Fix NVFP4 QAT convert path (#3450)

**Summary:** The previous convert path through `QATConfig` did
not swap `NVFP4FakeQuantizedLinear` back to `torch.nn.Linear`.
The numerics tests still passed because this fake quantized
linear happen to match the PTQ numerics exactly.

**Test Plan:**
```
python test/quantization/test_qat.py -k test_qat_nvfp4
```

Author: andrewor14

test/quantization/test_qat.py

@@ -2106,23 +2106,28 @@ def test_qat_nvfp4(self, use_per_tensor_scale: bool):
         Test QAT with `NVFP4FakeQuantizeConfig`.
         """
         from torchao.prototype.mx_formats import NVFP4DynamicActivationNVFP4WeightConfig
-        from torchao.prototype.qat import NVFP4FakeQuantizeConfig
+        from torchao.prototype.mx_formats.nvfp4_tensor import NVFP4Tensor
+        from torchao.prototype.qat import (
+            NVFP4FakeQuantizeConfig,
+            NVFP4FakeQuantizedLinear,
+        )
 
         torch.manual_seed(self.SEED)
         m = M().cuda()
         baseline_model = copy.deepcopy(m)
-        quantize_(
-            baseline_model,
-            NVFP4DynamicActivationNVFP4WeightConfig(
-                use_dynamic_per_tensor_scale=use_per_tensor_scale
-            ),
+        base_config = NVFP4DynamicActivationNVFP4WeightConfig(
+            use_dynamic_per_tensor_scale=use_per_tensor_scale
         )
+        quantize_(baseline_model, base_config)
         qat_config = QATConfig(
             activation_config=NVFP4FakeQuantizeConfig(use_per_tensor_scale),
             weight_config=NVFP4FakeQuantizeConfig(use_per_tensor_scale),
             step="prepare",
         )
         quantize_(m, qat_config)
+        self.assertEqual(type(m.linear1), NVFP4FakeQuantizedLinear)
+        self.assertEqual(type(m.linear2), NVFP4FakeQuantizedLinear)
+        self.assertEqual(type(m.sub.linear), NVFP4FakeQuantizedLinear)
 
         # Compare prepared values
         torch.manual_seed(self.SEED)
@@ -2132,6 +2137,18 @@ def test_qat_nvfp4(self, use_per_tensor_scale: bool):
         sqnr = compute_error(out, baseline_out).item()
         self.assertGreaterEqual(sqnr, float("inf"))
 
+        # Compare converted values
+        quantize_(m, QATConfig(base_config, step="convert"))
+        self.assertEqual(type(m.linear1), torch.nn.Linear)
+        self.assertEqual(type(m.linear2), torch.nn.Linear)
+        self.assertEqual(type(m.sub.linear), torch.nn.Linear)
+        self.assertEqual(type(m.linear1.weight), NVFP4Tensor)
+        self.assertEqual(type(m.linear2.weight), NVFP4Tensor)
+        self.assertEqual(type(m.sub.linear.weight), NVFP4Tensor)
+        out = m(*x)
+        sqnr = compute_error(out, baseline_out).item()
+        self.assertGreaterEqual(sqnr, float("inf"))
+
     @unittest.skipIf(not _CUDA_IS_AVAILABLE, "skipping when cuda is not available")
     @unittest.skipIf(
         not _is_fbgemm_gpu_genai_available(), "Requires fbgemm-gpu-genai >= 1.2.0"

torchao/prototype/qat/nvfp4.py

@@ -162,6 +162,22 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
         else:
             return fq
 
+    def to_linear(self) -> torch.nn.Linear:
+        new_linear = torch.nn.Linear(
+            self.in_features,
+            self.out_features,
+            self.bias is not None,
+            device=self.weight.device,
+            dtype=self.weight.dtype,
+        )
+        # In distributed training, the model may be instantiated
+        # on the meta device, in which case there is no need to
+        # copy the weights, and doing so will result in an error
+        if self.weight.device != torch.device("meta"):
+            new_linear.weight = self.weight
+            new_linear.bias = self.bias
+        return new_linear
+
     @classmethod
     def from_linear(
         cls,

torchao/quantization/qat/api.py

@@ -198,6 +198,13 @@ def _qat_config_transform(
     modules to the corresponding built-in `torch.nn.Module`s, then apply the
     base config directly to quantize the module.
     """
+    # TODO: rewrite this using a registration API so
+    # specific quantization schemes do not leak here
+    from torchao.prototype.qat import (
+        NVFP4FakeQuantizeConfig,
+        NVFP4FakeQuantizedLinear,
+    )
+
     # Prepare step
     # Swap nn.Linear -> FakeQuantizedLinear
     # Swap nn.Embedding -> FakeQuantizedEmbedding
@@ -210,13 +217,6 @@ def _qat_config_transform(
             act_config = config.activation_config
             weight_config = config.weight_config
         if isinstance(module, torch.nn.Linear):
-            # TODO: rewrite this using a registration API so
-            # specific quantization schemes do not leak here
-            from torchao.prototype.qat import (
-                NVFP4FakeQuantizeConfig,
-                NVFP4FakeQuantizedLinear,
-            )
-
             if isinstance(weight_config, NVFP4FakeQuantizeConfig):
                 assert act_config is None or isinstance(
                     act_config, NVFP4FakeQuantizeConfig
@@ -245,17 +245,20 @@ def _qat_config_transform(
         assert config.weight_config is None, "unexpected `weight_config`"
 
         # Ignore unrelated modules
-        if not isinstance(module, (FakeQuantizedLinear, FakeQuantizedEmbedding)):
+        if not isinstance(
+            module,
+            (FakeQuantizedLinear, FakeQuantizedEmbedding, NVFP4FakeQuantizedLinear),
+        ):
             return module
 
         # Optionally pass custom scales and zero points to base config handler
         # This is only for range learning and only applies to weights
         kwargs = {}
         has_custom_scale_and_zero_point = False
-        weight_config = module.weight_fake_quantizer.config
         if (
-            isinstance(weight_config, IntxFakeQuantizeConfig)
-            and weight_config.range_learning
+            hasattr(module, "weight_fake_quantizer")
+            and isinstance(module.weight_fake_quantizer.config, IntxFakeQuantizeConfig)
+            and module.weight_fake_quantizer.config.range_learning
         ):
             kwargs["custom_scale"] = module.weight_fake_quantizer.scale
             kwargs["custom_zero_point"] = module.weight_fake_quantizer.zero_point
@@ -265,7 +268,7 @@ def _qat_config_transform(
         # Swap FakeQuantizedEmbedding -> nn.Embedding
         # Then apply the base config's transform function to quantize the model
         # If there is no base config, then simply perform the module swap
-        if isinstance(module, FakeQuantizedLinear):
+        if isinstance(module, (FakeQuantizedLinear, NVFP4FakeQuantizedLinear)):
             module = module.to_linear()
         elif isinstance(module, FakeQuantizedEmbedding):
             module = module.to_embedding()