# -*- coding: utf-8 -*- # Copyright (c) Alibaba, Inc. and its affiliates. import copy # pylint: disable=R1702,too-many-branches,too-many-return-statements # pylint: disable=too-many-statements def merge_single_response( # noqa: E501 parsed_response, accumulated_data, n=1, ): """Merge a single response chunk with accumulated data. Args: parsed_response: The response chunk to merge accumulated_data: Dictionary storing accumulated data for each choice n: Number of expected choices (default 1) Returns: bool or list: True if this response should be yielded normally, False if filtered, or a list of responses for n>1 with non-stop finish reasons """ # Check if all choices have been sent (for n > 1 case) if n > 1 and accumulated_data: all_sent = all( data.get("all_choices_sent", False) for data in accumulated_data.values() if isinstance(data, dict) and "all_choices_sent" in data ) if all_sent: return False # Track usage for each choice index when n > 1 # Each streaming packet contains usage info for one specific choice if ( n > 1 and parsed_response.usage and parsed_response.output and parsed_response.output.choices and len(parsed_response.output.choices) > 0 ): if "usage_by_index" not in accumulated_data: accumulated_data["usage_by_index"] = {} # Get the choice index from the first (and typically only) choice in this packet # noqa: E501 # pylint: disable=line-too-long try: first_choice = parsed_response.output.choices[0] choice_idx = ( first_choice.index if hasattr( first_choice, "index", ) and "index" in first_choice else 0 ) # Store only output_tokens for this choice index if "output_tokens" in parsed_response.usage: accumulated_data["usage_by_index"][choice_idx] = dict( parsed_response.usage, ) except (KeyError, AttributeError, IndexError): pass # Handle output.text accumulation when choices is null if ( parsed_response.output and hasattr(parsed_response.output, "text") and ( not parsed_response.output.choices or parsed_response.output.choices is None ) ): choice_idx = 0 if choice_idx not in accumulated_data: accumulated_data[choice_idx] = { "content": "", "reasoning_content": "", "tool_calls": [], "logprobs": {"content": []}, "finished": False, "finish_reason": None, "all_choices_sent": False, "role": None, } # Accumulate text if not empty if parsed_response.output.text: accumulated_data[choice_idx][ "content" ] += parsed_response.output.text # Always set accumulated content back to response parsed_response.output.text = accumulated_data[choice_idx]["content"] return True # Process each choice in the choices array if parsed_response.output and parsed_response.output.choices: choices = parsed_response.output.choices # Filter out empty choices array if not choices: return False for choice_enum_idx, choice in enumerate(choices): # Use choice.index if available, otherwise use enumerate index try: choice_idx = ( choice.index if hasattr(choice, "index") and "index" in choice else choice_enum_idx ) except (KeyError, AttributeError): choice_idx = choice_enum_idx # Initialize accumulated data for this choice if not exists if choice_idx not in accumulated_data: accumulated_data[choice_idx] = { "content": "", "reasoning_content": "", "tool_calls": [], "logprobs": {"content": []}, "finished": False, "finish_reason": None, "all_choices_sent": False, "role": None, } # Handle message field - create if null if not choice.message: # Create message object with accumulated data choice.message = { "role": accumulated_data[choice_idx]["role"] if accumulated_data[choice_idx]["role"] else "assistant", "content": accumulated_data[choice_idx]["content"], } if accumulated_data[choice_idx]["reasoning_content"]: choice.message["reasoning_content"] = accumulated_data[ choice_idx ]["reasoning_content"] if accumulated_data[choice_idx]["tool_calls"]: choice.message["tool_calls"] = accumulated_data[ choice_idx ]["tool_calls"] else: # Save role if present if hasattr(choice.message, "role") and choice.message.role: accumulated_data[choice_idx]["role"] = choice.message.role # Handle content accumulation if "content" in choice.message: current_content = choice.message.content if current_content: # Check if content is multimodal format if isinstance(current_content, list): # Handle multimodal content (array format) # Initialize accumulated content as array if not already # noqa: E501 if not isinstance( accumulated_data[choice_idx]["content"], list, ): accumulated_data[choice_idx]["content"] = [] # Ensure accumulated content list has enough elements # noqa: E501 while len( accumulated_data[choice_idx]["content"], ) < len(current_content): accumulated_data[choice_idx]["content"].append( {"text": ""}, ) # Merge each content element for content_idx, content_item in enumerate( current_content, ): if ( isinstance(content_item, dict) and "text" in content_item ): if content_item["text"]: # Accumulate text content accumulated_data[choice_idx][ "content" ][content_idx]["text"] += content_item[ "text" ] # Update the current response with accumulated content # noqa: E501 for content_idx in range( len(accumulated_data[choice_idx]["content"]), ): if content_idx < len(choice.message.content): choice.message.content[content_idx][ "text" ] = accumulated_data[choice_idx][ "content" ][ content_idx ][ "text" ] else: # Handle regular content (string format) # Initialize accumulated content as string if isinstance( accumulated_data[choice_idx]["content"], list, ): accumulated_data[choice_idx]["content"] = "" # Accumulate content if not empty accumulated_data[choice_idx][ "content" ] += current_content # Always set accumulated content back to response if not isinstance( accumulated_data[choice_idx]["content"], list, ): choice.message.content = accumulated_data[choice_idx][ "content" ] else: # For multimodal content, ensure message.content # exists if not isinstance(choice.message.content, list): choice.message.content = accumulated_data[ choice_idx ]["content"] # Handle reasoning_content accumulation if "reasoning_content" in choice.message: current_reasoning_content = ( choice.message.reasoning_content ) if current_reasoning_content: accumulated_data[choice_idx][ "reasoning_content" ] += current_reasoning_content # Always set the accumulated reasoning_content back if we # have any, even if current response doesn't have it if accumulated_data[choice_idx]["reasoning_content"]: choice.message.reasoning_content = accumulated_data[ choice_idx ]["reasoning_content"] # Handle tool_calls accumulation if ( "tool_calls" in choice.message and choice.message.tool_calls ): current_tool_calls = choice.message.tool_calls # For each current tool call, accumulate its arguments for current_call in current_tool_calls: if ( isinstance(current_call, dict) and "index" in current_call ): idx = current_call["index"] # Find existing accumulated call with same index existing_call = None for acc_call in accumulated_data[choice_idx][ "tool_calls" ]: if ( isinstance(acc_call, dict) and acc_call.get("index") == idx ): existing_call = acc_call break if existing_call: # Accumulate function fields from current call if ( "function" in current_call and current_call["function"] ): if "function" not in existing_call: existing_call["function"] = {} # Accumulate function.name if "name" in current_call["function"]: if ( "name" not in existing_call["function"] ): existing_call["function"][ "name" ] = "" existing_call["function"][ "name" ] += current_call["function"]["name"] # Accumulate function.arguments if "arguments" in current_call["function"]: if ( "arguments" not in existing_call["function"] ): existing_call["function"][ "arguments" ] = "" existing_call["function"][ "arguments" ] += current_call["function"][ "arguments" ] # Update other fields with latest values existing_call.update( { k: v for k, v in current_call.items() if k != "function" and v }, ) if ( "function" in current_call and current_call["function"] ): existing_call["function"].update( { k: v for k, v in current_call[ "function" ].items() if k not in ["arguments", "name"] and v }, ) else: # Add new tool call accumulated_data[choice_idx][ "tool_calls" ].append(dict(current_call)) # Update choice with accumulated tool_calls choice.message.tool_calls = accumulated_data[choice_idx][ "tool_calls" ] elif accumulated_data[choice_idx]["tool_calls"]: # If current response has no tool_calls but we have # accumulated tool_calls, restore them choice.message.tool_calls = accumulated_data[choice_idx][ "tool_calls" ] # Restore role if we have it if accumulated_data[choice_idx]["role"] and ( not hasattr(choice.message, "role") or not choice.message.role ): choice.message.role = accumulated_data[choice_idx]["role"] # Handle logprobs accumulation (only if logprobs exists) try: if ( "logprobs" in choice and choice.logprobs and isinstance(choice.logprobs, dict) and "content" in choice.logprobs ): current_logprobs_content = choice.logprobs["content"] if current_logprobs_content and isinstance( current_logprobs_content, list, ): # Initialize logprobs content if not exists if "logprobs" not in accumulated_data[choice_idx]: accumulated_data[choice_idx]["logprobs"] = { "content": [], } elif ( "content" not in accumulated_data[choice_idx]["logprobs"] ): accumulated_data[choice_idx]["logprobs"][ "content" ] = [] # Extend the accumulated logprobs content array accumulated_data[choice_idx]["logprobs"][ "content" ].extend(current_logprobs_content) except (KeyError, AttributeError, TypeError): # logprobs field might not exist or be in unexpected format, safely skip # noqa: E501 # pylint: disable=line-too-long pass # Always set accumulated logprobs if we have any if ( accumulated_data[choice_idx]["logprobs"]["content"] and hasattr(choice, "logprobs") and choice.logprobs ): choice.logprobs["content"] = accumulated_data[choice_idx][ "logprobs" ]["content"] # Handle finish_reason for n > 1 case if ( n > 1 and hasattr(choice, "finish_reason") and choice.finish_reason and choice.finish_reason != "null" ): accumulated_data[choice_idx][ "finish_reason" ] = choice.finish_reason accumulated_data[choice_idx]["finished"] = True # Handle n > 1 case: different strategies for different finish_reason if n > 1: # Count finished choices finished_count = sum( 1 for data in accumulated_data.values() if isinstance(data, dict) and data.get("finished", False) ) # Find all finished choices in current packet finished_choices_in_packet = [] for choice in choices: if ( hasattr(choice, "finish_reason") and choice.finish_reason and choice.finish_reason != "null" ): choice_idx = ( choice.index if hasattr(choice, "index") and "index" in choice else 0 ) finish_reason = choice.finish_reason finished_choices_in_packet.append( (choice_idx, finish_reason, choice), ) # No finish_reason in current packet: return as is if not finished_choices_in_packet: return True # Get finish_reason type from first finished choice first_finish_reason = finished_choices_in_packet[0][1] # For stop: wait all choices, then merge into one result if first_finish_reason == "stop": if finished_count < n: # Hide finish_reason until all finished for choice in choices: if ( hasattr(choice, "finish_reason") and choice.finish_reason and choice.finish_reason != "null" ): choice.finish_reason = "null" else: # All finished: merge all choices into one result for data in accumulated_data.values(): if ( isinstance(data, dict) and "all_choices_sent" in data ): data["all_choices_sent"] = True # Return final result with all choices all_choices = [] # Sort by choice_idx to ensure correct order sorted_items = sorted( [ (idx, data) for idx, data in accumulated_data.items() if isinstance(data, dict) and "finished" in data ], key=lambda x: x[0], ) for choice_idx, data in sorted_items: # Create a new choice object final_choice_dict = { "index": choice_idx, "finish_reason": data["finish_reason"], } # Create message message_dict = { "role": data["role"] if data["role"] else "assistant", } if data["content"]: message_dict["content"] = ( data["content"] if isinstance(data["content"], str) else data["content"] ) if data["reasoning_content"]: message_dict["reasoning_content"] = data[ "reasoning_content" ] if data["tool_calls"]: message_dict["tool_calls"] = data["tool_calls"] final_choice_dict["message"] = message_dict # Add logprobs if present if data["logprobs"]["content"]: final_choice_dict["logprobs"] = { "content": data["logprobs"]["content"], } all_choices.append(final_choice_dict) # Update output choices with all accumulated choices parsed_response.output.choices = all_choices # Aggregate usage from all choice indices if ( "usage_by_index" in accumulated_data and accumulated_data["usage_by_index"] ): aggregated_usage = {} usage_by_idx = accumulated_data["usage_by_index"] # Sum output_tokens and recalculate total_tokens total_output_tokens = 0 input_tokens = None prompt_tokens_details = None for idx, usage in usage_by_idx.items(): if "output_tokens" in usage: total_output_tokens += usage["output_tokens"] # input_tokens should be the same for all indices if ( input_tokens is None and "input_tokens" in usage ): input_tokens = usage["input_tokens"] # Keep prompt_tokens_details from any index # (should be same) if ( prompt_tokens_details is None and "prompt_tokens_details" in usage ): prompt_tokens_details = usage[ "prompt_tokens_details" ] # Build aggregated usage if input_tokens is not None: aggregated_usage["input_tokens"] = input_tokens aggregated_usage["output_tokens"] = total_output_tokens if input_tokens is not None: aggregated_usage["total_tokens"] = ( input_tokens + total_output_tokens ) if prompt_tokens_details is not None: aggregated_usage[ "prompt_tokens_details" ] = prompt_tokens_details # Update response usage with aggregated values parsed_response.usage = aggregated_usage else: # For non-stop (e.g., tool_calls): output each choice separately # noqa: E501 responses_to_yield = [] for ( choice_idx, finish_reason, choice, ) in finished_choices_in_packet: current_data = accumulated_data.get(choice_idx) if current_data is None or current_data.get( "all_choices_sent", False, ): continue current_data["all_choices_sent"] = True # Create a new response for this choice if responses_to_yield: # Clone the response for additional choices new_response = copy.deepcopy(parsed_response) else: # Use the original response for the first choice new_response = parsed_response # Deep copy choice to avoid modifying accumulated_data choice_copy = copy.deepcopy(choice) # Set only this choice in the response new_response.output.choices = [choice_copy] # Update usage with this choice's output tokens if ( new_response.usage and "usage_by_index" in accumulated_data and choice_idx in accumulated_data["usage_by_index"] ): current_usage = accumulated_data["usage_by_index"][ choice_idx ] if "output_tokens" in current_usage: new_response.usage[ "output_tokens" ] = current_usage["output_tokens"] if "input_tokens" in current_usage: new_response.usage["total_tokens"] = ( current_usage["input_tokens"] + current_usage["output_tokens"] ) responses_to_yield.append(new_response) # Return list of responses if we have any if responses_to_yield: return responses_to_yield else: return False return True # pylint: disable=R1702,too-many-branches,too-many-statements def merge_multimodal_single_response( # noqa: E501 parsed_response, accumulated_data, n=1, ): """Merge a single response chunk with accumulated data. Args: parsed_response: The response chunk to merge accumulated_data: Dictionary storing accumulated data for each choice n: Number of expected choices (default 1) Returns: bool: True if this response should be yielded, False if filtered """ # Check if all choices have been sent (for n > 1 case) if n > 1 and accumulated_data: all_sent = any( data.get("all_choices_sent", False) for data in accumulated_data.values() ) if all_sent: return False # Track usage for each choice index when n > 1 # Each streaming packet contains usage info for one specific choice if ( n > 1 and parsed_response.usage and parsed_response.output and parsed_response.output.choices and len(parsed_response.output.choices) > 0 ): if "usage_by_index" not in accumulated_data: accumulated_data["usage_by_index"] = {} # Get the choice index from the first (and typically only) choice in this packet # noqa: E501 # pylint: disable=line-too-long try: first_choice = parsed_response.output.choices[0] choice_idx = ( first_choice.index if hasattr( first_choice, "index", ) and "index" in first_choice else 0 ) # Store only output_tokens for this choice index if "output_tokens" in parsed_response.usage: accumulated_data["usage_by_index"][choice_idx] = dict( parsed_response.usage, ) except (KeyError, AttributeError, IndexError): pass # Handle output.text accumulation when choices is null if ( parsed_response.output and hasattr(parsed_response.output, "text") and ( not parsed_response.output.choices or parsed_response.output.choices is None ) ): choice_idx = 0 if choice_idx not in accumulated_data: accumulated_data[choice_idx] = { "content": "", "reasoning_content": "", "tool_calls": [], "logprobs": {"content": []}, "finished": False, "finish_reason": None, "all_choices_sent": False, "role": None, } # Accumulate text if not empty if parsed_response.output.text: accumulated_data[choice_idx][ "content" ] += parsed_response.output.text # Always set accumulated content back to response parsed_response.output.text = accumulated_data[choice_idx]["content"] return True # Process each choice in the choices array if parsed_response.output and parsed_response.output.choices: choices = parsed_response.output.choices # Filter out empty choices array if not choices: return False for choice_enum_idx, choice in enumerate(choices): # Use choice.index if available, otherwise use enumerate index try: choice_idx = ( choice.index if hasattr(choice, "index") and "index" in choice else choice_enum_idx ) except (KeyError, AttributeError): choice_idx = choice_enum_idx # Initialize accumulated data for this choice if not exists if choice_idx not in accumulated_data: accumulated_data[choice_idx] = { "content": "", "reasoning_content": "", "tool_calls": [], "logprobs": {"content": []}, "finished": False, "finish_reason": None, "all_choices_sent": False, "role": None, } # Handle message field - create if null if not choice.message: # Create message object with accumulated data choice.message = { "role": accumulated_data[choice_idx]["role"] if accumulated_data[choice_idx]["role"] else "assistant", "content": accumulated_data[choice_idx]["content"], } if accumulated_data[choice_idx]["reasoning_content"]: choice.message["reasoning_content"] = accumulated_data[ choice_idx ]["reasoning_content"] if accumulated_data[choice_idx]["tool_calls"]: choice.message["tool_calls"] = accumulated_data[ choice_idx ]["tool_calls"] else: # Save role if present if hasattr(choice.message, "role") and choice.message.role: accumulated_data[choice_idx]["role"] = choice.message.role # Handle content accumulation if "content" in choice.message: current_content = choice.message.content # Check if content is multimodal format if isinstance(current_content, list): # Handle multimodal content (array format) # Initialize accumulated content as array if not already # noqa: E501 if not isinstance( accumulated_data[choice_idx]["content"], list, ): accumulated_data[choice_idx]["content"] = [] # Only process if current_content is not empty if current_content: # Ensure accumulated content list has enough elements # noqa: E501 while len( accumulated_data[choice_idx]["content"], ) < len(current_content): accumulated_data[choice_idx]["content"].append( {"text": ""}, ) # Merge each content element for content_idx, content_item in enumerate( current_content, ): if ( isinstance(content_item, dict) and "text" in content_item ): if content_item["text"]: # Accumulate text content accumulated_data[choice_idx][ "content" ][content_idx]["text"] += content_item[ "text" ] # Always set accumulated content back to response choice.message.content = accumulated_data[choice_idx][ "content" ] elif current_content: # Handle regular content (string format) # Initialize accumulated content as string if isinstance( accumulated_data[choice_idx]["content"], list, ): accumulated_data[choice_idx]["content"] = "" # Accumulate content if not empty accumulated_data[choice_idx][ "content" ] += current_content # Set accumulated content back to response choice.message.content = accumulated_data[choice_idx][ "content" ] elif ( not current_content and accumulated_data[choice_idx]["content"] ): # Current content is empty but we have accumulated content, restore it # noqa: E501 # pylint: disable=line-too-long choice.message.content = accumulated_data[choice_idx][ "content" ] # Handle reasoning_content accumulation if "reasoning_content" in choice.message: current_reasoning_content = ( choice.message.reasoning_content ) if current_reasoning_content: accumulated_data[choice_idx][ "reasoning_content" ] += current_reasoning_content # Always set the accumulated reasoning_content back if we # have any, even if current response doesn't have it if accumulated_data[choice_idx]["reasoning_content"]: choice.message.reasoning_content = accumulated_data[ choice_idx ]["reasoning_content"] # Handle tool_calls accumulation if ( "tool_calls" in choice.message and choice.message.tool_calls ): current_tool_calls = choice.message.tool_calls # For each current tool call, accumulate its arguments for current_call in current_tool_calls: if ( isinstance(current_call, dict) and "index" in current_call ): idx = current_call["index"] # Find existing accumulated call with same index existing_call = None for acc_call in accumulated_data[choice_idx][ "tool_calls" ]: if ( isinstance(acc_call, dict) and acc_call.get("index") == idx ): existing_call = acc_call break if existing_call: # Accumulate function fields from current call if ( "function" in current_call and current_call["function"] ): if "function" not in existing_call: existing_call["function"] = {} # Accumulate function.name if "name" in current_call["function"]: if ( "name" not in existing_call["function"] ): existing_call["function"][ "name" ] = "" existing_call["function"][ "name" ] += current_call["function"]["name"] # Accumulate function.arguments if "arguments" in current_call["function"]: if ( "arguments" not in existing_call["function"] ): existing_call["function"][ "arguments" ] = "" existing_call["function"][ "arguments" ] += current_call["function"][ "arguments" ] # Update other fields with latest values existing_call.update( { k: v for k, v in current_call.items() if k != "function" and v }, ) if ( "function" in current_call and current_call["function"] ): existing_call["function"].update( { k: v for k, v in current_call[ "function" ].items() if k not in ["arguments", "name"] and v }, ) else: # Add new tool call accumulated_data[choice_idx][ "tool_calls" ].append(dict(current_call)) # Update choice with accumulated tool_calls choice.message.tool_calls = accumulated_data[choice_idx][ "tool_calls" ] elif accumulated_data[choice_idx]["tool_calls"]: # If current response has no tool_calls but we have accumulated tool_calls, restore them # noqa: E501 # pylint: disable=line-too-long choice.message.tool_calls = accumulated_data[choice_idx][ "tool_calls" ] # Restore role if we have it if accumulated_data[choice_idx]["role"] and ( not hasattr(choice.message, "role") or not choice.message.role ): choice.message.role = accumulated_data[choice_idx]["role"] # Handle logprobs accumulation (only if logprobs exists) try: if ( "logprobs" in choice and choice.logprobs and isinstance(choice.logprobs, dict) and "content" in choice.logprobs ): current_logprobs_content = choice.logprobs["content"] if current_logprobs_content and isinstance( current_logprobs_content, list, ): # Initialize logprobs content if not exists if "logprobs" not in accumulated_data[choice_idx]: accumulated_data[choice_idx]["logprobs"] = { "content": [], } elif ( "content" not in accumulated_data[choice_idx]["logprobs"] ): accumulated_data[choice_idx]["logprobs"][ "content" ] = [] # Extend the accumulated logprobs content array accumulated_data[choice_idx]["logprobs"][ "content" ].extend(current_logprobs_content) except (KeyError, AttributeError, TypeError): # logprobs field might not exist or be in unexpected format, safely skip # noqa: E501 # pylint: disable=line-too-long pass # Always set accumulated logprobs if we have any if ( accumulated_data[choice_idx]["logprobs"]["content"] and hasattr(choice, "logprobs") and choice.logprobs ): choice.logprobs["content"] = accumulated_data[choice_idx][ "logprobs" ]["content"] # Handle finish_reason for n > 1 case if ( n > 1 and hasattr(choice, "finish_reason") and choice.finish_reason and choice.finish_reason != "null" ): accumulated_data[choice_idx][ "finish_reason" ] = choice.finish_reason accumulated_data[choice_idx]["finished"] = True # Handle n > 1 case: different strategies for different # finish_reason if n > 1: # Count finished choices finished_count = sum( 1 for data in accumulated_data.values() if isinstance(data, dict) and data.get("finished", False) ) # Find all finished choices in current packet finished_choices_in_packet = [] for choice in choices: if ( hasattr(choice, "finish_reason") and choice.finish_reason and choice.finish_reason != "null" ): choice_idx = ( choice.index if hasattr(choice, "index") and "index" in choice else 0 ) finish_reason = choice.finish_reason finished_choices_in_packet.append( (choice_idx, finish_reason, choice), ) # No finish_reason in current packet: return as is if not finished_choices_in_packet: return True # Get finish_reason type from first finished choice first_finish_reason = finished_choices_in_packet[0][1] # For stop: wait all choices, then merge into one result if first_finish_reason == "stop": if finished_count < n: # Hide finish_reason until all finished for choice in choices: if ( hasattr(choice, "finish_reason") and choice.finish_reason and choice.finish_reason != "null" ): choice.finish_reason = "null" else: # All finished: merge all choices into one result for data in accumulated_data.values(): if ( isinstance(data, dict) and "all_choices_sent" in data ): data["all_choices_sent"] = True # Return final result with all choices all_choices = [] # Sort by choice_idx to ensure correct order sorted_items = sorted( [ (idx, data) for idx, data in accumulated_data.items() if isinstance(data, dict) and "finished" in data ], key=lambda x: x[0], ) for choice_idx, data in sorted_items: # Create a new choice object final_choice_dict = { "index": choice_idx, "finish_reason": data["finish_reason"], } # Create message message_dict = { "role": data["role"] if data["role"] else "assistant", } if data["content"]: message_dict["content"] = ( data["content"] if isinstance( data["content"], str, ) else data["content"] ) if data["reasoning_content"]: message_dict["reasoning_content"] = data[ "reasoning_content" ] if data["tool_calls"]: message_dict["tool_calls"] = data["tool_calls"] final_choice_dict["message"] = message_dict # Add logprobs if present if data["logprobs"]["content"]: final_choice_dict["logprobs"] = { "content": data["logprobs"]["content"], } all_choices.append(final_choice_dict) # Update output choices with all accumulated choices parsed_response.output.choices = all_choices # Aggregate usage from all choice indices if ( "usage_by_index" in accumulated_data and accumulated_data["usage_by_index"] ): aggregated_usage = {} usage_by_idx = accumulated_data["usage_by_index"] # Sum output_tokens and recalculate total_tokens total_output_tokens = 0 input_tokens = None prompt_tokens_details = None for idx, usage in usage_by_idx.items(): if "output_tokens" in usage: total_output_tokens += usage["output_tokens"] # input_tokens should be the same for all indices if ( input_tokens is None and "input_tokens" in usage ): input_tokens = usage["input_tokens"] # Keep prompt_tokens_details from any index # (should be same) if ( prompt_tokens_details is None and "prompt_tokens_details" in usage ): prompt_tokens_details = usage[ "prompt_tokens_details" ] # Build aggregated usage if input_tokens is not None: aggregated_usage["input_tokens"] = input_tokens aggregated_usage["output_tokens"] = total_output_tokens if input_tokens is not None: aggregated_usage["total_tokens"] = ( input_tokens + total_output_tokens ) if prompt_tokens_details is not None: aggregated_usage[ "prompt_tokens_details" ] = prompt_tokens_details # Update response usage with aggregated values parsed_response.usage = aggregated_usage else: # For non-stop (e.g., tool_calls): output each choice # separately responses_to_yield = [] for ( choice_idx, finish_reason, choice, ) in finished_choices_in_packet: current_data = accumulated_data.get(choice_idx) if current_data is None or current_data.get( "all_choices_sent", False, ): continue current_data["all_choices_sent"] = True # Create a new response for this choice if responses_to_yield: # Clone the response for additional choices new_response = copy.deepcopy(parsed_response) else: # Use the original response for the first choice new_response = parsed_response # Deep copy choice to avoid modifying accumulated_data choice_copy = copy.deepcopy(choice) # Set only this choice in the response new_response.output.choices = [choice_copy] # Update usage with this choice's output tokens if ( new_response.usage and "usage_by_index" in accumulated_data and choice_idx in accumulated_data["usage_by_index"] ): current_usage = accumulated_data["usage_by_index"][ choice_idx ] if "output_tokens" in current_usage: new_response.usage[ "output_tokens" ] = current_usage["output_tokens"] if "input_tokens" in current_usage: new_response.usage["total_tokens"] = ( current_usage["input_tokens"] + current_usage["output_tokens"] ) responses_to_yield.append(new_response) # Return list of responses if we have any if responses_to_yield: return responses_to_yield else: return False return True