From a2451ef3840dc4b840c828ad72b5507ee63b47ec Mon Sep 17 00:00:00 2001
From: s-zx <2575376715@qq.com>
Date: Wed, 11 Mar 2026 22:10:20 +0100
Subject: [PATCH] fix: drop dimensions in hosted_vllm when drop_params=True and
 model is not text-embedding-3

- Add dimensions drop logic in HostedVLLMEmbeddingConfig.map_openai_params()
- Update test cases to use groq provider per Greptile feedback
---
 .../hosted_vllm/embedding/transformation.py     | 12 ++++++++++--
 tests/test_litellm/test_utils.py                | 17 ++++++-----------
 2 files changed, 16 insertions(+), 13 deletions(-)

diff --git a/litellm/llms/hosted_vllm/embedding/transformation.py b/litellm/llms/hosted_vllm/embedding/transformation.py
index 9c3e8c6c7cc..0f3fbb74cbd 100644
--- a/litellm/llms/hosted_vllm/embedding/transformation.py
+++ b/litellm/llms/hosted_vllm/embedding/transformation.py
@@ -163,8 +163,16 @@ class HostedVLLMEmbeddingConfig(BaseEmbeddingConfig):
         Map OpenAI parameters to Hosted VLLM format.
         """
         for param, value in non_default_params.items():
-            if param in self.get_supported_openai_params(model):
-                optional_params[param] = value
+            if param not in self.get_supported_openai_params(model):
+                continue
+            # Drop dimensions when drop_params=True and model is not text-embedding-3
+            if (
+                param == "dimensions"
+                and drop_params
+                and "text-embedding-3" not in model
+            ):
+                continue
+            optional_params[param] = value
         return optional_params
 
     def get_error_class(
diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py
index 5929a445a30..cf354e88de2 100644
--- a/tests/test_litellm/test_utils.py
+++ b/tests/test_litellm/test_utils.py
@@ -477,17 +477,12 @@ def test_cohere_embedding_optional_params():
 @pytest.mark.parametrize(
     "model,custom_llm_provider,drop_params,expect_dimensions",
     [
-        # hosted_vllm with a non-text-embedding-3 model and drop_params=True → drop
-        ("intfloat/e5-large-v2", "hosted_vllm", True, False),
-        # hosted_vllm with a non-text-embedding-3 model and drop_params=False → keep
-        ("intfloat/e5-large-v2", "hosted_vllm", False, True),
-        # text-embedding-3 model should always keep dimensions
-        ("text-embedding-3-small", "openai_compatible_providers", True, True),
-        ("text-embedding-3-small", "openai_compatible_providers", False, True),
-        # openrouter with a non-text-embedding-3 model and drop_params=True → drop
-        ("jina-embeddings-v3", "openai_compatible_providers", True, False),
-        # openrouter without drop_params → keep (don't block providers that support dimensions)
-        ("jina-embeddings-v3", "openai_compatible_providers", False, True),
+        ("intfloat/e5-large-v2", "groq", True, False),   # drop when drop_params=True
+        ("intfloat/e5-large-v2", "groq", False, True),   # keep when drop_params=False
+        ("text-embedding-3-small", "groq", True, True),  # text-embedding-3 always kept
+        ("text-embedding-3-small", "groq", False, True),
+        ("jina-embeddings-v3", "groq", True, False),
+        ("jina-embeddings-v3", "groq", False, True),
     ],
 )
 def test_openai_compatible_embedding_dimensions_drop_params(