Address review feedback: always emit qk_norm_epsilon, cache mxfp4 weight map, drop stray whitespace

- base.py: always add qk_norm_epsilon GQA attribute (matches convention of unconditionally-added attributes; value only used when q/k norm weights are present)
- base.py: remove whitespace-only line after CudaQuantizer import
- gptoss.py: add explicit _mxfp4_weight_map_loaded flag so the no-index case caches instead of re-globbing the snapshot dir on every tensor load

Author: tianleiwu

src/python/py/models/builders/base.py

@@ -33,7 +33,7 @@
 )
 
 from .cuda_quantizer import CudaQuantizer
-                                                                                                                                                                                                                        
+
 
 class Model:
     def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
@@ -2751,9 +2751,8 @@ def make_group_query_attention(self, name, **kwargs):
             "softcap": self.attention_attrs["softcap"],
             "do_rotary": self.attention_attrs["use_rope_in_attn"],
             "rotary_interleaved": self.rope_attrs["interleaved"],
+            "qk_norm_epsilon": kwargs.get("qk_norm_epsilon", self.attention_attrs["qk_norm_epsilon"]),
         }
-        if q_norm_weight:
-            attributes["qk_norm_epsilon"] = kwargs.get("qk_norm_epsilon", self.attention_attrs["qk_norm_epsilon"])
         self.make_node(
             "GroupQueryAttention",
             inputs=inputs,

src/python/py/models/builders/gptoss.py

@@ -22,6 +22,7 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
         self.moe_attrs["normalize_routing_weights"] = True
         self.moe_attrs["swiglu_fusion"] = 1
         self._mxfp4_weight_map = None
+        self._mxfp4_weight_map_loaded = False
         self._mxfp4_snapshot_dir = None
         self._mxfp4_tensor_cache = {}
 
@@ -50,7 +51,7 @@ def _get_mxfp4_snapshot_dir(self):
         return self._mxfp4_snapshot_dir
 
     def _get_mxfp4_weight_map(self):
-        if self._mxfp4_weight_map is not None:
+        if self._mxfp4_weight_map_loaded:
             return self._mxfp4_weight_map
 
         snapshot_dir = self._get_mxfp4_snapshot_dir()
@@ -67,6 +68,7 @@ def _get_mxfp4_weight_map(self):
                     f"Could not locate original GPT-OSS MXFP4 safetensors index in {snapshot_dir}."
                 )
             self._mxfp4_weight_map = None
+        self._mxfp4_weight_map_loaded = True
         return self._mxfp4_weight_map
 
     def _load_original_mxfp4_tensor(self, tensor_name):