[CPU] Fix torch version in x86 CPU backend (#19258)

Signed-off-by: jiang1.li <jiang1.li@intel.com>

tests/models/multimodal/generation/test_common.py

@@ -107,6 +107,8 @@ VLM_TEST_SETTINGS = {
             ),
             limit_mm_per_prompt={"image": 4},
         )],
+        # TODO: Revert to "auto" when CPU backend can use torch > 2.6
+        dtype="bfloat16" if current_platform.is_cpu() else "auto",
         marks=[pytest.mark.core_model, pytest.mark.cpu_model],
     ),
     "paligemma": VLMTestInfo(

tests/models/multimodal/generation/vlm_utils/builders.py

@@ -203,6 +203,9 @@ def build_embedding_inputs_from_test_info(
 
     images = [asset.pil_image for asset in image_assets]
     embeds = test_info.convert_assets_to_embeddings(image_assets)
+    if test_info.dtype != "auto":
+        dtype = getattr(torch, test_info.dtype)  # type: ignore
+        embeds = [e.to(dtype=dtype) for e in embeds]
     assert len(images) == len(model_prompts)
 
     inputs = build_single_image_inputs(images, model_prompts, size_wrapper)