mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-09 03:18:44 +00:00
fix(minimax): handle reasoning field in streaming chunks
- Added MinimaxChatResponseIterator extending OpenAIChatCompletionStreamingHandler - Override chunk_parser to clear content when reasoning/reasoning_content is present - MiniMax uses 'reasoning' field (not 'reasoning_content') in streaming chunks - Handle <think>...</think> tags in content, extract to reasoning_content - Strip stray </think> tags from content when reasoning finishes - Override get_model_response_iterator in MinimaxChatConfig to use custom iterator - Both streaming and non-streaming paths verified working
This commit is contained in:
renearaos
parent
e4a53f50de
commit
8b6dffa619
1 changed files with 93 additions and 29 deletions
|
|
@ -2,12 +2,17 @@
|
|||
MiniMax OpenAI transformation config - extends OpenAI chat config for MiniMax's OpenAI-compatible API
|
||||
"""
|
||||
|
||||
from typing import List, Optional, Tuple
|
||||
import re
|
||||
from typing import Any, Dict, List, Optional, Tuple
|
||||
|
||||
import litellm
|
||||
from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig
|
||||
from litellm.llms.openai.chat.gpt_transformation import (
|
||||
OpenAIGPTConfig,
|
||||
OpenAIChatCompletionStreamingHandler,
|
||||
)
|
||||
from litellm.secret_managers.main import get_secret_str
|
||||
from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam
|
||||
from litellm.types.utils import ModelResponseStream
|
||||
|
||||
|
||||
class MinimaxChatConfig(OpenAIGPTConfig):
|
||||
|
|
@ -25,20 +30,12 @@ class MinimaxChatConfig(OpenAIGPTConfig):
|
|||
|
||||
@staticmethod
|
||||
def get_api_key(api_key: Optional[str] = None) -> Optional[str]:
|
||||
"""
|
||||
Get MiniMax API key from environment or parameters.
|
||||
"""
|
||||
return api_key or get_secret_str("MINIMAX_API_KEY") or litellm.api_key
|
||||
|
||||
@staticmethod
|
||||
def get_api_base(
|
||||
api_base: Optional[str] = None,
|
||||
) -> str:
|
||||
"""
|
||||
Get MiniMax API base URL.
|
||||
Defaults to international endpoint: https://api.minimax.io/v1
|
||||
For China, set to: https://api.minimaxi.com/v1
|
||||
"""
|
||||
return (
|
||||
api_base
|
||||
or get_secret_str("MINIMAX_API_BASE")
|
||||
|
|
@ -54,14 +51,7 @@ class MinimaxChatConfig(OpenAIGPTConfig):
|
|||
litellm_params: dict,
|
||||
stream: Optional[bool] = None,
|
||||
) -> str:
|
||||
"""
|
||||
Get the complete URL for MiniMax OpenAI API.
|
||||
Override to ensure we use MiniMax's endpoint.
|
||||
"""
|
||||
# Get the base URL (either provided or default MiniMax endpoint)
|
||||
base_url = self.get_api_base(api_base=api_base)
|
||||
|
||||
# Ensure it ends with /chat/completions
|
||||
if base_url.endswith("/chat/completions"):
|
||||
return base_url
|
||||
elif base_url.endswith("/v1"):
|
||||
|
|
@ -77,26 +67,100 @@ class MinimaxChatConfig(OpenAIGPTConfig):
|
|||
messages: List[AllMessageValues],
|
||||
tools: Optional[List[ChatCompletionToolParam]] = None,
|
||||
) -> Tuple[List[AllMessageValues], Optional[List[ChatCompletionToolParam]]]:
|
||||
"""
|
||||
Override to preserve cache_control for MiniMax.
|
||||
MiniMax supports cache_control - don't strip it.
|
||||
"""
|
||||
# MiniMax supports cache_control, so return messages and tools unchanged
|
||||
return messages, tools
|
||||
|
||||
def get_supported_openai_params(self, model: str) -> list:
|
||||
"""
|
||||
Get supported OpenAI parameters for MiniMax.
|
||||
Adds reasoning_split and thinking to the list of supported params.
|
||||
"""
|
||||
base_params = super().get_supported_openai_params(model=model)
|
||||
additional_params = ["reasoning_split"]
|
||||
|
||||
# Add thinking parameter if model supports reasoning
|
||||
try:
|
||||
if litellm.supports_reasoning(model=model, custom_llm_provider="minimax"):
|
||||
additional_params.append("thinking")
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
return base_params + additional_params
|
||||
|
||||
def get_model_response_iterator(
|
||||
self,
|
||||
streaming_response: Any,
|
||||
sync_stream: Optional[bool] = None,
|
||||
json_mode: Optional[bool] = None,
|
||||
):
|
||||
return MinimaxChatResponseIterator(
|
||||
streaming_response=streaming_response,
|
||||
sync_stream=sync_stream,
|
||||
json_mode=json_mode,
|
||||
)
|
||||
|
||||
|
||||
class MinimaxChatResponseIterator(OpenAIChatCompletionStreamingHandler):
|
||||
"""
|
||||
MiniMax streaming response iterator that handles reasoning_content extraction.
|
||||
|
||||
MiniMax streaming responses can contain:
|
||||
1. reasoning_details array - converts to reasoning_content
|
||||
2. <think>...</think> tags in content - extracts to reasoning_content
|
||||
3. "reasoning" field (alias for reasoning_content) - clears from content
|
||||
"""
|
||||
|
||||
started_reasoning_content: bool = False
|
||||
finished_reasoning_content: bool = False
|
||||
pending_reasoning: str = ""
|
||||
|
||||
def chunk_parser(self, chunk: dict) -> ModelResponseStream:
|
||||
try:
|
||||
choices = chunk.get("choices", [])
|
||||
|
||||
for choice in choices:
|
||||
delta = choice.get("delta", {})
|
||||
|
||||
# MiniMax uses "reasoning" (not "reasoning_content") in streaming chunks
|
||||
# The parent class maps "reasoning" -> "reasoning_content"
|
||||
# We clear content when either reasoning field is present
|
||||
has_reasoning = (
|
||||
delta.get("reasoning_content")
|
||||
or delta.get("reasoning")
|
||||
)
|
||||
if has_reasoning:
|
||||
delta["content"] = None
|
||||
self.started_reasoning_content = True
|
||||
|
||||
reasoning_details = delta.get("reasoning_details")
|
||||
if reasoning_details and isinstance(reasoning_details, list):
|
||||
reasoning_text = ""
|
||||
for detail in reasoning_details:
|
||||
if isinstance(detail, dict) and detail.get("text"):
|
||||
reasoning_text += detail.get("text", "")
|
||||
if reasoning_text:
|
||||
delta["reasoning_content"] = reasoning_text
|
||||
delta["content"] = None
|
||||
self.started_reasoning_content = True
|
||||
|
||||
content = delta.get("content", "")
|
||||
if content and isinstance(content, str):
|
||||
reasoning_matches = re.findall(
|
||||
r"<think>(.*?)</think>", content, re.DOTALL
|
||||
)
|
||||
if reasoning_matches:
|
||||
self.pending_reasoning += "".join(reasoning_matches)
|
||||
clean_content = re.sub(
|
||||
r"<think>.*?</think>", "", content, flags=re.DOTALL
|
||||
)
|
||||
delta["content"] = clean_content if clean_content else None
|
||||
if self.pending_reasoning:
|
||||
delta["reasoning_content"] = self.pending_reasoning
|
||||
self.started_reasoning_content = True
|
||||
elif self.started_reasoning_content and not self.finished_reasoning_content:
|
||||
self.finished_reasoning_content = True
|
||||
self.pending_reasoning = ""
|
||||
clean_content = re.sub(
|
||||
r"</?think>", "", content
|
||||
)
|
||||
delta["content"] = clean_content if clean_content else None
|
||||
elif not self.started_reasoning_content:
|
||||
if "</think>" in content:
|
||||
clean_content = content.replace("</think>", "").replace("<think>", "")
|
||||
delta["content"] = clean_content if clean_content else None
|
||||
|
||||
return super().chunk_parser(chunk)
|
||||
except Exception:
|
||||
return super().chunk_parser(chunk)
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue