Rename control_server to api_server for general purpose

miles/ray/train/cell.py

@@ -15,7 +15,7 @@
     StatePending,
     StateStopped,
 )
-from miles.utils.ft_utils.control_server.models import CellStatus
+from miles.utils.ft_utils.api_server.models import CellStatus
 from miles.utils.ft_utils.health_checker import BaseHealthChecker
 from miles.utils.ft_utils.indep_dp import IndepDPInfo
 from miles.utils.tracking_utils.structured_log import log_structured

miles/ray/train/cell_monitor.py

@@ -8,7 +8,7 @@
     StatePending,
     StateStopped,
 )
-from miles.utils.ft_utils.control_server.models import CellCondition, CellStatus, TriState
+from miles.utils.ft_utils.api_server.models import CellCondition, CellStatus, TriState
 from miles.utils.ft_utils.health_checker import SimpleHealthChecker, SimpleHealthCheckerConfig
 
 if TYPE_CHECKING:

miles/utils/arguments.py

@@ -992,10 +992,10 @@ def add_fault_tolerance_arguments(parser):
                 help="Initial grace period (in seconds) before starting health checks. This allows time for model compilation and initialization. Increase this value significantly when using deepgemm.",
             )
             parser.add_argument(
-                "--control-server-port",
+                "--api-server-port",
                 type=int,
                 default=0,
-                help="Port for HTTP control server. 0 = disabled.",
+                help="Port for HTTP api server. 0 = disabled.",
             )
             parser.add_argument(
                 "--mini-ft-controller-enable",
@@ -2901,8 +2901,8 @@ def miles_validate_args(args):
     args.ft_components = _resolve_ft_components(args)
     args.eval_datasets = _resolve_eval_datasets(args)
 
-    if args.mini_ft_controller_enable and args.control_server_port == 0:
-        raise ValueError("--mini-ft-controller-enable requires --control-server-port to be set (non-zero)")
+    if args.mini_ft_controller_enable and args.api_server_port == 0:
+        raise ValueError("--mini-ft-controller-enable requires --api-server-port to be set (non-zero)")
 
     if "train" in args.ft_components:
         args.indep_dp = True

miles/utils/ft_utils/control_server/handles.py → miles/utils/ft_utils/api_server/handles.py

@@ -3,7 +3,7 @@
 import abc
 
 from miles.ray.train.group import RayTrainGroup
-from miles.utils.ft_utils.control_server.models import Cell, CellCondition, CellMetadata, CellSpec, CellStatus
+from miles.utils.ft_utils.api_server.models import Cell, CellCondition, CellMetadata, CellSpec, CellStatus
 from miles.utils.test_utils.fault_injector import FailureMode
 
 

miles/utils/ft_utils/control_server/registry.py → miles/utils/ft_utils/api_server/registry.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from miles.utils.ft_utils.control_server.handles import _CellHandle
+from miles.utils.ft_utils.api_server.handles import _CellHandle
 
 
 class _CellRegistry:

miles/utils/ft_utils/control_server/server.py → miles/utils/ft_utils/api_server/server.py

@@ -9,24 +9,17 @@
 from starlette.responses import JSONResponse
 
 from miles.ray.train.group import RayTrainGroup
-from miles.utils.ft_utils.control_server.handles import _ActorCellHandle, _CellHandle, _RolloutCellHandle
-from miles.utils.ft_utils.control_server.models import (
-    Cell,
-    CellList,
-    CellPatch,
-    FaultInjection,
-    K8sStatus,
-    _OkResponse,
-)
-from miles.utils.ft_utils.control_server.registry import _CellRegistry
+from miles.utils.ft_utils.api_server.handles import _ActorCellHandle, _CellHandle, _RolloutCellHandle
+from miles.utils.ft_utils.api_server.models import Cell, CellList, CellPatch, FaultInjection, K8sStatus, _OkResponse
+from miles.utils.ft_utils.api_server.registry import _CellRegistry
 
 logger = logging.getLogger(__name__)
 
 
 # -------------------------- entrypoint ------------------------------
 
 
-def start_control_server(
+def start_api_server(
     *,
     actor_model: RayTrainGroup,
     inference_controller: object,
@@ -49,24 +42,24 @@ def start_control_server(
                 )
             )
 
-    _start_control_server_raw(registry=registry, port=port)
+    _start_api_server_raw(registry=registry, port=port)
 
 
-def _start_control_server_raw(registry: _CellRegistry, port: int) -> None:
-    app = _create_control_app(registry)
+def _start_api_server_raw(registry: _CellRegistry, port: int) -> None:
+    app = _create_api_app(registry)
 
     def _run() -> None:
         uvicorn.run(app, host="0.0.0.0", port=port)
 
     thread = threading.Thread(target=_run, daemon=True)
     thread.start()
-    logger.info("Control server started on port %d", port)
+    logger.info("Api server started on port %d", port)
 
 
 # -------------------------- main app ------------------------------
 
 
-def _create_control_app(registry: _CellRegistry) -> FastAPI:
+def _create_api_app(registry: _CellRegistry) -> FastAPI:
     app = FastAPI()
 
     # -------------------------- exceptions ------------------------------

miles/utils/ft_utils/health_checker.py

@@ -8,7 +8,7 @@
 from typing import Any
 
 from miles.backends.sglang_utils.sglang_api_client import SGLangApiClient
-from miles.utils.ft_utils.control_server.models import TriState
+from miles.utils.ft_utils.api_server.models import TriState
 from miles.utils.pydantic_utils import StrictBaseModel
 from miles.utils.test_utils.clock import Clock, RealClock
 from miles.utils.tracking_utils.structured_log import log_structured

miles/utils/ft_utils/mini_ft_controller.py

@@ -11,7 +11,7 @@
 
 import httpx
 
-from miles.utils.ft_utils.control_server.models import Cell, CellList, CellPatch, CellPatchSpec, TriState
+from miles.utils.ft_utils.api_server.models import Cell, CellList, CellPatch, CellPatchSpec, TriState
 from miles.utils.pydantic_utils import StrictBaseModel
 from miles.utils.tracking_utils.structured_log import log_structured
 
@@ -26,7 +26,7 @@ def maybe_start_mini_ft_controller(args: Any) -> None:
         return
 
     runner = _MiniFTControllerRunner(
-        control_server_url=f"http://127.0.0.1:{args.control_server_port}",
+        api_server_url=f"http://127.0.0.1:{args.api_server_port}",
         poll_interval=args.mini_ft_controller_poll_interval,
         resume_delay=args.mini_ft_controller_resume_delay,
     )
@@ -46,11 +46,11 @@ class _MiniFTControllerRunner:
     def __init__(
         self,
         *,
-        control_server_url: str,
+        api_server_url: str,
         poll_interval: float,
         resume_delay: float,
     ) -> None:
-        url = control_server_url.rstrip("/")
+        url = api_server_url.rstrip("/")
         self._client = httpx.AsyncClient(base_url=url, timeout=30.0)
         self._controller = _MiniFTController(
             get_cells=self._get_cells,

tests/e2e/ft/README.md

@@ -239,13 +239,13 @@ Type: non-comparison (no baseline, no compare)
 Steps: 30 (default), configurable via --num-steps
 
 Architecture (external fault injection, not inside training loop):
-  1. Start training with indep_dp + control server (port 18080) + mini FT controller
+  1. Start training with indep_dp + api server (port 18080) + mini FT controller
   2. Start a background daemon thread that:
      a. Sleeps a random interval (exponential, mean = 60s / crash_probability ≈ 120s at the default)
      b. GET /api/v1/cells — read each cell's Healthy condition
      c. Count the genuinely-alive cells — reported Healthy, minus cells we injected that have
         not finished a down->up recovery (RecoveryGate) — and skip if injecting would leave
-        <=1 of them. The control server reports a just-killed cell Healthy for ~95s >> the
+        <=1 of them. The api server reports a just-killed cell Healthy for ~95s >> the
         inject interval, so excluding still-recovering cells is what keeps >=1 live replica
         (indep_dp cannot heal from zero survivors).
      d. Otherwise POST /api/v1/cells/{name}/inject-fault with a random failure mode
@@ -273,7 +273,7 @@ Type: non-comparison (no baseline run; reference = the baseline test's wandb cur
 Recipe: Qwen2.5-0.5B, GRPO, 250 rollouts; parallelism mirrors dp2_cp2_real_rollout
         (2 cells x CP2 on 4 train GPUs + 4 rollout engines x 1 GPU, disaggregated)
 Faults: same external random injection loop as scenario_ft_random
-        (train cells via control server)
+        (train cells via api server)
 
 Assertion: --ci-metric-checker-key eval/gsm8k with a threshold that must stay
   identical to the no-fault baseline's (0.55): fault recovery must not cost

tests/e2e/ft/conftest_ft/execution.py

@@ -155,7 +155,7 @@ def get_common_train_args(
 
 
 def get_ft_args(mode: FTTestMode) -> str:
-    return "--use-fault-tolerance " "--ft-components train " "--control-server-port 0 "
+    return "--use-fault-tolerance " "--ft-components train " "--api-server-port 0 "
 
 
 # Required for reproducibility (ref: https://github.com/THUDM/slime/pull/370)

tests/e2e/ft/conftest_ft/fault_injection.py

@@ -13,7 +13,7 @@
 
 logger = logging.getLogger(__name__)
 
-CONTROL_SERVER_PORT: int = 18080
+API_SERVER_PORT: int = 18080
 MEAN_INTERVAL_SECONDS: float = 60.0
 # Poll cell liveness this often so the gate tracks a crash->detect->heal cycle even when it
 # happens entirely between two (much sparser) injections; injections still fire on the long
@@ -27,7 +27,7 @@ def cell_is_alive(cell: dict) -> bool:
 
 
 class _CellState(enum.Enum):
-    INJECTED = enum.auto()  # we crashed it; the control server may still report it Healthy
+    INJECTED = enum.auto()  # we crashed it; the api server may still report it Healthy
     RECOVERING = enum.auto()  # observed unhealthy; awaiting its return to Healthy
 
 
@@ -76,7 +76,7 @@ def run_fault_injection_loop(
             resp.raise_for_status()
             cells = resp.json()["items"]
         except Exception:
-            logger.info("Failed to list cells from control server", exc_info=True)
+            logger.info("Failed to list cells from api server", exc_info=True)
             continue
 
         # Track recovery on every poll so a crash->detect->heal cycle that completes between two
@@ -139,7 +139,7 @@ def _on_successful_injection(self) -> None:
 
 
 def spawn_fault_injector(*, seed: int, mean_interval_seconds: float) -> FaultInjectorHandle:
-    base_url = f"http://localhost:{CONTROL_SERVER_PORT}"
+    base_url = f"http://localhost:{API_SERVER_PORT}"
     handle = FaultInjectorHandle(base_url=base_url, seed=seed, mean_interval_seconds=mean_interval_seconds)
     handle.start()
     return handle

tests/e2e/ft/conftest_ft/scenario_ft_random.py

@@ -14,7 +14,7 @@
     prepare,
     run_training,
 )
-from tests.e2e.ft.conftest_ft.fault_injection import CONTROL_SERVER_PORT, MEAN_INTERVAL_SECONDS, spawn_fault_injector
+from tests.e2e.ft.conftest_ft.fault_injection import API_SERVER_PORT, MEAN_INTERVAL_SECONDS, spawn_fault_injector
 from tests.e2e.ft.conftest_ft.modes import FTTestMode, resolve_mode
 
 from miles.utils.test_utils.reconfigure_assertions import assert_soak_reconfigure_events
@@ -34,7 +34,7 @@ def run_ci(
     """Random failure soak test.
 
     Starts a background thread that injects faults at random intervals via the
-    control server HTTP API. The mini FT controller auto-recovers; the test passes
+    api server HTTP API. The mini FT controller auto-recovers; the test passes
     if training completes without hanging.
 
     Doubles as the per-mode CI entry point: a CI file calls ``run_ci(mode)`` (defaults);
@@ -54,7 +54,7 @@ def run_ci(
             ft_mode, dump_dir=dump_dir, num_steps=num_steps, debug_rollout_data_dir=debug_rollout_data_dir
         )
         + get_ft_args(ft_mode)
-        + f"--control-server-port {CONTROL_SERVER_PORT} "
+        + f"--api-server-port {API_SERVER_PORT} "
         + "--mini-ft-controller-enable "
     )
 

tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py

@@ -8,7 +8,7 @@
 import typer
 
 from tests.e2e.ft.conftest_ft.app import resolve_dump_dir
-from tests.e2e.ft.conftest_ft.fault_injection import CONTROL_SERVER_PORT, MEAN_INTERVAL_SECONDS, spawn_fault_injector
+from tests.e2e.ft.conftest_ft.fault_injection import API_SERVER_PORT, MEAN_INTERVAL_SECONDS, spawn_fault_injector
 
 import miles.utils.external_utils.command_utils as U
 
@@ -146,7 +146,7 @@ def _get_gsm8k_train_args(*, seed: int, num_rollout: int, metric_threshold: floa
     fault_tolerance_args = (
         "--use-fault-tolerance "
         "--ft-components train "
-        f"--control-server-port {CONTROL_SERVER_PORT} "
+        f"--api-server-port {API_SERVER_PORT} "
         "--mini-ft-controller-enable "
     )
 

tests/fast/e2e/ft/test_fault_injection.py

@@ -38,7 +38,7 @@ def test_fresh_gate_counts_every_healthy_cell_as_alive() -> None:
 
 
 def test_injected_cell_is_excluded_while_its_crash_is_still_undetected() -> None:
-    """The control server's stale 'still healthy' view must not count a just-killed cell."""
+    """The api server's stale 'still healthy' view must not count a just-killed cell."""
     gate = RecoveryGate()
     cells = [_cell("c0", healthy=True), _cell("c1", healthy=True)]
     gate.note_injected("c0")

tests/fast/ray/train/test_cell_monitor.py

@@ -11,7 +11,7 @@
     StatePending,
     StateStopped,
 )
-from miles.utils.ft_utils.control_server.models import TriState
+from miles.utils.ft_utils.api_server.models import TriState
 from miles.utils.ft_utils.health_checker import SimpleHealthCheckerConfig
 from miles.utils.ft_utils.indep_dp import IndepDPInfo
 

tests/fast/utils/control_server/conftest.py → tests/fast/utils/api_server/conftest.py

@@ -5,9 +5,9 @@
 import httpx
 import pytest
 
-from miles.utils.ft_utils.control_server.models import Cell, CellCondition, CellMetadata, CellSpec, CellStatus
-from miles.utils.ft_utils.control_server.registry import _CellRegistry
-from miles.utils.ft_utils.control_server.server import _create_control_app
+from miles.utils.ft_utils.api_server.models import Cell, CellCondition, CellMetadata, CellSpec, CellStatus
+from miles.utils.ft_utils.api_server.registry import _CellRegistry
+from miles.utils.ft_utils.api_server.server import _create_api_app
 
 
 class MockHandle:
@@ -146,7 +146,7 @@ def is_stopped(self) -> bool:
         return self._is_stopped
 
     def cell_status(self) -> CellStatus:
-        from miles.utils.ft_utils.control_server.models import CellCondition, CellStatus
+        from miles.utils.ft_utils.api_server.models import CellCondition, CellStatus
 
         return CellStatus(
             phase=self._phase,
@@ -171,6 +171,6 @@ def registry() -> _CellRegistry:
 
 @pytest.fixture
 def async_client(registry: _CellRegistry) -> httpx.AsyncClient:
-    app = _create_control_app(registry)
+    app = _create_api_app(registry)
     transport = httpx.ASGITransport(app=app)
     return httpx.AsyncClient(transport=transport, base_url="http://test")

tests/fast/utils/control_server/test_handles.py → tests/fast/utils/api_server/test_handles.py

@@ -5,7 +5,7 @@
 import pytest
 
 from miles.ray.train.group import RayTrainGroup
-from miles.utils.ft_utils.control_server.handles import _ActorCellHandle, _CellHandle, _RolloutCellHandle
+from miles.utils.ft_utils.api_server.handles import _ActorCellHandle, _CellHandle, _RolloutCellHandle
 from miles.utils.test_utils.fault_injector import FailureMode
 
 from .conftest import MockInferenceController, MockRayTrainCell, make_mock_group

tests/fast/utils/control_server/test_registry.py → tests/fast/utils/api_server/test_registry.py

@@ -2,7 +2,7 @@
 
 import pytest
 
-from miles.utils.ft_utils.control_server.registry import _CellRegistry
+from miles.utils.ft_utils.api_server.registry import _CellRegistry
 
 from .conftest import MockHandle
 

tests/fast/utils/control_server/test_server.py → tests/fast/utils/api_server/test_server.py

@@ -3,7 +3,7 @@
 import httpx
 import pytest
 
-from miles.utils.ft_utils.control_server.registry import _CellRegistry
+from miles.utils.ft_utils.api_server.registry import _CellRegistry
 
 from .conftest import MockHandle
 

tests/fast/utils/test_health_checker.py

@@ -1,6 +1,6 @@
 import asyncio
 
-from miles.utils.ft_utils.control_server.models import TriState
+from miles.utils.ft_utils.api_server.models import TriState
 from miles.utils.ft_utils.health_checker import NoopHealthChecker, SimpleHealthChecker
 from miles.utils.test_utils.clock import FakeClock