compare prepared vs. converted outputs for Embedding (pytorch#2087)

Summary:

Fixed the embedding op and updated the test.

Reviewed By: telgamal-1, jerryzh168

Differential Revision: D73266106

Author: Naveen Suda

test/quantization/test_qat.py

@@ -808,6 +808,8 @@ def test_qat_4w_embedding(self):
             _quantized_decomposed_quantize_per_channel_group_wrapper,
         )
         from torchao.quantization.qat import Int4WeightOnlyEmbeddingQATQuantizer
+        from torchao.quantization.utils import compute_error
+
 
         group_size = 256
         model = M2()
@@ -816,9 +818,9 @@ def test_qat_4w_embedding(self):
         quantizer = Int4WeightOnlyEmbeddingQATQuantizer(group_size)
         prepared = quantizer.prepare(model)
         prepared_embedding_weight = copy.deepcopy(prepared.embedding.weight)
-        prepared(*x)
-        converted = quantizer.convert(model)
-        converted(*x)
+        prepared_output = prepared(*x)
+        converted = quantizer.convert(copy.deepcopy(prepared))
+        converted_output = converted(*x)
 
         # Assert the scales, zero points, and weights are correct after convert
         qmin, qmax = -8, 7
@@ -837,9 +839,12 @@ def test_qat_4w_embedding(self):
             torch.int8,
             group_size,
         )
+        sqnr = compute_error(prepared_output, converted_output).detach().item()
         torch.testing.assert_close(converted.embedding.weight, q_weight)
         torch.testing.assert_close(converted.embedding.scale, s)
         torch.testing.assert_close(converted.embedding.zero_point, zp)
+        torch.testing.assert_close(sqnr, float('inf'))
+
 
     def test_fake_quantize_config_granularity(self):
         """

torchao/quantization/qat/embedding.py

@@ -251,6 +251,7 @@ def _convert_helper(self, module: torch.nn.Module):
                     group_size=group_size,
                     scale_precision=scale_precision,
                     zero_point_precision=zero_point_precision,
+                    weight_original_precision=child.weight.dtype,
                     device=child.weight.device,
                 )
                 setattr(module, name, quantized_embedding)
@@ -336,6 +337,7 @@ def __init__(
         group_size: int = 32,
         scale_precision: torch.dtype = torch.float32,
         zero_point_precision: torch.dtype = torch.int32,
+        weight_original_precision: torch.dtype = torch.float32,
         device: torch.device = None,
     ):
         super().__init__()
@@ -354,6 +356,7 @@ def __init__(
         self.group_size = group_size
         self.scale_precision = scale_precision
         self.zero_point_precision = zero_point_precision
+        self.weight_original_precision = weight_original_precision
 
         # currently storing unpacked int8 weights
         self.register_buffer(
@@ -393,7 +396,7 @@ def forward(self, x):
             qmax,
             torch.int8,
             self.group_size,
-            x.dtype,
+            self.weight_original_precision,
         )
         return F.embedding(
             x,