dosubot[bot]
commented
1 week ago Hello @k3ntar0! I'm here to help you with any bugs, questions, or contributions you may have. Let's work together to solve the issue you're experiencing.
To resolve the issue where the text generation process using Azure OpenAI GPT-4 in your self-hosted Dify version 0.6.12-fix1 setup is interrupted with the error 'NoneType' object has no attribute 'tool_calls', you need to ensure that the tool_calls attribute is properly initialized and handled in the _handle_generate_response and _handle_generate_stream_response methods.
The error suggests that the tool_calls attribute is being accessed on a NoneType object, which means that the tool_calls attribute is not being set correctly in some cases. Here is a potential fix:
- Ensure that
tool_callsis initialized as an empty list if it is not provided. - Update the
_handle_generate_responseand_handle_generate_stream_responsemethods to handle cases wheretool_callsmight beNone.
Here is the updated code:
def _handle_generate_response(self, model: str, credentials: dict, response: requests.Response,
prompt_messages: list[PromptMessage]) -> LLMResult:
response_json = response.json()
completion_type = LLMMode.value_of(credentials['mode'])
output = response_json['choices'][0]
response_content = ''
tool_calls = None
function_calling_type = credentials.get('function_calling_type', 'no_call')
if completion_type is LLMMode.CHAT:
response_content = output.get('message', {}).get('content', '')
if function_calling_type == 'tool_call':
tool_calls = output.get('message', {}).get('tool_calls', [])
elif function_calling_type == 'function_call':
tool_calls = output.get('message', {}).get('function_call', [])
elif completion_type is LLMMode.COMPLETION:
response_content = output.get('text', '')
assistant_message = AssistantPromptMessage(content=response_content, tool_calls=[])
if tool_calls:
if function_calling_type == 'tool_call':
assistant_message.tool_calls = self._extract_response_tool_calls(tool_calls)
elif function_calling_type == 'function_call':
assistant_message.tool_calls = [self._extract_response_function_call(tool_calls)]
usage = response_json.get("usage")
if usage:
# transform usage
prompt_tokens = usage["prompt_tokens"]
completion_tokens = usage["completion_tokens"]
else:
# calculate num tokens
prompt_tokens = self._num_tokens_from_string(model, prompt_messages[0].content)
completion_tokens = self._num_tokens_from_string(model, assistant_message.content)
# transform usage
usage = self._calc_response_usage(model, credentials, prompt_tokens, completion_tokens)
# transform response
result = LLMResult(
model=response_json["model"],
prompt_messages=prompt_messages,
message=assistant_message,
usage=usage,
)
return result
def _handle_generate_stream_response(self, model: str, credentials: dict, response: requests.Response,
prompt_messages: list[PromptMessage]) -> Generator:
full_assistant_content = ''
chunk_index = 0
def create_final_llm_result_chunk(index: int, message: AssistantPromptMessage, finish_reason: str) \
-> LLMResultChunk:
# calculate num tokens
prompt_tokens = self._num_tokens_from_string(model, prompt_messages[0].content)
completion_tokens = self._num_tokens_from_string(model, full_assistant_content)
# transform usage
usage = self._calc_response_usage(model, credentials, prompt_tokens, completion_tokens)
return LLMResultChunk(
model=model,
prompt_messages=prompt_messages,
delta=LLMResultChunkDelta(
index=index,
message=message,
finish_reason=finish_reason,
usage=usage
)
)
# delimiter for stream response, need unicode_escape
import codecs
delimiter = credentials.get("stream_mode_delimiter", "\n\n")
delimiter = codecs.decode(delimiter, "unicode_escape")
tools_calls: list[AssistantPromptMessage.ToolCall] = []
def increase_tool_call(new_tool_calls: list[AssistantPromptMessage.ToolCall]):
def get_tool_call(tool_call_id: str):
if not tool_call_id:
return tools_calls[-1]
tool_call = next((tool_call for tool_call in tools_calls if tool_call.id == tool_call_id), None)
if tool_call is None:
tool_call = AssistantPromptMessage.ToolCall(
id=tool_call_id,
type="function",
function=AssistantPromptMessage.ToolCall.ToolCallFunction(
name="",
arguments=""
)
)
tools_calls.append(tool_call)
return tool_call
for new_tool_call in new_tool_calls:
# get tool call
tool_call = get_tool_call(new_tool_call.function.name)
# update tool call
if new_tool_call.id:
tool_call.id = new_tool_call.id
if new_tool_call.type:
tool_call.type = new_tool_call.type
if new_tool_call.function.name:
tool_call.function.name = new_tool_call.function.name
if new_tool_call.function.arguments:
tool_call.function.arguments += new_tool_call.function.arguments
finish_reason = 'Unknown'
for chunk in response.iter_lines(decode_unicode=True, delimiter=delimiter):
chunk = chunk.strip()
if chunk:
# ignore sse comments
if chunk.startswith(':'):
continue
decoded_chunk = chunk.strip().lstrip('data: ').lstrip()
try:
chunk_json = json.loads(decoded_chunk)
# stream ended
except json.JSONDecodeError as e:
yield create_final_llm_result_chunk(
index=chunk_index + 1,
message=AssistantPromptMessage(content=""),
finish_reason="Non-JSON encountered."
)
break
if not chunk_json or len(chunk_json['choices']) == 0:
continue
choice = chunk_json['choices'][0]
finish_reason = chunk_json['choices'][0].get('finish_reason')
chunk_index += 1
if 'delta' in choice:
delta = choice['delta']
delta_content = delta.get('content')
assistant_message_tool_calls = None
if 'tool_calls' in delta and credentials.get('function_calling_type', 'no_call') == 'tool_call':
assistant_message_tool_calls = delta.get('tool_calls', None)
elif 'function_call' in delta and credentials.get('function_calling_type', 'no_call') == 'function_call':
assistant_message_tool_calls = [{
'id': 'tool_call_id',
'type': 'function',
'function': delta.get('function_call', {})
}]
# extract tool calls from response
if assistant_message_tool_calls:
tool_calls = self._extract_response_tool_calls(assistant_message_tool_calls)
increase_tool_call(tool_calls)
if delta_content is None or delta_content == '':
continue
# transform assistant message to prompt message
assistant_prompt_message = AssistantPromptMessage(
content=delta_content,
)
# reset tool calls
tool_calls = []
full_assistant_content += delta_content
elif 'text' in choice:
choice_text = choice.get('text', '')
if choice_text == '':
continue
# transform assistant message to prompt message
assistant_prompt_message = AssistantPromptMessage(content=choice_text)
full_assistant_content += choice_text
else:
continue
yield LLMResultChunk(
model=model,
prompt_messages=prompt_messages,
delta=LLMResultChunkDelta(
index=chunk_index,
message=assistant_prompt_message,
)
)
chunk_index += 1
if tools_calls:
yield LLMResultChunk(
model=model,
prompt_messages=prompt_messages,
delta=LLMResultChunkDelta(
index=chunk_index,
message=AssistantPromptMessage(
tool_calls=tools_calls,
content=""
),
)
)
yield create_final_llm_result_chunk(
index=chunk_index,
message=AssistantPromptMessage(content=""),
finish_reason=finish_reason
)By ensuring that tool_calls is initialized as an empty list if it is not provided, you can avoid the 'NoneType' object has no attribute 'tool_calls' error. Additionally, the _extract_response_tool_calls and _extract_response_function_call methods should be used to handle the extraction of tool calls from the response, ensuring that the tool_calls attribute is always properly initialized [1][2][3].
To continue talking to Dosu, mention @dosu.
laipz8200
k3ntar0
AlexSu1108
shelltea
real-zony
jumpei-tsutsui
Self Checks
Dify version
0.6.12-fix1 This issue does not occur in version 0.6.11 or earlier on my environment.
Cloud or Self Hosted
Self Hosted (Docker)
Steps to reproduce
I have implemented a workflow using Azure OpenAI GPT-4o to generate responses. The initial part of the text generation process, where the text is streamed, runs without any issues. However, during the streaming, I encounter an interruption in the generation process with the following error, causing the processing at the LLM node to halt:
[azure_openai] Error: 'NoneType' object has no attribute 'tool_calls'
✔️ Expected Behavior
The generation by the LLM should complete and proceed to the next node in the workflow.
❌ Actual Behavior
The generation by the LLM is interrupted, and the workflow stops.