mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-12 08:42:18 +00:00
fix: 精细化国内模型参数过滤
移除过度激进的参数过滤，只保留必要的： 必须过滤： - client_metadata: LiteLLM 内部报错 - coding_plan: Codex 特有参数 - functions/function_call: 旧格式 - LiteLLM 内部参数（shared_session 等） 不再过滤（厂商支持或忽略）： - reasoning_effort: DeepSeek/小米支持 - parallel_tool_calls: 厂商支持 - frequency_penalty, presence_penalty: 厂商支持 - response_format, seed, logprobs 等: 厂商支持/忽略 测试验证： - 小米 mimo-v2.5-pro: 缓存命中正常 - DeepSeek Pro: reasoning_effort 正常 - MiniMax/Doubao: 正常响应 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
oadank
parent
308b38e3f8
commit
c63807c832
1 changed files with 34 additions and 54 deletions
|
|
@ -247,40 +247,22 @@ class LiteLLMCompletionTransformationHandler:
|
|||
actual_model = litellm_params_model
|
||||
|
||||
if is_domestic_model_or_endpoint(actual_model, api_base):
|
||||
# DEBUG logger（先定义，后面使用）
|
||||
# DEBUG logger
|
||||
logger = logging.getLogger("LiteLLM.DomesticFilter")
|
||||
|
||||
# Codex CLI 特有参数
|
||||
# ========== 必须过滤的参数 ==========
|
||||
|
||||
# Codex CLI 特有参数 - LiteLLM 内部报错
|
||||
completion_args.pop("client_metadata", None)
|
||||
# OpenAI 扩展参数（国内模型不支持）
|
||||
completion_args.pop("reasoning_effort", None) # reasoning 模式
|
||||
completion_args.pop("reasoning", None) # reasoning 参数 (Codex 0.130.0+)
|
||||
|
||||
# Responses API 专用参数（国内厂商 Chat API 不支持）
|
||||
completion_args.pop("coding_plan", None) # Codex coding plan
|
||||
completion_args.pop("parallel_tool_calls", None) # 并行工具调用
|
||||
# 旧格式参数（国内模型只支持 tools/tool_choice，不支持 functions/function_call）
|
||||
completion_args.pop("functions", None) # OpenAI 旧格式
|
||||
completion_args.pop("function_call", None) # OpenAI 旧格式
|
||||
# 其他国内模型可能不支持的参数
|
||||
completion_args.pop("stream_options", None) # stream 选项
|
||||
completion_args.pop("modalities", None) # 多模态参数
|
||||
completion_args.pop("prediction", None) # 预测参数
|
||||
completion_args.pop("audio", None) # 音频参数
|
||||
completion_args.pop("store", None) # 存储参数
|
||||
completion_args.pop("include", None) # 包含参数
|
||||
completion_args.pop("prompt_cache_key", None) # 缓存键
|
||||
# OpenAI 高级参数（国内模型可能不支持）
|
||||
completion_args.pop("frequency_penalty", None) # 频率惩罚
|
||||
completion_args.pop("presence_penalty", None) # 存在惩罚
|
||||
completion_args.pop("logprobs", None) # logprobs
|
||||
completion_args.pop("top_logprobs", None) # top_logprobs
|
||||
completion_args.pop("response_format", None) # 响应格式（如 json_object）
|
||||
completion_args.pop("seed", None) # 随机种子
|
||||
completion_args.pop("logit_bias", None) # logit bias
|
||||
completion_args.pop("n", None) # 返回多个结果
|
||||
completion_args.pop("service_tier", None) # 服务等级
|
||||
# 注意：litellm_* 内部参数（litellm_metadata, litellm_logging_obj, litellm_call_id 等）
|
||||
# 必须保留！这些是代理会计/spend attribution 必需的参数，不应删除。
|
||||
# 只有发送给上游 API 的参数才需要过滤，LiteLLM 内部参数会被 litellm.completion 正确处理。
|
||||
|
||||
# 旧格式参数（国内模型只支持 tools/tool_choice）
|
||||
completion_args.pop("functions", None)
|
||||
completion_args.pop("function_call", None)
|
||||
|
||||
# LiteLLM 内部参数（不应发送给上游）
|
||||
completion_args.pop("shared_session", None)
|
||||
completion_args.pop("model_info", None)
|
||||
completion_args.pop("secret_fields", None)
|
||||
|
|
@ -288,35 +270,36 @@ class LiteLLMCompletionTransformationHandler:
|
|||
completion_args.pop("use_litellm_proxy", None)
|
||||
completion_args.pop("merge_reasoning_content_in_choices", None)
|
||||
completion_args.pop("supports_function_calling", None)
|
||||
completion_args.pop("max_retries", None)
|
||||
|
||||
# ========== 厂商忽略的参数（保留过滤，避免无意义传输）==========
|
||||
# 这些参数厂商 Chat API 不支持，发送也不会报错但无意义
|
||||
completion_args.pop("stream_options", None)
|
||||
completion_args.pop("modalities", None)
|
||||
completion_args.pop("prediction", None)
|
||||
completion_args.pop("audio", None)
|
||||
completion_args.pop("store", None)
|
||||
completion_args.pop("include", None)
|
||||
completion_args.pop("prompt_cache_key", None)
|
||||
completion_args.pop("caching", None)
|
||||
completion_args.pop("extra_body", None)
|
||||
# 注意：custom_llm_provider 必须保留，LiteLLM 需要它识别提供商
|
||||
completion_args.pop("max_retries", None)
|
||||
|
||||
# ========== 不过滤的参数（厂商支持或忽略）==========
|
||||
# reasoning_effort - DeepSeek/Xiaomi 支持，其他厂商忽略
|
||||
# parallel_tool_calls - 厂商支持/忽略
|
||||
# frequency_penalty, presence_penalty - 厂商支持
|
||||
# logprobs, top_logprobs - 厂商支持/忽略
|
||||
# response_format - 厂商支持（json_object 等）
|
||||
# seed - 厂商支持/忽略
|
||||
# logit_bias, n, service_tier - 厂商支持/忽略
|
||||
|
||||
# tool_choice 兼容：国内模型不支持 "required"，改成 "auto"
|
||||
tool_choice = completion_args.get("tool_choice")
|
||||
logger.info(f"[DomesticFilter] tool_choice before fix: {tool_choice}")
|
||||
if tool_choice == "required":
|
||||
completion_args["tool_choice"] = "auto"
|
||||
logger.info("[DomesticFilter] tool_choice converted: required -> auto")
|
||||
|
||||
# DEBUG: 国内模型参数过滤验证（脱敏 api_base，只显示域名）
|
||||
# 安全：api_base 可能包含凭证或内部主机名，只显示基础域名
|
||||
safe_api_base = (
|
||||
api_base.split("//")[-1].split("/")[0].split("?")[0]
|
||||
if api_base
|
||||
else "none"
|
||||
)
|
||||
logger.info(
|
||||
f"[DomesticFilter] actual_model={actual_model}, api_base_host={safe_api_base}, "
|
||||
f"is_domestic={is_domestic_model_or_endpoint(actual_model, api_base)}"
|
||||
)
|
||||
logger.info(
|
||||
f"[DomesticFilter] completion_args keys after filter: "
|
||||
f"{list(completion_args.keys())}"
|
||||
)
|
||||
|
||||
# 确保所有历史消息中的 tool_calls arguments 是有效 JSON 格式
|
||||
# 国内模型要求 function.arguments 必须是严格 JSON
|
||||
# 确保 tool_calls arguments 是有效 JSON
|
||||
messages = completion_args.get("messages")
|
||||
if messages:
|
||||
completion_args["messages"] = (
|
||||
|
|
@ -324,9 +307,6 @@ class LiteLLMCompletionTransformationHandler:
|
|||
messages
|
||||
)
|
||||
)
|
||||
logger.info(
|
||||
"[DomesticFilter] Validated tool_calls arguments in messages"
|
||||
)
|
||||
|
||||
litellm_completion_response: Union[
|
||||
ModelResponse, litellm.CustomStreamWrapper
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue