[Bugfix] Fix broadcasting logic for multi_modal_kwargs (#6836)

tests/distributed/test_multimodal_broadcast.py

@@ -19,10 +19,10 @@ from vllm.utils import cuda_device_count_stateless
 
 model = os.environ["TEST_DIST_MODEL"]
 
-if model.startswith("llava-hf/llava"):
+if model.startswith("llava-hf/llava-1.5"):
     from ..models.test_llava import models, run_test
-elif model.startswith("microsoft/Phi-3-vision"):
-    from ..models.test_phi3v import models, run_test
+elif model.startswith("llava-hf/llava-v1.6"):
+    from ..models.test_llava_next import models, run_test
 else:
     raise NotImplementedError(f"Unsupported model: {model}")
 
@@ -45,7 +45,8 @@ def test_models(hf_runner, vllm_runner, image_assets,
         vllm_runner,
         image_assets,
         model=models[0],
-        size_factors=[1.0],
+        # So that LLaVA-NeXT processor may return nested list
+        size_factors=[0.25, 0.5, 1.0],
         dtype=dtype,
         max_tokens=max_tokens,
         num_logprobs=num_logprobs,

tests/distributed/test_parallel_state.py

@@ -1,57 +0,0 @@
-from typing import Any, Dict
-
-import pytest
-import torch
-
-from vllm.distributed.parallel_state import (_split_tensor_dict,
-                                             _update_nested_dict)
-
-
-def test_split_tensor_dict():
-    test_dict = {
-        "key_a": "a",
-        "key_b": torch.arange(8, dtype=torch.float32),
-        "key_c": {
-            "key_1": torch.arange(5, dtype=torch.float32),
-            "key_2": torch.tensor([], dtype=torch.float32),
-            "key_3": 123,
-        },
-        "key_d": {},
-    }
-    metadata_list, tensor_list = _split_tensor_dict(test_dict)
-    assert len(metadata_list) == 6
-    assert torch.allclose(tensor_list[0], test_dict["key_b"])
-    assert torch.allclose(tensor_list[1], test_dict["key_c"]["key_1"])
-    assert torch.allclose(tensor_list[2], test_dict["key_c"]["key_2"])
-
-
-def test_split_tensor_dict_invalid_key():
-    test_dict = {
-        "a%b": "a",
-    }
-    with pytest.raises(AssertionError):
-        _split_tensor_dict(test_dict)
-
-
-def test_update_nested_dict():
-    flattened_keys_values = [("key1%key2%key3", "value1"),
-                             ("key1%key2%key4", "value2"),
-                             ("key1%key5", "value3"), ("key6%key7", "value4"),
-                             ("key8", "value5")]
-    res: Dict[str, Any] = {}
-
-    for flat_key, value in flattened_keys_values:
-        _update_nested_dict(res, flat_key, value)
-    assert res == {
-        "key1": {
-            "key2": {
-                "key3": "value1",
-                "key4": "value2"
-            },
-            "key5": "value3"
-        },
-        "key6": {
-            "key7": "value4"
-        },
-        "key8": "value5"
-    }