[chore] harden checkpoint quantization metadata parsing (#36922)

test/registered/unit/model_loader/test_modelopt_loader.py

@@ -14,6 +14,7 @@ from unittest.mock import MagicMock, patch
 
 import torch
 import torch.nn as nn
+from transformers import PretrainedConfig
 
 from sglang.srt.configs.device_config import DeviceConfig
 from sglang.srt.configs.load_config import LoadConfig
@@ -657,7 +658,7 @@ class TestModelOptFp4LoaderSelection(CustomTestCase):
                     quantization="modelopt_fp4",
                     is_draft_model=True,
                     is_draft_quantization_explicit=is_explicit,
-                    hf_config=SimpleNamespace(
+                    hf_config=PretrainedConfig(
                         quantization_config={
                             "quant_algo": "NVFP4",
                             "group_size": 16,
@@ -757,7 +758,7 @@ class TestModelOptMixedPrecisionConfig(CustomTestCase):
                 with self.subTest(inline_config=inline_config):
                     model_config = SimpleNamespace(
                         quantization="modelopt_mixed",
-                        hf_config=SimpleNamespace(
+                        hf_config=PretrainedConfig(
                             quantization_config=inline_config,
                         ),
                         model_path=model_path,
@@ -787,7 +788,7 @@ class TestModelOptMixedPrecisionConfig(CustomTestCase):
         }
         model_config = SimpleNamespace(
             quantization="modelopt_mixed",
-            hf_config=SimpleNamespace(
+            hf_config=PretrainedConfig(
                 quantization_config={
                     "quant_method": "modelopt_mixed",
                     "quant_algo": "MIXED_PRECISION",

test/registered/unit/test_checkpoint_quantization.py

@@ -2,6 +2,8 @@
 
 import unittest
 
+from transformers import PretrainedConfig
+
 from sglang.srt.layers.modelopt_utils import canonicalize_modelopt_quant_algo
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.model_loader.checkpoint_quantization import (
@@ -14,19 +16,6 @@ from sglang.test.test_utils import CustomTestCase
 register_cpu_ci(est_time=5, suite="base-a-test-cpu")
 
 
-class _ConfigObject:
-    def __init__(self, **values):
-        self.__dict__.update(values)
-
-
-class _QuantConfigObject:
-    def __init__(self, values):
-        self._values = values
-
-    def to_dict(self):
-        return self._values
-
-
 class TestResolveCheckpointQuantSpec(CustomTestCase):
     def test_modelopt_quant_algo_canonicalization(self):
         cases = {
@@ -75,9 +64,9 @@ class TestResolveCheckpointQuantSpec(CustomTestCase):
             ),
         )
 
-    def test_text_config_fallback_supports_config_objects(self):
-        config = _ConfigObject(
-            text_config=_ConfigObject(
+    def test_text_config_fallback_supports_pretrained_configs(self):
+        config = PretrainedConfig(
+            text_config=PretrainedConfig(
                 quantization_config={"quant_method": "gptq", "bits": 4}
             ),
             compression_config={"quant_method": "compressed-tensors"},
@@ -90,7 +79,7 @@ class TestResolveCheckpointQuantSpec(CustomTestCase):
         self.assertEqual(spec.source, "text_config.quantization_config")
 
     def test_compression_config_fallback(self):
-        config = _ConfigObject(
+        config = PretrainedConfig(
             compression_config={"quant_method": "compressed-tensors"}
         )
 
@@ -114,18 +103,6 @@ class TestResolveCheckpointQuantSpec(CustomTestCase):
         self.assertIsNone(spec.declared_method)
         self.assertEqual(spec.config["quant_algo"], "FP8")
 
-    def test_quant_config_object_is_converted(self):
-        config = _ConfigObject(
-            quantization_config=_QuantConfigObject(
-                {"quant_method": "bitsandbytes", "load_in_4bit": True}
-            )
-        )
-
-        spec = resolve_checkpoint_quant_spec(config)
-
-        self.assertIsNotNone(spec)
-        self.assertEqual(spec.config["load_in_4bit"], True)
-
     def test_lookup_priority_matches_srt_loader(self):
         config = {
             "quantization_config": {},