api: use fp32 for base64 embeddings

aphrodite/endpoints/openai/serving_embedding.py

@@ -32,7 +32,9 @@ def request_output_to_embedding_response(
         prompt_token_ids = final_res.prompt_token_ids
         embedding = final_res.outputs.embedding
         if encoding_format == "base64":
-            embedding_bytes = np.array(embedding).tobytes()
+            # Force to use float32 for base64 encoding
+            # to match the OpenAI python client behavior
+            embedding_bytes = np.array(embedding, dtype="float32").tobytes()
             embedding = base64.b64encode(embedding_bytes).decode("utf-8")
         embedding_data = EmbeddingResponseData(index=idx, embedding=embedding)
         data.append(embedding_data)

examples/openai_api/embedding.py

@@ -19,7 +19,6 @@
         "The weather is nice today",
     ],
     model=model,
-    encoding_format="float",
 )
 
 for data in responses.data:

tests/endpoints/openai/test_embedding.py

@@ -128,9 +128,17 @@ async def test_batch_base64_embedding(embedding_client: openai.AsyncOpenAI,
     for data in responses_base64.data:
         decoded_responses_base64_data.append(
             np.frombuffer(base64.b64decode(data.embedding),
-                          dtype="float").tolist())
+                          dtype="float32").tolist())
 
     assert responses_float.data[0].embedding == decoded_responses_base64_data[
         0]
     assert responses_float.data[1].embedding == decoded_responses_base64_data[
         1]
+
+    # Default response is float32 decoded from base64 by OpenAI Client
+    responses_default = await embedding_client.embeddings.create(
+        input=input_texts, model=model_name)
+    assert responses_float.data[0].embedding == responses_default.data[
+        0].embedding
+    assert responses_float.data[1].embedding == responses_default.data[
+        1].embedding