Fix AzureOpenAI streaming token usage

llama-index-integrations/llms/llama-index-llms-openai/llama_index/llms/openai/base.py

@@ -77,7 +77,7 @@
     update_tool_calls,
     is_json_schema_supported,
 )
-from openai import AsyncOpenAI, AzureOpenAI, AsyncAzureOpenAI
+from openai import AsyncOpenAI
 from openai import OpenAI as SyncOpenAI
 from openai.types.chat.chat_completion_chunk import (
     ChatCompletionChunk,
@@ -532,10 +532,12 @@ def gen() -> ChatResponseGen:
                 if len(response.choices) > 0:
                     delta = response.choices[0].delta
                 else:
-                    if isinstance(client, AzureOpenAI):
+                    # Choices could be empty if we receive usage details or content filtering result.
+                    # In the case of usage details, we simply generate an empty delta. The usage delta comes in the last
+                    # chunk of the response and avoid yielding in other cases.
+                    if response.usage is None:
                         continue
-                    else:
-                        delta = ChoiceDelta()
+                    delta = ChoiceDelta()
 
                 if delta is None:
                     continue
@@ -801,10 +803,12 @@ async def gen() -> ChatResponseAsyncGen:
                         continue
                     delta = response.choices[0].delta
                 else:
-                    if isinstance(aclient, AsyncAzureOpenAI):
+                    # Choices could be empty if we receive usage details or content filtering result.
+                    # In the case of usage details, we simply generate an empty delta. The usage delta comes in the last
+                    # chunk of the response and avoid yielding in other cases.
+                    if response.usage is None:
                         continue
-                    else:
-                        delta = ChoiceDelta()
+                    delta = ChoiceDelta()
                 first_chat_chunk = False
 
                 if delta is None:

llama-index-integrations/llms/llama-index-llms-openai/pyproject.toml

@@ -27,7 +27,7 @@ dev = [
 
 [project]
 name = "llama-index-llms-openai"
-version = "0.5.2"
+version = "0.5.3"
 description = "llama-index llms openai integration"
 authors = [{name = "llama-index"}]
 requires-python = ">=3.9,<4.0"