sycommon-python-lib 0.2.5a16__tar.gz → 0.2.5a17__tar.gz

{sycommon_python_lib-0.2.5a16 → sycommon_python_lib-0.2.5a17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.5a16
+Version: 0.2.5a17
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

{sycommon_python_lib-0.2.5a16 → sycommon_python_lib-0.2.5a17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sycommon-python-lib"
-version = "0.2.5a16"
+version = "0.2.5a17"
 description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.11"

{sycommon_python_lib-0.2.5a16 → sycommon_python_lib-0.2.5a17}/src/sycommon/agent/summarization_utils.py

@@ -62,6 +62,7 @@ def _patched_compute_summarization_defaults(model):
 
     # 60% 触发（120K/200K）：实测模型在 input≈137K 时开始退化，
     # 在 120K 触发压缩留 ~17K 安全余量给工具 schema 等未计入开销
+    # output 预算由 LLMWithTokenTracking._clamp_max_tokens 动态管理，不在此预留
     trigger = int(max_tokens * 0.60)
     keep = int(max_tokens * 0.10)
     return {
@@ -168,6 +169,7 @@ def build_summarization_middleware(
     except Exception:
         max_tokens = default_max_tokens
 
+    # output 预算由 LLMWithTokenTracking._clamp_max_tokens 动态管理，不在此预留
     trigger_tokens = int(max_tokens * trigger_fraction)
     keep_tokens = int(max_tokens * keep_fraction)
 

{sycommon_python_lib-0.2.5a16 → sycommon_python_lib-0.2.5a17}/src/sycommon/llm/llm_with_token_tracking.py

@@ -96,11 +96,35 @@ class LLMWithTokenTracking(BaseChatModel):
             from sycommon.logging.kafka_log import SYLogger
             SYLogger.warning(f"Token 使用量记录失败: {e}")
 
+    def _clamp_max_tokens(self, messages, **kwargs):
+        """动态调整 max_tokens，确保 input + max_output 不超过上下文窗口
+
+        每次 LLM 调用前，估算当前 input tokens，
+        将 max_tokens = min(配置的 maxOutputTokens, context_window - input - buffer)
+        """
+        if not self.llmConfig or not self.llmConfig.maxTokens:
+            return
+        max_output = getattr(self.llm, 'max_tokens', None)
+        if not max_output:
+            return
+
+        from langchain_core.messages.utils import count_tokens_approximately
+        try:
+            input_est = count_tokens_approximately(messages, chars_per_token=2.0)
+        except Exception:
+            return
+
+        buffer = 4096
+        available = self.llmConfig.maxTokens - input_est - buffer
+        dynamic_max = min(max_output, max(available, 1024))
+        self.llm.max_tokens = dynamic_max
+
     def _generate(self, messages, stop=None, run_manager=None, **kwargs):
         """
         同步生成 - 核心方法
         BaseChatModel 的 invoke/batch/stream 最终都会调用此方法
         """
+        self._clamp_max_tokens(messages, **kwargs)
         result = self.llm._generate(
             messages, stop=stop, run_manager=run_manager, **kwargs)
         return self._inject_token_usage(result, is_async=False)
@@ -110,6 +134,7 @@ class LLMWithTokenTracking(BaseChatModel):
         异步生成 - 核心方法
         BaseChatModel 的 ainvoke/abatch/astream 最终都会调用此方法
         """
+        self._clamp_max_tokens(messages, **kwargs)
         result = await self.llm._agenerate(messages, stop=stop, run_manager=run_manager, **kwargs)
         return self._inject_token_usage(result, is_async=True)
 

{sycommon_python_lib-0.2.5a16 → sycommon_python_lib-0.2.5a17}/src/sycommon/llm/usage_token.py

@@ -164,12 +164,35 @@ class LLMWithAutoTokenUsage(BaseChatModel):
             is_native_mode=False
         )
 
+    # ========== 动态 max_tokens 管理 ==========
+
+    def _clamp_max_tokens(self, messages, **kwargs):
+        """动态调整 max_tokens，确保 input + max_output 不超过上下文窗口"""
+        if not self.llmConfig or not self.llmConfig.maxTokens:
+            return
+        max_output = getattr(self.llm, 'max_tokens', None)
+        if not max_output:
+            return
+
+        from langchain_core.messages.utils import count_tokens_approximately
+        try:
+            input_est = count_tokens_approximately(messages, chars_per_token=2.0)
+        except Exception:
+            return
+
+        buffer = 4096
+        available = self.llmConfig.maxTokens - input_est - buffer
+        dynamic_max = min(max_output, max(available, 1024))
+        self.llm.max_tokens = dynamic_max
+
     # ========== 实现BaseChatModel抽象方法 ==========
     def _generate(self, messages, stop=None, run_manager=None, **kwargs):
+        self._clamp_max_tokens(messages, **kwargs)
         return self.llm._generate(messages, stop=stop, run_manager=run_manager, **kwargs)
 
     async def _agenerate(self, messages, stop=None, run_manager=None, **kwargs):
         """异步生成 - 委托给底层 LLM"""
+        self._clamp_max_tokens(messages, **kwargs)
         return await self.llm._agenerate(messages, stop=stop, run_manager=run_manager, **kwargs)
 
     def _stream(self, messages, stop=None, run_manager=None, **kwargs):

sycommon_python_lib-0.2.5a17/src/sycommon/tests/test_dynamic_max_tokens.py

@@ -0,0 +1,372 @@
+"""
+测试动态 max_tokens 调整机制
+
+验证 LLMWithTokenTracking 和 LLMWithAutoTokenUsage 的 _clamp_max_tokens 逻辑：
+- 每次 LLM 调用前动态计算 max_tokens = min(配置值, context_window - input - buffer)
+- 确保 input + max_output 永远不超过模型上下文窗口
+
+同时验证通过 create_deep_agent 创建的 agent 能正常应用动态计算规则。
+"""
+
+import asyncio
+import sys
+import os
+import pytest
+
+# 添加项目路径
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from unittest.mock import MagicMock, patch, AsyncMock
+from langchain_core.messages import HumanMessage, AIMessage, SystemMessage
+from langchain_core.language_models import BaseChatModel
+from langchain_core.outputs import ChatResult, ChatGeneration
+
+from sycommon.config.LLMConfig import LLMConfig
+from sycommon.llm.llm_with_token_tracking import LLMWithTokenTracking
+from sycommon.llm.usage_token import LLMWithAutoTokenUsage
+
+
+def _make_llm_config(max_tokens=196608, max_output=128000):
+    return LLMConfig(
+        model="glm-5.1",
+        provider="openai",
+        baseUrl="http://10.10.6.132:3000/v1",
+        apiKey="sk-test",
+        maxTokens=max_tokens,
+        vision=False,
+        callFunction=True,
+        maxOutputTokens=max_output,
+    )
+
+
+class MockChatModel(BaseChatModel):
+    """最小 mock，满足 pydantic 校验"""
+    max_tokens: int = 128000
+
+    @property
+    def _llm_type(self) -> str:
+        return "mock"
+
+    def _generate(self, messages, stop=None, run_manager=None, **kwargs):
+        return ChatResult(generations=[ChatGeneration(message=AIMessage(content="ok"))])
+
+    async def _agenerate(self, messages, stop=None, run_manager=None, **kwargs):
+        return ChatResult(generations=[ChatGeneration(message=AIMessage(content="ok"))])
+
+
+def _make_long_messages(target_chars):
+    """生成指定字符数的消息列表（模拟大 input）"""
+    msg = "这是一段测试文本。" * (target_chars // 8)
+    return [
+        SystemMessage(content="你是一个智能助手"),
+        HumanMessage(content="你好"),
+        AIMessage(content=msg),
+        HumanMessage(content="请回答我"),
+    ]
+
+
+# ============== LLMWithTokenTracking 测试 ==============
+
+
+class TestLLMWithTokenTrackingClamp:
+    """测试 LLMWithTokenTracking._clamp_max_tokens"""
+
+    def test_small_input_keeps_max_output(self):
+        """input 小时，max_tokens 保持配置的 maxOutputTokens"""
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithTokenTracking(llm=llm, llmConfig=config)
+
+        messages = [HumanMessage(content="你好")]
+        wrapper._clamp_max_tokens(messages)
+
+        # input 很小，available = 196608 - ~2 - 4096 ≈ 192510，min(128000, 192510) = 128000
+        assert llm.max_tokens == 128000
+
+    def test_large_input_reduces_max_output(self):
+        """input 大时，max_tokens 自动缩小"""
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithTokenTracking(llm=llm, llmConfig=config)
+
+        # 模拟 ~100K 字符 ≈ ~50K tokens 的 input
+        messages = _make_long_messages(100000)
+        wrapper._clamp_max_tokens(messages)
+
+        # available = 196608 - ~50K - 4096 ≈ 142K, min(128000, 142K) = 128000
+        assert llm.max_tokens == 128000
+
+    def test_near_limit_input_clamps_hard(self):
+        """input 接近上限时，max_tokens 被大幅压缩"""
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithTokenTracking(llm=llm, llmConfig=config)
+
+        # 模拟 ~160K 字符 ≈ ~80K tokens 的 input
+        messages = _make_long_messages(160000)
+        wrapper._clamp_max_tokens(messages)
+
+        # available = 196608 - ~80K - 4096 ≈ 112K, min(128000, 112K) = ~112K
+        assert llm.max_tokens < 128000
+        assert llm.max_tokens > 1024  # 至少保留 1024
+
+    def test_extremely_large_input_floors_at_1024(self):
+        """input 超过上限时，max_tokens 至少保留 1024"""
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithTokenTracking(llm=llm, llmConfig=config)
+
+        # 模拟 ~400K 字符 ≈ ~200K tokens 的 input（超过上下文窗口）
+        messages = _make_long_messages(400000)
+        wrapper._clamp_max_tokens(messages)
+
+        assert llm.max_tokens == 1024
+
+    def test_no_max_output_configured_skips(self):
+        """没有配置 maxOutputTokens 时跳过调整"""
+        llm = MockChatModel(max_tokens=0)
+        config = _make_llm_config(max_tokens=196608, max_output=0)
+        wrapper = LLMWithTokenTracking(llm=llm, llmConfig=config)
+
+        original_max = llm.max_tokens
+        messages = [HumanMessage(content="你好")]
+        wrapper._clamp_max_tokens(messages)
+
+        assert llm.max_tokens == original_max
+
+    def test_generate_calls_clamp(self):
+        """_generate 调用前会触发 _clamp_max_tokens"""
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithTokenTracking(llm=llm, llmConfig=config)
+
+        wrapper._generate([HumanMessage(content="测试")])
+
+        # _clamp_max_tokens 被调用，llm.max_tokens 可能被修改
+        # 关键是没抛异常
+        assert llm.max_tokens > 0
+
+
+# ============== LLMWithAutoTokenUsage 测试 ==============
+
+
+class TestLLMWithAutoTokenUsageClamp:
+    """测试 LLMWithAutoTokenUsage._clamp_max_tokens"""
+
+    def test_small_input_keeps_max_output(self):
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithAutoTokenUsage(llm=llm, langfuse=None, llmConfig=config, summary_prompt=None)
+
+        messages = [HumanMessage(content="你好")]
+        wrapper._clamp_max_tokens(messages)
+
+        assert llm.max_tokens == 128000
+
+    def test_large_input_reduces_max_output(self):
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithAutoTokenUsage(llm=llm, langfuse=None, llmConfig=config, summary_prompt=None)
+
+        messages = _make_long_messages(160000)
+        wrapper._clamp_max_tokens(messages)
+
+        assert llm.max_tokens < 128000
+        assert llm.max_tokens > 1024
+
+    def test_agenerate_calls_clamp(self):
+        llm = MockChatModel(max_tokens=128000)
+        config = _make_llm_config(max_tokens=196608, max_output=128000)
+        wrapper = LLMWithAutoTokenUsage(llm=llm, langfuse=None, llmConfig=config, summary_prompt=None)
+
+        asyncio.get_event_loop().run_until_complete(
+            wrapper._agenerate([HumanMessage(content="测试")])
+        )
+
+        assert llm.max_tokens > 0
+
+
+# ============== create_deep_agent 集成测试 ==============
+
+
+@pytest.mark.asyncio
+async def test_deep_agent_dynamic_max_tokens():
+    """验证通过 create_deep_agent 创建的 agent 走 get_llm -> LLMWithTokenTracking 路径
+
+    模拟场景：多轮对话后 input 累积到接近上下文窗口上限，
+    验证 _clamp_max_tokens 能正确缩小 max_tokens，避免 400 报错。
+    """
+    from langchain.chat_models import init_chat_model
+
+    # 用真实 API 创建 model（和 create_deep_agent 内部一致）
+    model = init_chat_model(
+        model="glm-5.1",
+        model_provider="openai",
+        streaming=False,
+        base_url="http://10.10.6.132:3000/v1",
+        api_key="sk-3hXhO7uCG5CRi7t3THE2cWTKVJqesXJ38cKstHZhDVXHQOIn",
+        temperature=0.7,
+        max_tokens=128000,
+    )
+
+    # 包装为 LLMWithTokenTracking（和 get_llm(wrap_structured=False) 一致）
+    config = _make_llm_config(max_tokens=196608, max_output=128000)
+    wrapper = LLMWithTokenTracking(llm=model, llmConfig=config)
+
+    # 第一轮：小 input，max_tokens 应保持 128000
+    messages_1 = [
+        SystemMessage(content="你是智能助手"),
+        HumanMessage(content="你好，简单回答即可"),
+    ]
+    wrapper._clamp_max_tokens(messages_1)
+    max_tokens_1 = model.max_tokens
+    print(f"[轮次1] input 小, max_tokens = {max_tokens_1}")
+    assert max_tokens_1 == 128000, f"小 input 时 max_tokens 应保持 128000，实际 {max_tokens_1}"
+
+    # 第二轮：中等 input
+    messages_2 = messages_1 + [
+        AIMessage(content="你好！有什么可以帮助你的吗？"),
+        HumanMessage(content="请帮我写一首很长的诗"),
+    ]
+    wrapper._clamp_max_tokens(messages_2)
+    max_tokens_2 = model.max_tokens
+    print(f"[轮次2] input 中等, max_tokens = {max_tokens_2}")
+    assert max_tokens_2 == 128000, f"中等 input 时 max_tokens 应保持 128000，实际 {max_tokens_2}"
+
+    # 第三轮：模拟大 input（~200K 字符 ≈ 100K tokens，超过 context - max_output = 68K）
+    big_content = "这是一段很长的对话历史。" * 17000
+    messages_3 = messages_2 + [
+        AIMessage(content=big_content),
+        HumanMessage(content="请总结上面的内容"),
+    ]
+    wrapper._clamp_max_tokens(messages_3)
+    max_tokens_3 = model.max_tokens
+    print(f"[轮次3] input 大(~60K tokens), max_tokens = {max_tokens_3}")
+    assert max_tokens_3 < 128000, f"大 input 时 max_tokens 应缩小，实际 {max_tokens_3}"
+    assert max_tokens_3 > 1024, f"max_tokens 不应低于 1024，实际 {max_tokens_3}"
+
+    # 验证 input + max_tokens 不会超过 context window
+    from langchain_core.messages.utils import count_tokens_approximately
+    input_tokens = count_tokens_approximately(messages_3, chars_per_token=2.0)
+    total = input_tokens + max_tokens_3 + 4096
+    print(f"[轮次3] input_tokens ≈ {input_tokens}, max_output = {max_tokens_3}, total = {total}")
+    assert total <= 196608, f"input + max_output + buffer 不应超过 196608，实际 {total}"
+
+    print("\n所有本地计算测试通过！")
+
+
+@pytest.mark.asyncio
+async def test_real_api_small_input():
+    """真实 API 测试：小 input，正常调用不报错"""
+    from langchain.chat_models import init_chat_model
+
+    model = init_chat_model(
+        model="glm-5.1",
+        model_provider="openai",
+        streaming=False,
+        base_url="http://10.10.6.132:3000/v1",
+        api_key="sk-3hXhO7uCG5CRi7t3THE2cWTKVJqesXJ38cKstHZhDVXHQOIn",
+        temperature=0.7,
+        max_tokens=128000,
+    )
+    config = _make_llm_config(max_tokens=196608, max_output=128000)
+    wrapper = LLMWithTokenTracking(llm=model, llmConfig=config)
+
+    messages = [
+        SystemMessage(content="你是智能助手，用一句话回答"),
+        HumanMessage(content="1+1等于几？"),
+    ]
+    wrapper._clamp_max_tokens(messages)
+    print(f"[真实API-小input] max_tokens={model.max_tokens}")
+
+    result = await wrapper._agenerate(messages)
+    content = result.generations[0].message.content
+    print(f"[真实API-小input] 回复: {content}")
+    assert content, "小 input 应该正常返回"
+    assert model.max_tokens == 128000, "小 input 时 max_tokens 应保持 128000"
+
+
+@pytest.mark.asyncio
+async def test_real_api_large_input():
+    """真实 API 测试：大 input，动态缩小 max_tokens 后不 400"""
+    from langchain.chat_models import init_chat_model
+    from langchain_core.messages.utils import count_tokens_approximately
+
+    model = init_chat_model(
+        model="glm-5.1",
+        model_provider="openai",
+        streaming=False,
+        base_url="http://10.10.6.132:3000/v1",
+        api_key="sk-3hXhO7uCG5CRi7t3THE2cWTKVJqesXJ38cKstHZhDVXHQOIn",
+        temperature=0.7,
+        max_tokens=128000,
+    )
+    config = _make_llm_config(max_tokens=196608, max_output=128000)
+    wrapper = LLMWithTokenTracking(llm=model, llmConfig=config)
+
+    # 构造 ~100K tokens 的 input
+    big_content = "这是一段很长的对话历史。" * 17000
+    messages = [
+        SystemMessage(content="你是智能助手，用一句话回答"),
+        HumanMessage(content="你好"),
+        AIMessage(content=big_content),
+        HumanMessage(content="请用一句话总结上面的内容"),
+    ]
+
+    input_tokens = count_tokens_approximately(messages, chars_per_token=2.0)
+    print(f"[真实API-大input] input_tokens ≈ {input_tokens}")
+
+    wrapper._clamp_max_tokens(messages)
+    print(f"[真实API-大input] clamp后 max_tokens={model.max_tokens}")
+    assert model.max_tokens < 128000, f"大 input 时 max_tokens 应缩小，实际 {model.max_tokens}"
+
+    total = input_tokens + model.max_tokens + 4096
+    print(f"[真实API-大input] total = {input_tokens} + {model.max_tokens} + 4096 = {total}")
+    assert total <= 196608
+
+    result = await wrapper._agenerate(messages)
+    content = result.generations[0].message.content
+    print(f"[真实API-大input] 回复: {content[:100]}")
+    assert content, "大 input + 动态 max_tokens 应该正常返回，不应 400"
+
+
+@pytest.mark.asyncio
+async def test_real_api_without_clamp_would_fail():
+    """对比测试：同样的大 input，不用 _clamp_max_tokens 会 400 报错"""
+    from langchain.chat_models import init_chat_model
+    from openai import BadRequestError
+
+    model = init_chat_model(
+        model="glm-5.1",
+        model_provider="openai",
+        streaming=False,
+        base_url="http://10.10.6.132:3000/v1",
+        api_key="sk-3hXhO7uCG5CRi7t3THE2cWTKVJqesXJ38cKstHZhDVXHQOIn",
+        temperature=0.7,
+        max_tokens=128000,
+    )
+    # 不用 wrapper，直接裸调
+    big_content = "这是一段很长的对话历史。" * 17000
+    messages = [
+        SystemMessage(content="你是智能助手"),
+        HumanMessage(content="你好"),
+        AIMessage(content=big_content),
+        HumanMessage(content="总结一下"),
+    ]
+
+    try:
+        await model._agenerate(messages)
+        print("[对比测试] 未预期地成功了，可能模型上下文足够大或估算偏保守")
+    except (BadRequestError, Exception) as e:
+        error_msg = str(e)
+        print(f"[对比测试] 不 clamp 时果然报错: {error_msg[:200]}")
+        assert "400" in error_msg or "context length" in error_msg.lower() or "token" in error_msg.lower(), \
+            f"期望是 400 token 超限错误，实际: {error_msg[:200]}"
+
+
+if __name__ == "__main__":
+    print("===== 单元测试 =====")
+    pytest.main([__file__, "-v", "-x", "-k", "not test_real_api"])
+
+    print("\n===== 真实 API 测试 =====")
+    pytest.main([__file__, "-v", "-s", "-k", "test_real_api"])

{sycommon_python_lib-0.2.5a16 → sycommon_python_lib-0.2.5a17}/src/sycommon_python_lib.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.5a16
+Version: 0.2.5a17
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

{sycommon_python_lib-0.2.5a16 → sycommon_python_lib-0.2.5a17}/src/sycommon_python_lib.egg-info/SOURCES.txt

@@ -260,6 +260,7 @@ src/sycommon/synacos/param.py
 src/sycommon/tests/deep_agent_server.py
 src/sycommon/tests/test_consumer_recovery.py
 src/sycommon/tests/test_deep_agent.py
+src/sycommon/tests/test_dynamic_max_tokens.py
 src/sycommon/tests/test_email.py
 src/sycommon/tests/test_encoding_cache.py
 src/sycommon/tests/test_mcp_server.py