Ensure we can call custom ops from torch cadence lib (#14034)

Summary:

Fixes mismatches between op registration names and implementation names, fixes some type issues in tests where unexpected types are passed in given the op definition. Also fixes an incorrect layernorm meta op (normalized_shape should be list, not int). Tests corrected as well.

Tests now use the torch cadence custom op library.

Reviewed By: hsharma35

Differential Revision: D81738196

Author: Andrew Grebenisan

backends/cadence/aot/TARGETS

@@ -614,6 +614,7 @@ python_unittest(
     typing = True,
     deps = [
         ":typing_stubs",
+        "//executorch/backends/cadence/aot:ops_registrations",
         "//executorch/backends/cadence/aot:ref_implementations",
         "//caffe2:torch",
     ]

backends/cadence/aot/ops_registrations.py

@@ -1449,7 +1449,7 @@ def quantized_layer_norm_meta(
     input: torch.Tensor,
     X_scale: torch.Tensor,
     X_zero_point: torch.Tensor,
-    normalized_shape: int,
+    normalized_shape: list[int],
     weight: torch.Tensor,
     bias: torch.Tensor,
     eps: float,
@@ -1464,7 +1464,7 @@ def quantized_layer_norm_per_tensor_meta(
     input: torch.Tensor,
     X_scale: float,
     X_zero_point: int,
-    normalized_shape: int,
+    normalized_shape: list[int],
     weight: torch.Tensor,
     bias: torch.Tensor,
     eps: float,

backends/cadence/aot/ref_implementations.py

@@ -64,9 +64,9 @@ def quantize_per_tensor(
             f"Unsupported dtype to quantize to. Supported dtypes must be one of {supported_quant_types}"
         )
 
-    dequantized = torch.round(input_tensor * scale + zero_point).to(dtype)
+    quantized = torch.round(input_tensor * scale + zero_point).to(dtype)
     return torch.max(
-        torch.min(dequantized, torch.tensor(quant_max)),
+        torch.min(quantized, torch.tensor(quant_max)),
         torch.tensor(quant_min),
     )
 
@@ -247,12 +247,12 @@ def quantized_linear(
     ).reshape(*leading_dims, N)
 
 
-@impl(m, "quantized_layer_norm_per_tensor")
+@impl(m, "quantized_layer_norm.per_tensor")
 def quantized_layer_norm_per_tensor(
     input_tensor: torch.Tensor,
     X_scale: float,
     X_zero_point: int,
-    normalized_shape: int,
+    normalized_shape: list[int],
     weight: torch.Tensor,
     bias: torch.Tensor,
     eps: float,
@@ -283,7 +283,7 @@ def quantized_layer_norm_per_tensor(
         input_tensor, X_scale, X_zero_point, -128, 127, torch.float32
     )
     out = torch.nn.functional.layer_norm(
-        float_input_tensor, (normalized_shape,), weight, bias, eps=eps
+        float_input_tensor, normalized_shape, weight, bias, eps=eps
     )
 
     return quantize_per_tensor(
@@ -365,7 +365,7 @@ def quantized_conv_per_tensor(
     )
 
 
-@impl(m, "quantized_conv_nchw_per_tensor")
+@impl(m, "quantized_conv_nchw.per_tensor")
 def quantized_conv_nchw_per_tensor(
     input_tensor: torch.Tensor,
     weight: torch.Tensor,
@@ -421,7 +421,7 @@ def quantized_conv_nchw_per_tensor(
     )
 
 
-@impl(m, "quantized_conv_nhwc_per_tensor")
+@impl(m, "quantized_conv_nhwc.per_tensor")
 def quantized_conv_nhwc_per_tensor(
     input_tensor: torch.Tensor,
     weight: torch.Tensor,
@@ -558,62 +558,62 @@ def variant(
     return decorator
 
 
-@impl(m, "quantized_conv_nchw_asym8sxsym8s_asym8s_per_tensor")
+@impl(m, "quantized_conv_nchw_asym8sxsym8s_asym8s.per_tensor")
 @quantized_conv_variant("nchw", torch.int8, torch.int8)
 def quantized_conv_nchw_asym8sxsym8s_asym8s_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nchw_asym8uxsym8u_asym8u_per_tensor")
+@impl(m, "quantized_conv_nchw_asym8uxsym8u_asym8u.per_tensor")
 @quantized_conv_variant("nchw", torch.uint8, torch.uint8)
 def quantized_conv_nchw_asym8uxsym8u_asym8u_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nhwc_asym8sxsym8s_asym8s_per_tensor")
+@impl(m, "quantized_conv_nhwc_asym8sxsym8s_asym8s.per_tensor")
 @quantized_conv_variant("nhwc", torch.int8, torch.int8)
 def quantized_conv_nhwc_asym8sxsym8s_asym8s_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nhwc_asym8uxsym8u_asym8u_per_tensor")
+@impl(m, "quantized_conv_nhwc_asym8uxsym8u_asym8u.per_tensor")
 @quantized_conv_variant("nhwc", torch.uint8, torch.uint8)
 def quantized_conv_nhwc_asym8uxsym8u_asym8u_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nchw_dilated_asym8sxsym8s_asym8s_per_tensor")
+@impl(m, "quantized_conv_nchw_dilated_asym8sxsym8s_asym8s.per_tensor")
 @quantized_conv_variant("nchw", torch.int8, torch.int8)
 def quantized_conv_nchw_dilated_asym8sxsym8s_asym8s_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nchw_dilated_asym8uxsym8u_asym8u_per_tensor")
+@impl(m, "quantized_conv_nchw_dilated_asym8uxsym8u_asym8u.per_tensor")
 @quantized_conv_variant("nchw", torch.uint8, torch.uint8)
 def quantized_conv_nchw_dilated_asym8uxsym8u_asym8u_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s_per_tensor")
+@impl(m, "quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s.per_tensor")
 @quantized_conv_variant("nhwc", torch.int8, torch.int8)
 def quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u_per_tensor")
+@impl(m, "quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u.per_tensor")
 @quantized_conv_variant("nhwc", torch.uint8, torch.uint8)
 def quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nchw_depthwise_asym8sxsym8s_asym8s_per_tensor")
+@impl(m, "quantized_conv_nchw_depthwise_asym8sxsym8s_asym8s.per_tensor")
 @quantized_conv_variant("nchw", torch.int8, torch.int8)
 def quantized_conv_nchw_depthwise_asym8sxsym8s_asym8s_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nchw_depthwise_asym8uxsym8u_asym8u_per_tensor")
+@impl(m, "quantized_conv_nchw_depthwise_asym8uxsym8u_asym8u.per_tensor")
 @quantized_conv_variant("nchw", torch.uint8, torch.uint8)
 def quantized_conv_nchw_depthwise_asym8uxsym8u_asym8u_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nhwc_depthwise_asym8sxsym8s_asym8s_per_tensor")
+@impl(m, "quantized_conv_nhwc_depthwise_asym8sxsym8s_asym8s.per_tensor")
 @quantized_conv_variant("nhwc", torch.int8, torch.int8)
 def quantized_conv_nhwc_depthwise_asym8sxsym8s_asym8s_per_tensor() -> torch.Tensor: ...
 
 
-@impl(m, "quantized_conv_nhwc_depthwise_asym8uxsym8u_asym8u_per_tensor")
+@impl(m, "quantized_conv_nhwc_depthwise_asym8uxsym8u_asym8u.per_tensor")
 @quantized_conv_variant("nhwc", torch.uint8, torch.uint8)
 def quantized_conv_nhwc_depthwise_asym8uxsym8u_asym8u_per_tensor() -> torch.Tensor: ...