isagellm-core 0.2.2.8__tar.gz → 0.3.0.2__tar.gz

{isagellm_core-0.2.2.8/src/isagellm_core.egg-info → isagellm_core-0.3.0.2}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: isagellm-core
-Version: 0.2.2.8
-Summary: sageLLM core runtime (config/observability/registries), plugin-ready
+Version: 0.3.0.2
+Summary: sageLLM core runtime with PD separation (MVP)
 Author: IntelliStream Team
 License: Proprietary - IntelliStream
 Classifier: Development Status :: 3 - Alpha
@@ -13,8 +13,10 @@ Requires-Python: ==3.11.*
 Description-Content-Type: text/markdown
 Requires-Dist: pydantic>=2.0.0
 Requires-Dist: pyyaml>=6.0.0
-Requires-Dist: isagellm-protocol<0.2.0,>=0.1.0
-Requires-Dist: isagellm-backend<0.3.0,>=0.2.0.0
+Requires-Dist: isagellm-protocol<0.4.0,>=0.3.0.0
+Requires-Dist: isagellm-backend<0.4.0,>=0.3.0.0
+Requires-Dist: isagellm-comm<0.4.0,>=0.3.0.0
+Requires-Dist: isagellm-kv-cache<0.4.0,>=0.3.0.0
 Provides-Extra: dev
 Requires-Dist: pytest>=7.0.0; extra == "dev"
 Requires-Dist: pytest-cov>=4.0.0; extra == "dev"
@@ -23,6 +25,7 @@ Requires-Dist: ruff>=0.8.0; extra == "dev"
 Requires-Dist: mypy>=1.0.0; extra == "dev"
 Requires-Dist: types-PyYAML>=6.0.0; extra == "dev"
 Requires-Dist: pre-commit>=3.0.0; extra == "dev"
+Requires-Dist: isage-pypi-publisher>=0.2.0; extra == "dev"
 
 # sagellm-core
 

{isagellm_core-0.2.2.8 → isagellm_core-0.3.0.2}/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "isagellm-core"
-version = "0.2.2.8"
-description = "sageLLM core runtime (config/observability/registries), plugin-ready"
+version = "0.3.0.2"
+description = "sageLLM core runtime with PD separation (MVP)"
 readme = "README.md"
 requires-python = "==3.11.*"
 authors = [{ name = "IntelliStream Team" }]
@@ -20,8 +20,10 @@ classifiers = [
 dependencies = [
   "pydantic>=2.0.0",
   "pyyaml>=6.0.0",
-  "isagellm-protocol>=0.1.0,<0.2.0",
-  "isagellm-backend>=0.2.0.0,<0.3.0",
+  "isagellm-protocol>=0.3.0.0,<0.4.0",
+  "isagellm-backend>=0.3.0.0,<0.4.0",
+  "isagellm-comm>=0.3.0.0,<0.4.0",
+  "isagellm-kv-cache>=0.3.0.0,<0.4.0",
 ]
 
 [project.optional-dependencies]
@@ -33,6 +35,7 @@ dev = [
   "mypy>=1.0.0",
   "types-PyYAML>=6.0.0",
   "pre-commit>=3.0.0",
+  "isage-pypi-publisher>=0.2.0",
 ]
 
 [tool.setuptools]
@@ -51,7 +54,9 @@ line-length = 100
 
 [project.entry-points."sagellm.engines"]
 cpu = "sagellm_core.engines.cpu:create_cpu_engine"
+hf-cuda = "sagellm_core.engines.hf_cuda:create_hf_cuda_engine"
 pytorch = "sagellm_core.engines.pytorch:create_pytorch_engine"
+ascend = "sagellm_core.engines.ascend:create_ascend_engine"
 
 [tool.mypy]
 python_version = "3.10"

{isagellm_core-0.2.2.8 → isagellm_core-0.3.0.2/src/isagellm_core.egg-info}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: isagellm-core
-Version: 0.2.2.8
-Summary: sageLLM core runtime (config/observability/registries), plugin-ready
+Version: 0.3.0.2
+Summary: sageLLM core runtime with PD separation (MVP)
 Author: IntelliStream Team
 License: Proprietary - IntelliStream
 Classifier: Development Status :: 3 - Alpha
@@ -13,8 +13,10 @@ Requires-Python: ==3.11.*
 Description-Content-Type: text/markdown
 Requires-Dist: pydantic>=2.0.0
 Requires-Dist: pyyaml>=6.0.0
-Requires-Dist: isagellm-protocol<0.2.0,>=0.1.0
-Requires-Dist: isagellm-backend<0.3.0,>=0.2.0.0
+Requires-Dist: isagellm-protocol<0.4.0,>=0.3.0.0
+Requires-Dist: isagellm-backend<0.4.0,>=0.3.0.0
+Requires-Dist: isagellm-comm<0.4.0,>=0.3.0.0
+Requires-Dist: isagellm-kv-cache<0.4.0,>=0.3.0.0
 Provides-Extra: dev
 Requires-Dist: pytest>=7.0.0; extra == "dev"
 Requires-Dist: pytest-cov>=4.0.0; extra == "dev"
@@ -23,6 +25,7 @@ Requires-Dist: ruff>=0.8.0; extra == "dev"
 Requires-Dist: mypy>=1.0.0; extra == "dev"
 Requires-Dist: types-PyYAML>=6.0.0; extra == "dev"
 Requires-Dist: pre-commit>=3.0.0; extra == "dev"
+Requires-Dist: isage-pypi-publisher>=0.2.0; extra == "dev"
 
 # sagellm-core
 

{isagellm_core-0.2.2.8 → isagellm_core-0.3.0.2}/src/isagellm_core.egg-info/SOURCES.txt

@@ -17,31 +17,40 @@ src/sagellm_core/engine.pyc
 src/sagellm_core/engine_factory.pyc
 src/sagellm_core/factory.pyc
 src/sagellm_core/health.pyc
+src/sagellm_core/pd_executor.pyc
 src/sagellm_core/plugins.pyc
 src/sagellm_core/py.typed
 src/sagellm_core/runner.pyc
+src/sagellm_core/runtime.pyc
 src/sagellm_core/workload.pyc
 src/sagellm_core/__pycache__/__init__.cpython-311.pyc
+src/sagellm_core/__pycache__/base_engine.cpython-311.pyc
 src/sagellm_core/__pycache__/config.cpython-311.pyc
 src/sagellm_core/__pycache__/demo.cpython-311.pyc
 src/sagellm_core/__pycache__/engine.cpython-311.pyc
 src/sagellm_core/__pycache__/engine_factory.cpython-311.pyc
 src/sagellm_core/__pycache__/factory.cpython-311.pyc
 src/sagellm_core/__pycache__/health.cpython-311.pyc
+src/sagellm_core/__pycache__/mock_engine.cpython-311.pyc
+src/sagellm_core/__pycache__/pd_executor.cpython-311.pyc
 src/sagellm_core/__pycache__/plugins.cpython-311.pyc
 src/sagellm_core/__pycache__/runner.cpython-311.pyc
+src/sagellm_core/__pycache__/runtime.cpython-311.pyc
 src/sagellm_core/__pycache__/workload.cpython-311.pyc
 src/sagellm_core/engines/__init__.py
 src/sagellm_core/engines/__init__.pyc
+src/sagellm_core/engines/ascend.pyc
 src/sagellm_core/engines/cpu.pyc
 src/sagellm_core/engines/embedding.pyc
 src/sagellm_core/engines/hf_cuda.pyc
 src/sagellm_core/engines/pytorch.pyc
 src/sagellm_core/engines/pytorch_engine.pyc
 src/sagellm_core/engines/__pycache__/__init__.cpython-311.pyc
+src/sagellm_core/engines/__pycache__/ascend.cpython-311.pyc
 src/sagellm_core/engines/__pycache__/cpu.cpython-311.pyc
 src/sagellm_core/engines/__pycache__/embedding.cpython-311.pyc
 src/sagellm_core/engines/__pycache__/hf_cuda.cpython-311.pyc
+src/sagellm_core/engines/__pycache__/mock.cpython-311.pyc
 src/sagellm_core/engines/__pycache__/pytorch.cpython-311.pyc
 src/sagellm_core/engines/__pycache__/pytorch_engine.cpython-311.pyc
 tests/test_ci_smoke.py
@@ -53,5 +62,7 @@ tests/test_engine.py
 tests/test_engine_behavior_parity.py
 tests/test_engine_contract.py
 tests/test_engine_contract_simplified.py
+tests/test_pd_separation.py
 tests/test_pytorch_engine.py
+tests/test_streaming_pd.py
 tests/test_task0_10_workload.py

{isagellm_core-0.2.2.8 → isagellm_core-0.3.0.2}/src/isagellm_core.egg-info/entry_points.txt

@@ -1,3 +1,5 @@
 [sagellm.engines]
+ascend = sagellm_core.engines.ascend:create_ascend_engine
 cpu = sagellm_core.engines.cpu:create_cpu_engine
+hf-cuda = sagellm_core.engines.hf_cuda:create_hf_cuda_engine
 pytorch = sagellm_core.engines.pytorch:create_pytorch_engine

isagellm_core-0.3.0.2/src/isagellm_core.egg-info/requires.txt

@@ -0,0 +1,16 @@
+pydantic>=2.0.0
+pyyaml>=6.0.0
+isagellm-protocol<0.4.0,>=0.3.0.0
+isagellm-backend<0.4.0,>=0.3.0.0
+isagellm-comm<0.4.0,>=0.3.0.0
+isagellm-kv-cache<0.4.0,>=0.3.0.0
+
+[dev]
+pytest>=7.0.0
+pytest-cov>=4.0.0
+pytest-timeout>=2.0.0
+ruff>=0.8.0
+mypy>=1.0.0
+types-PyYAML>=6.0.0
+pre-commit>=3.0.0
+isage-pypi-publisher>=0.2.0

{isagellm_core-0.2.2.8 → isagellm_core-0.3.0.2}/src/sagellm_core/__init__.py

@@ -6,6 +6,8 @@
 - Engine 工厂函数
 - 插件系统
 - Demo Runner
+- 分布式 Runtime（PD 分离 MVP）
+- PD 分离执行器
 """
 
 from __future__ import annotations
@@ -23,17 +25,24 @@ from sagellm_core.config import (
 from sagellm_core.demo import main as demo_main
 from sagellm_core.engine_factory import EngineFactory
 from sagellm_core.engines import (
+    AscendEngine,
+    AscendEngineConfig,
     CPUEngine,
     EmbeddingEngine,
     EmbeddingEngineConfig,
     HFCudaEngine,
     HFCudaEngineConfig,
+    create_ascend_engine,
 )
 from sagellm_core.factory import create_backend, create_engine
 from sagellm_core.health import HealthStatus
 from sagellm_core.plugins import PluginResolutionError, list_entry_points, resolve_kind
 from sagellm_core.runner import DemoRunner, RunnerContext
 
+# PD 分离 MVP 模块
+from sagellm_core.runtime import DistributedConfig, DistributedRuntime, RuntimeState
+from sagellm_core.pd_executor import PDExecutionContext, PDSeparatedExecutor
+
 # PyTorch engine (optional, loaded lazily)
 PyTorchEngine = None
 PyTorchEngineConfig = None
@@ -61,6 +70,11 @@ try:
 except Exception:
     pass  # torch or CUDA not available
 
+try:
+    EngineFactory.register(AscendEngine)
+except Exception:
+    pass  # torch_npu or Ascend deps not available
+
 try:
     EngineFactory.register(EmbeddingEngine)
 except Exception:
@@ -72,7 +86,7 @@ try:
 except Exception:
     pass  # torch or transformers not available
 
-__version__ = "0.2.2.8"
+__version__ = "0.3.0.0"
 
 __all__ = [
     # Version
@@ -90,6 +104,9 @@ __all__ = [
     "EngineInstanceConfig",  # For runtime engine instantiation
     "HealthStatus",
     # Engine implementations
+    "AscendEngine",
+    "AscendEngineConfig",
+    "create_ascend_engine",
     "CPUEngine",
     "HFCudaEngine",
     "HFCudaEngineConfig",
@@ -111,4 +128,10 @@ __all__ = [
     "demo_main",
     "DemoRunner",
     "RunnerContext",
+    # PD Separation MVP
+    "DistributedConfig",
+    "DistributedRuntime",
+    "RuntimeState",
+    "PDExecutionContext",
+    "PDSeparatedExecutor",
 ]

{isagellm_core-0.2.2.8 → isagellm_core-0.3.0.2}/src/sagellm_core/engines/__init__.py

@@ -3,27 +3,39 @@
 This module provides built-in engine implementations:
 - CPUEngine: CPU-only inference with HuggingFace Transformers
 - HFCudaEngine: CUDA inference with HuggingFace Transformers
+- AscendEngine: Huawei Ascend NPU (PyTorch-NPU)
 - PyTorchEngine: Unified PyTorch engine (auto-detects CUDA/NPU/CPU)
 - EmbeddingEngine: Embedding model inference
 """
 
 from __future__ import annotations
 
-from sagellm_core.engines.cpu import CPUEngine
+from sagellm_core.engines.ascend import AscendEngine, AscendEngineConfig, create_ascend_engine
+from sagellm_core.engines.cpu import CPUEngine, create_cpu_engine
 from sagellm_core.engines.embedding import EmbeddingEngine, EmbeddingEngineConfig
-from sagellm_core.engines.hf_cuda import HFCudaEngine, HFCudaEngineInstanceConfig
+from sagellm_core.engines.hf_cuda import (
+    HFCudaEngine,
+    HFCudaEngineInstanceConfig,
+    create_hf_cuda_engine,
+)
 from sagellm_core.engines.pytorch_engine import PyTorchEngine, create_pytorch_engine
 
 # Export aliases for backward compatibility
 HFCudaEngineConfig = HFCudaEngineInstanceConfig
 
 __all__ = [
+    # Ascend engine
+    "AscendEngine",
+    "AscendEngineConfig",
+    "create_ascend_engine",
     # CPU engine
     "CPUEngine",
+    "create_cpu_engine",
     # CUDA engine
     "HFCudaEngine",
     "HFCudaEngineConfig",
     "HFCudaEngineInstanceConfig",
+    "create_hf_cuda_engine",
     # PyTorch engine (unified)
     "PyTorchEngine",
     "create_pytorch_engine",

{isagellm_core-0.2.2.8 → isagellm_core-0.3.0.2}/tests/test_engine.py

@@ -43,6 +43,28 @@ class TestCPUEngine(BaseEngine):
     async def health_check(self) -> bool:
         return self._is_running
 
+    async def prefill(self, request: Request) -> dict:
+        """Minimal prefill implementation for testing."""
+        if not self._is_running:
+            raise RuntimeError("not running")
+        return {
+            "kv_handle": {"test": "handle"},
+            "num_tokens": len(request.prompt.split()) if request.prompt else 0,
+            "first_token_id": 1,
+        }
+
+    async def decode(self, request: Request, kv_handle=None, max_new_tokens=None) -> dict:
+        """Minimal decode implementation for testing."""
+        if not self._is_running:
+            raise RuntimeError("not running")
+        num_tokens = max_new_tokens or request.max_tokens
+        return {
+            "output_tokens": [1, 2, 3][:num_tokens],
+            "output_text": "test output",
+            "finish_reason": "stop",
+            "num_tokens": min(3, num_tokens),
+        }
+
     async def execute(self, request: Request) -> Response:
         if not self._is_running:
             raise RuntimeError("not running")

isagellm_core-0.3.0.2/tests/test_pd_separation.py

@@ -0,0 +1,207 @@
+"""测试 PD 分离功能
+
+验证 CPUEngine 的 prefill() 和 decode() 方法是否正常工作。
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+
+import pytest
+
+from sagellm_protocol.types import Request
+from sagellm_core import PDSeparatedExecutor, DistributedRuntime, create_backend, create_engine
+from sagellm_core.config import EngineConfig
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+@pytest.mark.asyncio
+async def test_cpu_engine_prefill():
+    """测试 CPUEngine 的 prefill 方法"""
+    # 创建引擎 - 使用轻量级模型（约 50MB，快速下载）
+    backend = create_backend({"kind": "cpu"})
+    config = EngineConfig(
+        kind="cpu",
+        model="sshleifer/tiny-gpt2",  # 轻量级测试模型
+        device="cpu",
+    )
+    engine = create_engine(config, backend)
+
+    try:
+        await engine.start()
+
+        # 创建请求
+        request = Request(
+            request_id="req-prefill-001",
+            trace_id="trace-001",
+            model="sshleifer/tiny-gpt2",
+            prompt="Hello, how are you?",
+            max_tokens=10,
+            stream=False,
+        )
+
+        # 执行 Prefill
+        result = await engine.prefill(request)
+
+        # 验证结果
+        assert "kv_handle" in result
+        assert "num_tokens" in result
+        assert result["num_tokens"] > 0
+        assert "first_token_id" in result
+
+        logger.info(f"✓ Prefill completed: {result['num_tokens']} tokens processed")
+
+    finally:
+        await engine.stop()
+
+
+@pytest.mark.asyncio
+async def test_cpu_engine_decode():
+    """测试 CPUEngine 的 decode 方法"""
+    backend = create_backend({"kind": "cpu"})
+    config = EngineConfig(
+        kind="cpu",
+        model="sshleifer/tiny-gpt2",
+        device="cpu",
+    )
+    engine = create_engine(config, backend)
+
+    try:
+        await engine.start()
+
+        # 先执行 Prefill
+        request = Request(
+            request_id="req-decode-001",
+            trace_id="trace-001",
+            model="sshleifer/tiny-gpt2",
+            prompt="Hello!",
+            max_tokens=10,
+            stream=False,
+        )
+
+        prefill_result = await engine.prefill(request)
+        kv_handle = prefill_result["kv_handle"]
+
+        # 执行 Decode
+        decode_result = await engine.decode(request, kv_handle=kv_handle, max_new_tokens=10)
+
+        # 验证结果
+        assert "output_tokens" in decode_result
+        assert "output_text" in decode_result
+        assert "finish_reason" in decode_result
+        assert len(decode_result["output_tokens"]) > 0
+
+        logger.info(
+            f"✓ Decode completed: {len(decode_result['output_tokens'])} tokens generated, "
+            f"text='{decode_result['output_text']}'"
+        )
+
+    finally:
+        await engine.stop()
+
+
+@pytest.mark.asyncio
+async def test_pd_executor_hybrid():
+    """测试 PDSeparatedExecutor 的 Hybrid 模式"""
+    # 初始化 Runtime
+    runtime = DistributedRuntime()
+    await runtime.initialize()
+
+    # 创建引擎
+    backend = create_backend({"kind": "cpu"})
+    config = EngineConfig(
+        kind="cpu",
+        model="sshleifer/tiny-gpt2",
+        device="cpu",
+    )
+    engine = create_engine(config, backend)
+
+    try:
+        await engine.start()
+
+        # 创建 PD Executor
+        executor = PDSeparatedExecutor(engine=engine, runtime=runtime)
+
+        # 创建 Hybrid 请求
+        request = Request(
+            request_id="req-hybrid-001",
+            trace_id="trace-001",
+            model="sshleifer/tiny-gpt2",
+            prompt="Hello!",
+            max_tokens=10,
+            stream=False,
+            phase="decode",  # hybrid mode
+        )
+
+        # 执行
+        response = await executor.execute(request)
+
+        # 验证结果
+        assert response.request_id == "req-hybrid-001"
+        assert response.metrics.prefill_ms > 0
+        assert response.metrics.decode_ms > 0
+        assert len(response.output_text) > 0
+
+        logger.info("✓ Hybrid execution completed:")
+        logger.info(f"  - Prefill: {response.metrics.prefill_ms:.2f} ms")
+        logger.info(f"  - Decode: {response.metrics.decode_ms:.2f} ms")
+        logger.info(f"  - Output: '{response.output_text}'")
+
+    finally:
+        await engine.stop()
+        await runtime.shutdown()
+
+
+@pytest.mark.asyncio
+async def test_pd_executor_prefill_only():
+    """测试 PDSeparatedExecutor 的 Prefill-Only 模式"""
+    runtime = DistributedRuntime()
+    await runtime.initialize()
+
+    backend = create_backend({"kind": "cpu"})
+    config = EngineConfig(
+        kind="cpu",
+        model="sshleifer/tiny-gpt2",
+        device="cpu",
+    )
+    engine = create_engine(config, backend)
+
+    try:
+        await engine.start()
+        executor = PDSeparatedExecutor(engine=engine, runtime=runtime)
+
+        # Prefill-Only 请求
+        request = Request(
+            request_id="req-prefill-only-001",
+            trace_id="trace-001",
+            model="sshleifer/tiny-gpt2",
+            prompt="Hello!",
+            max_tokens=1,  # 最小值（但实际只做 prefill）
+            stream=False,
+            phase="prefill",
+        )
+
+        response = await executor.execute(request)
+
+        # 验证：只有 Prefill，没有 Decode（或 Decode 很少）
+        assert response.metrics.prefill_ms > 0
+        # decode_ms 可能为 0（prefill-only）或很小（生成了 1 token）
+
+        logger.info("✓ Prefill-Only execution completed:")
+        logger.info(f"  - Prefill: {response.metrics.prefill_ms:.2f} ms")
+
+    finally:
+        await engine.stop()
+        await runtime.shutdown()
+
+
+if __name__ == "__main__":
+    # 运行测试
+    asyncio.run(test_cpu_engine_prefill())
+    asyncio.run(test_cpu_engine_decode())
+    asyncio.run(test_pd_executor_hybrid())
+    asyncio.run(test_pd_executor_prefill_only())
+    print("\n✅ All PD separation tests passed!")

isagellm_core-0.3.0.2/tests/test_streaming_pd.py

@@ -0,0 +1,186 @@
+"""Tests for Streaming PD Separation.
+
+测试流式 Prefill-Decode 分离功能。
+"""
+
+from __future__ import annotations
+
+import pytest
+from sagellm_protocol.types import Request
+
+from sagellm_core.engines.cpu import CPUEngine, CPUEngineConfig
+from sagellm_core.pd_executor import PDSeparatedExecutor
+from sagellm_core.runtime import DistributedRuntime
+
+
+@pytest.mark.asyncio
+async def test_streaming_pd_separation():
+    """测试流式 PD 分离：验证事件顺序和 TTFT 在首 token"""
+    config = CPUEngineConfig(
+        engine_id="cpu-engine-stream-001",
+        model_path="sshleifer/tiny-gpt2",
+        torch_dtype="float32",
+        trust_remote_code=True,
+    )
+    engine = CPUEngine(config)
+    runtime = DistributedRuntime()
+
+    try:
+        await engine.start()
+        await runtime.initialize()
+
+        executor = PDSeparatedExecutor(engine=engine, runtime=runtime)
+
+        request = Request(
+            request_id="test-stream-001",
+            trace_id="trace-stream-001",
+            model="sshleifer/tiny-gpt2",
+            prompt="Once upon a time",
+            max_tokens=5,
+            stream=True,
+        )
+
+        events = []
+        async for event in executor.stream(request):
+            events.append(event)
+
+        # Verify Event Sequence
+        # ═════════════════════════════════════════════════════════════════
+        assert len(events) >= 3, "Should have at least start, delta, end"
+
+        # 1. Start Event
+        assert events[0].event == "start"
+        assert events[0].request_id == "test-stream-001"
+        assert events[0].trace_id == "trace-stream-001"
+
+        # 2. Delta Events (至少一个)
+        delta_events = [e for e in events if e.event == "delta"]
+        assert len(delta_events) > 0, "Should have at least one delta event"
+
+        # 每个 delta 应包含 chunk 和 chunk_tokens
+        for delta in delta_events:
+            assert hasattr(delta, "chunk"), "Delta should have chunk"
+            assert hasattr(delta, "chunk_tokens"), "Delta should have chunk_tokens"
+
+        # 3. End Event
+        assert events[-1].event == "end"
+        assert events[-1].finish_reason in ["stop", "length", "error"]
+        assert events[-1].metrics is not None, "End event should contain final metrics"
+
+        # Verify PD Metrics
+        # ═════════════════════════════════════════════════════════════════
+        final_metrics = events[-1].metrics
+        assert final_metrics.ttft_ms > 0, "TTFT should be positive"
+        assert final_metrics.prefill_ms > 0, "Prefill time should be positive"
+        assert final_metrics.decode_ms > 0, "Decode time should be positive"
+
+        # TBT 应该是 decode 平均时间
+        assert final_metrics.tbt_ms > 0, "TBT should be positive"
+
+    finally:
+        await engine.stop()
+        await runtime.shutdown()
+
+
+@pytest.mark.asyncio
+async def test_streaming_pd_event_content():
+    """测试流式事件内容：delta text 和 tokens"""
+    config = CPUEngineConfig(
+        engine_id="cpu-engine-stream-002",
+        model_path="sshleifer/tiny-gpt2",
+        torch_dtype="float32",
+        trust_remote_code=True,
+    )
+    engine = CPUEngine(config)
+    runtime = DistributedRuntime()
+
+    try:
+        await engine.start()
+        await runtime.initialize()
+
+        executor = PDSeparatedExecutor(engine=engine, runtime=runtime)
+
+        request = Request(
+            request_id="test-stream-002",
+            trace_id="trace-stream-002",
+            model="sshleifer/tiny-gpt2",
+            prompt="Hello world",
+            max_tokens=3,
+            stream=True,
+        )
+
+        events = []
+        async for event in executor.stream(request):
+            events.append(event)
+
+        # ═════════════════════════════════════════════════════════════════
+        # Verify Delta Content
+        # ═════════════════════════════════════════════════════════════════
+        delta_events = [e for e in events if e.event == "delta"]
+
+        for delta in delta_events:
+            # 每个 delta 应包含 chunk 和 chunk_tokens
+            assert hasattr(delta, "chunk"), "Delta should have chunk"
+            assert hasattr(delta, "chunk_tokens"), "Delta should have chunk_tokens"
+
+        # ═════════════════════════════════════════════════════════════════
+        # Verify End Content
+        # ═════════════════════════════════════════════════════════════════
+        end_event = events[-1]
+        assert end_event.output_text, "End event should contain full output text"
+        assert end_event.output_tokens, "End event should contain output tokens"
+
+    finally:
+        await engine.stop()
+        await runtime.shutdown()
+
+
+@pytest.mark.asyncio
+async def test_streaming_pd_metrics_consistency():
+    """测试流式 PD 指标一致性：首 delta 和 end 的 TTFT 应相同"""
+    config = CPUEngineConfig(
+        engine_id="cpu-engine-stream-003",
+        model_path="sshleifer/tiny-gpt2",
+        torch_dtype="float32",
+        trust_remote_code=True,
+    )
+    engine = CPUEngine(config)
+    runtime = DistributedRuntime()
+
+    try:
+        await engine.start()
+        await runtime.initialize()
+
+        executor = PDSeparatedExecutor(engine=engine, runtime=runtime)
+
+        request = Request(
+            request_id="test-stream-003",
+            trace_id="trace-stream-003",
+            model="sshleifer/tiny-gpt2",
+            prompt="Testing metrics",
+            max_tokens=4,
+            stream=True,
+        )
+
+        events = []
+        async for event in executor.stream(request):
+            events.append(event)
+
+        # ═════════════════════════════════════════════════════════════════
+        # Extract Metrics
+        # ═════════════════════════════════════════════════════════════════
+        end_metrics = events[-1].metrics
+
+        # Verify TTFT Consistency
+        # ═════════════════════════════════════════════════════════════════
+        assert end_metrics is not None
+
+        # TTFT 应相同（首 token 延迟不变）
+        assert end_metrics.ttft_ms > 0, "TTFT should be positive"
+
+        # Prefill 时间应相同
+        assert end_metrics.prefill_ms > 0, "Prefill time should be consistent"
+
+    finally:
+        await engine.stop()
+        await runtime.shutdown()

isagellm_core-0.2.2.8/src/isagellm_core.egg-info/requires.txt

@@ -1,13 +0,0 @@
-pydantic>=2.0.0
-pyyaml>=6.0.0
-isagellm-protocol<0.2.0,>=0.1.0
-isagellm-backend<0.3.0,>=0.2.0.0
-
-[dev]
-pytest>=7.0.0
-pytest-cov>=4.0.0
-pytest-timeout>=2.0.0
-ruff>=0.8.0
-mypy>=1.0.0
-types-PyYAML>=6.0.0
-pre-commit>=3.0.0