 384]" = torch.nn.functional.relu(o_3);  o_3 = None
-
-        o_5: "f32[8, 16, 96]" = torch._C._nn.linear(o_4, l_self_modules_w2_parameters_weight_, None);  o_4 = l_self_modules_w2_parameters_weight_ = None
-
-        o_6: "f32[8, 16, 96]" = o0 + o_5;  o0 = o_5 = None
-        return (o_6,)
-
+        getitem: "f32[8, 16, 16, 6]" = local_map_hop[0];  local_map_hop = None
+        permute_3: "f32[8, 16, 16, 6]" = getitem.permute(0, 2, 1, 3);  getitem = None
+        o: "f32[8, 16, 96]" = permute_3.flatten(-2);  permute_3 = None
+        o_1: "f32[8, 16, 96]" = torch._C._nn.linear(o, l_self_modules_wo_parameters_weight_, None);  o = l_self_modules_wo_parameters_weight_ = None
+        o0: "f32[8, 16, 96]" = o_1 + l_x_;  o_1 = l_x_ = None
+        o_2: "f32[8, 16, 384]" = torch._C._nn.linear(o0, l_self_modules_w1_parameters_weight_, None);  l_self_modules_w1_parameters_weight_ = None
+        o_3: "f32[8, 16, 384]" = torch.nn.functional.relu(o_2);  o_2 = None
+        o_4: "f32[8, 16, 96]" = torch._C._nn.linear(o_3, l_self_modules_w2_parameters_weight_, None);  o_3 = l_self_modules_w2_parameters_weight_ = None
+        o_5: "f32[8, 16, 96]" = o0 + o_4;  o0 = o_4 = None
+        return (o_5,)
     class subgraph_0(torch.nn.Module):
         def forward(self, q_1: "f32[1, 2, 4, 6]", k_1: "f32[1, 2, 16, 6]", v_1: "f32[1, 2, 16, 6]"):
             out: "f32[1, 2, 4, 6]" = torch._C._nn.scaled_dot_product_attention(query = q_1, key = k_1, value = v_1, is_causal = False);  q_1 = k_1 = v_1 = None
-            return (out,)
-""",
+            return (out,)""",
                 ignore_empty_lines=True,
             )