full_like to full decomposition moving to decomposition.py for dynamic case

Author: apbose

py/torch_tensorrt/dynamo/lowering/_decompositions.py

@@ -549,6 +549,18 @@ def scaled_dot_product_cudnn_attention_decomposition(
     return attn, None, None, None, 0, 0, None, None, None
 
 
+@register_torch_trt_decomposition(
+    torch.ops.aten.full_like, registry=TORCH_TRT_DECOMPOSITIONS
+)
+def full_like_decomposition(*args, **kwargs) -> torch.Tensor:
+    input = args[0]
+    shape = args[0].shape
+    fill_value = args[1]
+    kwargs["dtype"] = input.dtype
+    kwargs["device"] = to_torch_device(default_device())
+    return torch.full(shape, fill_value, dtype=kwargs["dtype"], device=kwargs["device"])
+
+
 def get_decompositions(
     enable_experimental_decompositions: bool = False,
 ) -> Dict[OpOverload, Callable[[Any], Any]]:

py/torch_tensorrt/dynamo/lowering/passes/_aten_lowering_pass.py

@@ -13,7 +13,6 @@
 from .remove_detach import remove_detach
 from .remove_input_alias_fixing_clones import remove_input_alias_fixing_clones
 from .repair_input_as_output import repair_input_as_output
-from .replace_full_like_with_full import replace_full_like_with_full
 from .replace_max_pool_with_indices import replace_max_pool_with_indices
 from .view_to_reshape import view_to_reshape
 
@@ -25,7 +24,6 @@
         lower_linear,
         fuse_prims_broadcast,
         replace_max_pool_with_indices,
-        replace_full_like_with_full,
         view_to_reshape,
         remove_assert_scalar,
         accumulate_fp32_matmul,

py/torch_tensorrt/dynamo/lowering/passes/replace_full_like_with_full.py

@@ -434,11 +434,13 @@ def __init__(self, *args, **kwargs) -> None:
                 super().__init__(*args, **kwargs)
 
             def forward(self, x):
-                y = torch.full_like(x, 2.0)
-                return y
+                c = torch.ops.aten.add(x, x)
+                y = torch.ops.aten.full_like.default(c, 2)
+                d = y + c
+                return d
 
         # Operations expected to be removed in the traced graph after decompositions
-        expected_ops = {torch.ops.aten.full.default}
+        expected_ops = {torch.ops.aten.add.Tensor}
         unexpected_ops = {torch.ops.aten.full_like.default}
 
         inputs = [torch.randn(3, 3, dtype=torch.float32).cuda()]