Fix Nano Nemotron VL regressions (#38655)

Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com>

tests/models/registry.py

@@ -7,6 +7,7 @@ from typing import Any, Literal
 
 import pytest
 from packaging.version import Version
+from transformers import PretrainedConfig
 from transformers import __version__ as TRANSFORMERS_VERSION
 
 from vllm.config.model import ModelDType, TokenizerMode
@@ -1004,7 +1005,26 @@ _MULTIMODAL_EXAMPLE_MODELS = {
         trust_remote_code=True,
     ),
     "NemotronH_Nano_VL_V2": _HfExamplesInfo(
-        "nano_vl_dummy", is_available_online=False, trust_remote_code=True
+        "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16",
+        max_model_len=4096,
+        # NemotronH layers are constructed via `hybrid_override_pattern`:
+        use_original_num_layers=True,
+        hf_overrides={
+            "vision_config": PretrainedConfig(
+                args={
+                    "min_num_patches": 1,  # Trigger image dynamic res
+                    "max_num_patches": 12,
+                    "model": "vit_huge_patch16_224",
+                },
+                # Trigger conv3d:
+                video_temporal_patch_size=2,
+            ),
+            "text_config": {
+                "num_hidden_layers": 2,
+                "hybrid_override_pattern": "M*",
+            },
+        },
+        trust_remote_code=True,
     ),
     "OpenCUAForConditionalGeneration": _HfExamplesInfo(
         "xlangai/OpenCUA-7B", trust_remote_code=True

tests/models/utils.py

@@ -447,9 +447,16 @@ def dummy_hf_overrides(
     Dummy HF overrides function used to create dummy model
     with only minimum nums of layer.
     """
-    hf_config.update(exist_overrides or {})
+    # Copy because this helper is called more than once
+    # while loading config, and we `.pop()`
+    exist_overrides = (exist_overrides or {}).copy()
+    text_config_override = exist_overrides.pop("text_config", None)
+    hf_config.update(exist_overrides)
 
     text_config = hf_config.get_text_config()
+    if text_config_override is not None:
+        # multimodal test models may override *some* text-model fields
+        text_config.update(text_config_override)
 
     # Ensure at least 2 expert per group
     # Since `grouped_topk` assumes top-2