Add parameter_name support to _nvfp4_inference_linear_transform (#3976)

* Add parameter_name support to _mx_inference_linear_transform

Add `parameter_name` keyword-only argument to support FqnToConfig
per-parameter quantization. Replace module.weight reads/writes with
getattr/setattr, and use _module_extra_repr partial pattern.

[ghstack-poisoned]

* Add parameter_name support to _nvfp4_inference_linear_transform

Add `parameter_name` keyword-only argument to the dynamic quantization
branch (step=None) to support FqnToConfig per-parameter quantization.
PREPARE and CONVERT branches are left unchanged as they are
Linear-specific observer flow.

[ghstack-poisoned]

* Update base for Update on "Add parameter_name support to _nvfp4_inference_linear_transform"


Add `parameter_name` keyword-only argument to the dynamic quantization
branch (step=None) to support FqnToConfig per-parameter quantization.
PREPARE and CONVERT branches are left unchanged as they are
Linear-specific observer flow.

[ghstack-poisoned]

Author: jcaip

test/quantization/test_quant_api.py

@@ -63,6 +63,7 @@
     is_sm_at_least_89,
     is_sm_at_least_90,
     is_sm_at_least_100,
+    torch_version_at_least,
     unwrap_tensor_subclass,
 )
 
@@ -75,6 +76,7 @@
 
 from torchao.prototype.mx_formats.inference_workflow import (
     MXDynamicActivationMXWeightConfig,
+    NVFP4DynamicActivationNVFP4WeightConfig,
 )
 
 
@@ -1057,6 +1059,8 @@ def test_fqn_to_config_non_weight_param(self):
         ]
         if is_sm_at_least_100():
             configs.append(MXDynamicActivationMXWeightConfig())
+        if is_sm_at_least_100() and torch_version_at_least("2.8.0"):
+            configs.append(NVFP4DynamicActivationNVFP4WeightConfig())
         for config in configs:
             with self.subTest(config=type(config).__name__):
                 model = torch.nn.Sequential(

torchao/prototype/mx_formats/inference_workflow.py

@@ -240,7 +240,10 @@ def __post_init__(self):
 
 @register_quantize_module_handler(NVFP4DynamicActivationNVFP4WeightConfig)
 def _nvfp4_inference_linear_transform(
-    module: torch.nn.Linear, config: NVFP4DynamicActivationNVFP4WeightConfig
+    module: torch.nn.Linear,
+    config: NVFP4DynamicActivationNVFP4WeightConfig,
+    *,
+    parameter_name: str = "weight",
 ):
     """Quantization handler for NVFP4DynamicActivationNVFP4WeightConfig
 
@@ -249,7 +252,7 @@ def _nvfp4_inference_linear_transform(
     - CONVERT: Extract amax from observer, compute static per_tensor_scale, quantize
     - None (default): Original dynamic quantization behavior
     """
-    weight = module.weight
+    weight = getattr(module, parameter_name)
     if weight.shape[-2] % 16 != 0 or weight.shape[-1] % 16 != 0:
         raise RuntimeError(
             f"NVFP4 only supports weight shape with last 2 dims divisible by 16, got {weight.shape}"
@@ -306,6 +309,8 @@ def _nvfp4_inference_linear_transform(
             "NVFP4 DYNAMIC mode is only supported on sm100+ machines"
         )
 
+        weight = getattr(module, parameter_name)
+
         per_tensor_scale = None
         if config.use_dynamic_per_tensor_scale:
             tensor_amax = torch.max(torch.abs(weight))
@@ -325,8 +330,19 @@ def _nvfp4_inference_linear_transform(
             act_quant_kwargs=act_quant_kwargs,
         )
         quantized_weight.use_triton_kernel = config.use_triton_kernel
-        module.weight = torch.nn.Parameter(quantized_weight, requires_grad=False)
-        module.extra_repr = types.MethodType(_linear_extra_repr, module)
+        setattr(
+            module,
+            parameter_name,
+            torch.nn.Parameter(quantized_weight, requires_grad=False),
+        )
+        module.extra_repr = types.MethodType(
+            partial(
+                _module_extra_repr,
+                original_extra_repr=module.extra_repr,
+                parameter_name=parameter_name,
+            ),
+            module,
+        )
         return module
 
     else: