Fix GPT-OSS swiglu_limit not passed in for MXFP4 #40197

Author: returnL

src/transformers/quantizers/quantizer_mxfp4.py

@@ -61,7 +61,16 @@ def validate_environment(self, *args, **kwargs):
             return
 
         if not torch.cuda.is_available():
-            if self.pre_quantized:
+            raise RuntimeError("Using MXFP4 quantized models requires a GPU")
+
+        if not is_accelerate_available():
+            raise ImportError("Using mxfp4 requires Accelerate: `pip install accelerate`")
+
+        compute_capability = torch.cuda.get_device_capability()
+        major, minor = compute_capability
+
+        if not is_triton_available("3.4.0") or not is_triton_kernels_availalble():
+            if self.pre_quantized and not self.quantization_config.dequantize:
                 logger.warning_once(
                     "Using MXFP4 quantized models requires a GPU, we will default to dequantizing the model to bf16"
                 )