[Model] VLM2Vec, the first multimodal embedding model in vLLM (#9303)

tests/models/embedding/utils.py

@@ -0,0 +1,29 @@
+from typing import List, Sequence
+
+import torch
+import torch.nn.functional as F
+
+
+def check_embeddings_close(
+    *,
+    embeddings_0_lst: Sequence[List[float]],
+    embeddings_1_lst: Sequence[List[float]],
+    name_0: str,
+    name_1: str,
+    tol: float = 1e-3,
+) -> None:
+    assert len(embeddings_0_lst) == len(embeddings_1_lst)
+
+    for prompt_idx, (embeddings_0, embeddings_1) in enumerate(
+            zip(embeddings_0_lst, embeddings_1_lst)):
+        assert len(embeddings_0) == len(embeddings_1)
+
+        sim = F.cosine_similarity(torch.tensor(embeddings_0),
+                                  torch.tensor(embeddings_1),
+                                  dim=0)
+
+        fail_msg = (f"Test{prompt_idx}:"
+                    f"\n{name_0}:\t{embeddings_0!r}"
+                    f"\n{name_1}:\t{embeddings_1!r}")
+
+        assert sim >= 1 - tol, fail_msg