Enable gpt-oss mxfp4 on older hardware (sm75+) (#39940)

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>

tests/quantization/mxfp4/test_mxfp4.py

@@ -107,18 +107,31 @@ class Mxfp4QuantizerTest(unittest.TestCase):
 
     def test_quantizer_validation_low_compute_capability(self):
         """Test quantizer validation with low compute capability"""
-        with patch("torch.cuda.get_device_capability", return_value=(8, 0)):
+        with patch("torch.cuda.get_device_capability", return_value=(7, 0)):
+            from transformers.quantizers.quantizer_mxfp4 import Mxfp4HfQuantizer
+
+            config = Mxfp4Config()
+            quantizer = Mxfp4HfQuantizer(config)
+            quantizer.pre_quantized = False
+
+            with self.assertRaises(ValueError):
+                quantizer.validate_environment()
+
+    def test_quantizer_validation_low_compute_capability_with_prequantized(self):
+        """Test quantizer validation with low compute capability"""
+        with patch("torch.cuda.get_device_capability", return_value=(7, 0)):
             from transformers.quantizers.quantizer_mxfp4 import Mxfp4HfQuantizer
 
             config = Mxfp4Config()
             quantizer = Mxfp4HfQuantizer(config)
 
-            with self.assertRaises(ValueError):
-                quantizer.validate_environment()
+            # Should automatically set dequantize=True and warn
+            quantizer.validate_environment()
+            self.assertTrue(quantizer.quantization_config.dequantize)
 
     def test_quantizer_validation_low_compute_capability_with_dequantize(self):
         """Test quantizer validation with low compute capability but dequantize enabled"""
-        with patch("torch.cuda.get_device_capability", return_value=(8, 0)):
+        with patch("torch.cuda.get_device_capability", return_value=(7, 0)):
             from transformers.quantizers.quantizer_mxfp4 import Mxfp4HfQuantizer
 
             config = Mxfp4Config(dequantize=True)