[testing] test num_hidden_layers being small in model tester (#40992)

ydshieh · web-flow · commit 46ea7e613d67 · 2025-09-19T11:45:07.000+02:00
fix

Co-authored-by: ydshieh &lt;ydshieh@users.noreply.github.com&gt;
diff --git a/tests/models/aya_vision/test_modeling_aya_vision.py b/tests/models/aya_vision/test_modeling_aya_vision.py
@@ -71,7 +71,7 @@ def __init__(
             "vocab_size": 99,
             "hidden_size": 128,
             "intermediate_size": 37,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "num_attention_heads": 4,
             "output_channels": 64,
             "hidden_act": "silu",
diff --git a/tests/models/bamba/test_modeling_bamba.py b/tests/models/bamba/test_modeling_bamba.py
@@ -73,7 +73,7 @@ def __init__(
         use_labels=True,
         vocab_size=99,
         hidden_size=32,
-        num_hidden_layers=4,
+        num_hidden_layers=2,
         num_attention_heads=4,
         num_key_value_heads=2,
         intermediate_size=64,
diff --git a/tests/models/bitnet/test_modeling_bitnet.py b/tests/models/bitnet/test_modeling_bitnet.py
@@ -49,7 +49,7 @@ def __init__(
         use_input_mask=True,
         vocab_size=99,
         hidden_size=64,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         num_key_value_heads=2,
         intermediate_size=37,
diff --git a/tests/models/bros/test_modeling_bros.py b/tests/models/bros/test_modeling_bros.py
@@ -49,7 +49,7 @@ def __init__(
         use_labels=True,
         vocab_size=99,
         hidden_size=64,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         intermediate_size=37,
         hidden_act="gelu",
diff --git a/tests/models/cohere/test_modeling_cohere.py b/tests/models/cohere/test_modeling_cohere.py
@@ -54,7 +54,7 @@ def __init__(
         use_labels=True,
         vocab_size=99,
         hidden_size=32,
-        num_hidden_layers=4,
+        num_hidden_layers=2,
         num_attention_heads=4,
         intermediate_size=37,
         hidden_act="gelu",
diff --git a/tests/models/cohere2_vision/test_modeling_cohere2_vision.py b/tests/models/cohere2_vision/test_modeling_cohere2_vision.py
@@ -65,7 +65,7 @@ def __init__(
             "vocab_size": 99,
             "hidden_size": 128,
             "intermediate_size": 37,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "num_attention_heads": 4,
             "output_channels": 64,
             "hidden_act": "silu",
diff --git a/tests/models/deepseek_v3/test_modeling_deepseek_v3.py b/tests/models/deepseek_v3/test_modeling_deepseek_v3.py
@@ -65,7 +65,7 @@ def __init__(
         hidden_size=32,
         intermediate_size=37,
         moe_intermediate_size=12,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         num_key_value_heads=4,
         n_shared_experts=1,
diff --git a/tests/models/eomt/test_modeling_eomt.py b/tests/models/eomt/test_modeling_eomt.py
@@ -47,7 +47,7 @@ def __init__(
         num_labels=4,
         hidden_size=8,
         num_attention_heads=2,
-        num_hidden_layers=4,
+        num_hidden_layers=2,
     ):
         self.parent = parent
         self.batch_size = batch_size
diff --git a/tests/models/falcon/test_modeling_falcon.py b/tests/models/falcon/test_modeling_falcon.py
@@ -208,7 +208,7 @@ def test_falcon_alibi_sdpa_matches_eager(self):
         config = FalconConfig(
             vocab_size=1000,
             hidden_size=64,
-            num_hidden_layers=3,
+            num_hidden_layers=2,
             num_attention_heads=4,
             new_decoder_architecture=True,
             alibi=True,
diff --git a/tests/models/falcon_h1/test_modeling_falcon_h1.py b/tests/models/falcon_h1/test_modeling_falcon_h1.py
@@ -55,7 +55,7 @@ def __init__(
         use_labels=True,
         vocab_size=99,
         hidden_size=32,
-        num_hidden_layers=4,
+        num_hidden_layers=2,
         num_attention_heads=4,
         num_key_value_heads=2,
         intermediate_size=64,
diff --git a/tests/models/got_ocr2/test_modeling_got_ocr2.py b/tests/models/got_ocr2/test_modeling_got_ocr2.py
@@ -59,7 +59,7 @@ def __init__(
             "vocab_size": 99,
             "hidden_size": 128,
             "intermediate_size": 37,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "num_attention_heads": 4,
             "num_key_value_heads": 2,
             "output_channels": 64,
diff --git a/tests/models/idefics/test_modeling_idefics.py b/tests/models/idefics/test_modeling_idefics.py
@@ -67,7 +67,7 @@ def __init__(
         use_labels=True,
         vocab_size=99,
         hidden_size=32,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         intermediate_size=37,
         hidden_act="gelu",
@@ -85,7 +85,7 @@ def __init__(
         vision_patch_size=2,
         vision_image_size=30,
         vision_num_attention_heads=4,
-        vision_num_hidden_layers=5,
+        vision_num_hidden_layers=2,
         vision_intermediate_size=37,
         perceiver_qk_layer_norms_perceiver=False,
         perceiver_resampler_depth=2,
diff --git a/tests/models/idefics2/test_modeling_idefics2.py b/tests/models/idefics2/test_modeling_idefics2.py
@@ -86,7 +86,7 @@ def __init__(
             "vocab_size": 100,
             "hidden_size": 64,
             "intermediate_size": 56,
-            "num_hidden_layers": 3,
+            "num_hidden_layers": 2,
             "num_attention_heads": 2,
             "num_key_value_heads": 2,
             "hidden_act": "silu",
diff --git a/tests/models/idefics3/test_modeling_idefics3.py b/tests/models/idefics3/test_modeling_idefics3.py
@@ -74,7 +74,7 @@ def __init__(
             "vocab_size": 100,
             "hidden_size": 64,
             "intermediate_size": 56,
-            "num_hidden_layers": 3,
+            "num_hidden_layers": 2,
             "num_attention_heads": 2,
             "num_key_value_heads": 2,
             "hidden_act": "silu",
diff --git a/tests/models/internvl/test_modeling_internvl.py b/tests/models/internvl/test_modeling_internvl.py
@@ -74,7 +74,7 @@ def __init__(
             "vocab_size": 99,
             "hidden_size": 128,
             "intermediate_size": 37,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "num_attention_heads": 4,
             "num_key_value_heads": 2,
             "output_channels": 64,
diff --git a/tests/models/longcat_flash/test_modeling_longcat_flash.py b/tests/models/longcat_flash/test_modeling_longcat_flash.py
@@ -60,7 +60,7 @@ def __init__(
         hidden_size=144,
         ffn_hidden_size=288,
         expert_ffn_hidden_size=48,
-        num_layers=2,
+        num_layers=1,  # We have `self.num_hidden_layers = 2 * num_layers` in the body. See `LongcatFlashConfig`.
         num_attention_heads=8,
         num_key_value_heads=8,
         kv_lora_rank=16,
@@ -96,7 +96,7 @@ def __init__(
         self.expert_ffn_hidden_size = expert_ffn_hidden_size
         self.num_layers = num_layers
         self.num_hidden_layers = 2 * num_layers  # for compatibility
-        self.expected_num_hidden_layers = 3  # embedding + 2 layers
+        self.expected_num_hidden_layers = 2  # embedding + 2 layers
         self.num_attention_heads = num_attention_heads
         self.num_key_value_heads = num_key_value_heads
         self.kv_lora_rank = kv_lora_rank
diff --git a/tests/models/lxmert/test_modeling_lxmert.py b/tests/models/lxmert/test_modeling_lxmert.py
@@ -59,7 +59,7 @@ def __init__(
         num_object_labels=16,
         num_attr_labels=4,
         num_visual_features=10,
-        l_layers=2,
+        l_layers=1,
         x_layers=1,
         r_layers=1,
         visual_feat_dim=128,
diff --git a/tests/models/mllama/test_modeling_mllama.py b/tests/models/mllama/test_modeling_mllama.py
@@ -145,7 +145,7 @@ def __init__(
             "model_type": "mllama",
             "vocab_size": 99,
             "hidden_size": 32,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "num_attention_heads": 4,
             "num_key_value_heads": 4,
             "intermediate_size": 37,
@@ -166,7 +166,7 @@ def __init__(
             "intermediate_layers_indices": [0],
             "vision_output_dim": 32,
             "projection_dim": 32,
-            "num_hidden_layers": 6,
+            "num_hidden_layers": 2,
             "num_global_layers": 2,
             "num_attention_heads": 4,
             "intermediate_size": 37,
diff --git a/tests/models/pop2piano/test_modeling_pop2piano.py b/tests/models/pop2piano/test_modeling_pop2piano.py
@@ -57,7 +57,7 @@ def __init__(
         use_attention_mask=True,
         use_labels=True,
         hidden_size=64,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         d_ff=37,
         relative_attention_num_buckets=8,
diff --git a/tests/models/qwen2_5_omni/test_modeling_qwen2_5_omni.py b/tests/models/qwen2_5_omni/test_modeling_qwen2_5_omni.py
@@ -99,7 +99,7 @@ def __init__(
             "vocab_size": 99,
             "hidden_size": 32,
             "intermediate_size": 37,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "num_attention_heads": 4,
             "num_key_value_heads": 2,
             "hidden_act": "silu",
diff --git a/tests/models/qwen2_5_vl/test_modeling_qwen2_5_vl.py b/tests/models/qwen2_5_vl/test_modeling_qwen2_5_vl.py
@@ -85,7 +85,7 @@ def __init__(
         max_window_layers=3,
         model_type="qwen2_5_vl",
         num_attention_heads=4,
-        num_hidden_layers=4,
+        num_hidden_layers=2,
         num_key_value_heads=2,
         rope_theta=10000,
         tie_word_embeddings=True,
diff --git a/tests/models/qwen2_vl/test_modeling_qwen2_vl.py b/tests/models/qwen2_vl/test_modeling_qwen2_vl.py
@@ -79,7 +79,7 @@ def __init__(
         max_window_layers=3,
         model_type="qwen2_vl",
         num_attention_heads=4,
-        num_hidden_layers=4,
+        num_hidden_layers=2,
         num_key_value_heads=2,
         rope_theta=10000,
         tie_word_embeddings=True,
diff --git a/tests/models/qwen3_vl/test_modeling_qwen3_vl.py b/tests/models/qwen3_vl/test_modeling_qwen3_vl.py
@@ -61,7 +61,7 @@ def __init__(
             "max_position_embeddings": 512,
             "model_type": "qwen3_vl",
             "num_attention_heads": 4,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "num_key_value_heads": 2,
             "rope_theta": 10000,
             "tie_word_embeddings": True,
diff --git a/tests/models/qwen3_vl_moe/test_modeling_qwen3_vl_moe.py b/tests/models/qwen3_vl_moe/test_modeling_qwen3_vl_moe.py
@@ -61,7 +61,7 @@ def __init__(
             "model_type": "qwen3_vl_moe",
             "num_attention_heads": 4,
             "num_key_value_heads": 2,
-            "num_hidden_layers": 4,
+            "num_hidden_layers": 2,
             "moe_intermediate_size": 16,
             "num_experts_per_tok": 4,
             "num_experts": 8,
diff --git a/tests/models/reformer/test_modeling_reformer.py b/tests/models/reformer/test_modeling_reformer.py
@@ -83,7 +83,7 @@ def __init__(
         axial_pos_embds=True,
         axial_pos_shape=[4, 8],
         axial_pos_embds_dim=[16, 16],
-        attn_layers=["local", "local", "local", "local"],
+        attn_layers=["local", "local"],
         pad_token_id=0,
         eos_token_id=2,
         scope=None,
diff --git a/tests/models/smolvlm/test_modeling_smolvlm.py b/tests/models/smolvlm/test_modeling_smolvlm.py
@@ -77,7 +77,7 @@ def __init__(
             "vocab_size": 100,
             "hidden_size": 64,
             "intermediate_size": 56,
-            "num_hidden_layers": 3,
+            "num_hidden_layers": 2,
             "num_attention_heads": 2,
             "num_key_value_heads": 2,
             "hidden_act": "silu",
diff --git a/tests/models/udop/test_modeling_udop.py b/tests/models/udop/test_modeling_udop.py
@@ -55,7 +55,7 @@ def __init__(
         use_attention_mask=True,
         use_labels=True,
         hidden_size=32,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         d_ff=37,
         relative_attention_num_buckets=32,
@@ -425,7 +425,7 @@ def __init__(
         is_training=False,
         use_attention_mask=True,
         hidden_size=32,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         decoder_layers=2,
         num_attention_heads=4,
         d_ff=37,
diff --git a/tests/models/vitpose/test_modeling_vitpose.py b/tests/models/vitpose/test_modeling_vitpose.py
@@ -51,7 +51,7 @@ def __init__(
         is_training=True,
         use_labels=True,
         hidden_size=32,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         intermediate_size=37,
         hidden_act="gelu",
diff --git a/tests/models/vitpose_backbone/test_modeling_vitpose_backbone.py b/tests/models/vitpose_backbone/test_modeling_vitpose_backbone.py
@@ -44,7 +44,7 @@ def __init__(
         is_training=True,
         use_labels=True,
         hidden_size=32,
-        num_hidden_layers=5,
+        num_hidden_layers=2,
         num_attention_heads=4,
         intermediate_size=37,
         hidden_act="gelu",
diff --git a/tests/models/vjepa2/test_modeling_vjepa2.py b/tests/models/vjepa2/test_modeling_vjepa2.py
@@ -61,7 +61,7 @@ def __init__(
         patch_size=16,
         num_channels=3,
         hidden_size=32,
-        num_hidden_layers=4,
+        num_hidden_layers=2,
         num_attention_heads=2,
         num_frames=2,
         mlp_ratio=1,
diff --git a/tests/models/xlnet/test_modeling_xlnet.py b/tests/models/xlnet/test_modeling_xlnet.py
@@ -80,7 +80,7 @@ def __init__(
         self.hidden_size = 32
         self.num_attention_heads = 4
         self.d_inner = 128
-        self.num_hidden_layers = 5
+        self.num_hidden_layers = 3
         self.type_sequence_label_size = 2
         self.bi_data = False
         self.same_length = False
diff --git a/tests/test_modeling_common.py b/tests/test_modeling_common.py
@@ -674,6 +674,45 @@ def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
 
         return inputs_dict
 
+    def test_num_layers_is_small(self):
+        # TODO (if possible): Avoid exceptional cases, especially for `OwlViT`.
+        # ⛔ DO NOT edit this list (unless there is really nothing to tweak in the model tester class and approved by the reviewer) ⛔!
+        exceptional_num_hidden_layers = {
+            # TODO: There might be some way to fix
+            "FunnelModelTest": 5,
+            "FunnelBaseModelTest": 4,
+            "GroupViTVisionModelTest": 12,
+            "OwlViTModelTest": 12,
+            "OwlViTTextModelTest": 12,
+            "OwlViTForObjectDetectionTest": 12,
+            "Owlv2ModelTest": 12,
+            "Owlv2TextModelTest": 12,
+            "Owlv2ForObjectDetectionTest": 12,
+            "SamHQModelTest": 12,
+            "Swin2SRModelTest": 3,
+            "XLNetModelTest": 3,
+            "DPTModelTest": 4,  # `test_modeling_dpt_hybrid.py`: not able to get it work after change `num_hidden_layers` and `neck_hidden_sizes`
+            # Nothing we can't do
+            "Gemma3nTextModelTest": 4,  # need to test KV shared layer for both types: `full_attention` and `sliding_attention`
+            "BeitModelTest": 4,  # BeitForSemanticSegmentation requires config.out_indices to be a list of 4 integers
+            "ZambaModelTest": 5,  # The minimum number to test beyond the initial ["mamba", "mamba", "hybrid"] in `ZambaConfig._layers_block_type`
+        }
+        target_num_hidden_layers = exceptional_num_hidden_layers.get(type(self).__name__, 2)
+
+        if hasattr(self.model_tester, "num_hidden_layers") and isinstance(self.model_tester.num_hidden_layers, int):
+            assert self.model_tester.num_hidden_layers <= target_num_hidden_layers
+
+        if hasattr(self.model_tester, "vision_config") and "num_hidden_layers" in self.model_tester.vision_config:
+            if isinstance(self.model_tester.vision_config, dict):
+                assert self.model_tester.vision_config["num_hidden_layers"] <= target_num_hidden_layers
+            else:
+                assert self.model_tester.vision_config.num_hidden_layers <= target_num_hidden_layers
+        if hasattr(self.model_tester, "text_config") and "num_hidden_layers" in self.model_tester.text_config:
+            if isinstance(self.model_tester.text_config, dict):
+                assert self.model_tester.text_config["num_hidden_layers"] <= target_num_hidden_layers
+            else:
+                assert self.model_tester.text_config.num_hidden_layers <= target_num_hidden_layers
+
     def test_save_load(self):
         def check_save_load(out1, out2):
             # make sure we don't have nans
