Revert "Move Qwen3VL imports back to module top level"

This reverts commit 63a229a.

Author: hychiang-git

tests/_test_utils/torch/transformers_models.py

@@ -32,11 +32,9 @@
     PreTrainedModel,
     Qwen3Config,
     Qwen3MoeConfig,
-    Qwen3VLConfig,
     T5Config,
     T5ForConditionalGeneration,
 )
-from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLForConditionalGeneration
 
 import modelopt.torch.opt as mto
 
@@ -125,6 +123,12 @@ def create_tiny_qwen3_moe_dir(
 
 ##### Qwen3-VL #####
 def get_tiny_qwen3vl(**config_kwargs) -> PreTrainedModel:
+    # Lazy imports — Qwen3VL classes live under transformers.models.qwen3_vl which
+    # may not exist in older transformers builds, and this module is imported by
+    # every test that uses transformers_models.py.
+    from transformers import Qwen3VLConfig
+    from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLForConditionalGeneration
+
     set_seed(SEED)
 
     # Defaults: hidden_size=num_attention_heads*head_dim (e.g. 4*8=32).

tests/gpu_megatron/torch/export/test_unified_export_megatron.py

@@ -30,7 +30,6 @@
 )
 from safetensors import safe_open
 from safetensors.torch import save_file
-from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLForConditionalGeneration
 
 import modelopt.torch.quantization as mtq
 import modelopt.torch.speculative as mtsp
@@ -198,6 +197,10 @@ def _test_unified_export_megatron(
                 "vision encoder keys missing from combined export"
             )
             # try to load the model and run a forward pass
+            from transformers.models.qwen3_vl.modeling_qwen3_vl import (
+                Qwen3VLForConditionalGeneration,
+            )
+
             vl_model = Qwen3VLForConditionalGeneration.from_pretrained(
                 tmp_export_dir, torch_dtype=torch.bfloat16
             ).cuda()