From d633a9be87c5a0181618c3918cd1e4a7ac6e8d31 Mon Sep 17 00:00:00 2001
From: Jiajun Li <guapisolo@gmail.com>
Date: Mon, 13 Jul 2026 14:52:09 -0700
Subject: [PATCH 1/4] ci: enable ft-long fault-tolerance soak tests

The run-ci-ft-long label already reaches the full-node H200 suite, but all three matching registrations are disabled. Remove those temporary gates so the label schedules every FT soak entry.
---
 tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py | 4 +---
 tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py      | 4 +---
 tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py             | 1 -
 3 files changed, 2 insertions(+), 7 deletions(-)

diff --git a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
index 0a307d724c..21f42fc543 100644
--- a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
+++ b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
@@ -6,9 +6,7 @@
 from tests.ci.ci_register import register_cuda_ci
 from tests.e2e.ft.conftest_ft.scenario_ft_random import run_ci
 
-register_cuda_ci(
-    est_time=2400, suite="stage-c-8-gpu-h200", labels=["ft-long"], disabled="FT soak tests pending CI infra support"
-)
+register_cuda_ci(est_time=2400, suite="stage-c-8-gpu-h200", labels=["ft-long"])
 
 _MODE: str = "dp2_cp2_real_rollout"
 
diff --git a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
index 2a9d0e7827..e1a46fc8bf 100644
--- a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
+++ b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
@@ -6,9 +6,7 @@
 from tests.ci.ci_register import register_cuda_ci
 from tests.e2e.ft.conftest_ft.scenario_ft_random import run_ci
 
-register_cuda_ci(
-    est_time=1800, suite="stage-c-8-gpu-h200", labels=["ft-long"], disabled="FT soak tests pending CI infra support"
-)
+register_cuda_ci(est_time=1800, suite="stage-c-8-gpu-h200", labels=["ft-long"])
 
 _MODE: str = "dp2_cp2_tp2_ep2"
 
diff --git a/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py b/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
index 3149b53d0e..356ffdf5c0 100644
--- a/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
+++ b/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
@@ -10,7 +10,6 @@
     est_time=9000,
     suite="stage-c-8-gpu-h200",
     labels=["ft-long", "long"],
-    disabled="FT soak tests pending CI infra support",
 )
 
 if __name__ == "__main__":

From e3d3b798643aa14614e4ff5b51f89535ca4d6570 Mon Sep 17 00:00:00 2001
From: Jiajun Li <guapisolo@gmail.com>
Date: Tue, 14 Jul 2026 17:55:39 +0000
Subject: [PATCH 2/4] ci: stop FT accuracy soak after metric success

The newly enabled ft-long tests exceeded their stale per-file budgets while still making progress. Propagate metric-checker success to the train loops, opt the realistic GSM8K soak into early completion without relaxing its threshold, and calibrate all three soak estimates from the observed H200 run.
---
 miles/ray/rollout/rollout_manager.py          |  6 ++--
 miles/utils/arguments.py                      | 13 ++++++++
 miles/utils/metric_checker.py                 |  3 +-
 tests/e2e/ft/README.md                        |  5 +--
 .../conftest_ft/scenario_realistic_gsm8k.py   |  1 +
 ..._trainer_ft_random_dp2_cp2_real_rollout.py |  2 +-
 .../test_trainer_ft_random_dp2_cp2_tp2_ep2.py |  2 +-
 .../e2e/ft/test_trainer_ft_realistic_gsm8k.py |  2 +-
 tests/fast/ray/rollout/conftest.py            |  3 ++
 .../rollout/real_ray/test_rollout_manager.py  | 26 +++++++++++++++
 tests/fast/utils/test_arguments.py            |  9 +++++
 tests/fast/utils/test_metric_checker.py       | 33 +++++++++++++++++++
 train.py                                      |  8 +++--
 train_async.py                                |  9 +++--
 14 files changed, 110 insertions(+), 12 deletions(-)
 create mode 100644 tests/fast/utils/test_metric_checker.py

diff --git a/miles/ray/rollout/rollout_manager.py b/miles/ray/rollout/rollout_manager.py
index ecaa9f41e5..8b96bfd173 100644
--- a/miles/ray/rollout/rollout_manager.py
+++ b/miles/ray/rollout/rollout_manager.py
@@ -133,7 +133,7 @@ async def generate(self, rollout_id):
     async def eval(self, rollout_id):
         if self.args.debug_train_only:
             # if debug train only, we don't generate evaluation data
-            return
+            return False
         self._health_monitoring_resume()
 
         if self.use_experimental_refactor:
@@ -148,7 +148,9 @@ async def eval(self, rollout_id):
         save_debug_rollout_data(self.args, data, rollout_id=rollout_id, evaluation=True)
         metrics = log_eval_rollout_data(rollout_id, self.args, data, result.metrics)
         if self._metric_checker is not None:
-            self._metric_checker.on_eval(metrics)
+            check_succeeded = self._metric_checker.on_eval(metrics)
+            return self.args.ci_metric_checker_stop_on_success and check_succeeded
+        return False
 
     async def _get_rollout_data(self, rollout_id):
         if self.args.load_debug_rollout_data:
diff --git a/miles/utils/arguments.py b/miles/utils/arguments.py
index 53dcc34f8e..f57baad918 100644
--- a/miles/utils/arguments.py
+++ b/miles/utils/arguments.py
@@ -1993,6 +1993,11 @@ def add_ci_arguments(parser):
                 type=float,
                 default=None,
             )
+            parser.add_argument(
+                "--ci-metric-checker-stop-on-success",
+                action="store_true",
+                help="Stop training after a CI metric evaluation reaches its threshold.",
+            )
             parser.add_argument(
                 "--ci-save-grad-norm",
                 type=str,
@@ -2266,6 +2271,14 @@ def miles_validate_args(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.ci_metric_checker_stop_on_success and (
+        not args.ci_test or args.ci_metric_checker_key is None or args.ci_metric_checker_threshold is None
+    ):
+        raise ValueError(
+            "--ci-metric-checker-stop-on-success requires --ci-test, "
+            "--ci-metric-checker-key, and --ci-metric-checker-threshold"
+        )
+
     if "train" in args.ft_components:
         args.indep_dp = True
         args.delay_split_train_data_by_dp = True
diff --git a/miles/utils/metric_checker.py b/miles/utils/metric_checker.py
index c4921d3b09..07811862a4 100644
--- a/miles/utils/metric_checker.py
+++ b/miles/utils/metric_checker.py
@@ -14,7 +14,7 @@ def __init__(self, args):
         self.args = args
         self._exists_check_success = False
 
-    def on_eval(self, metrics: dict[str, float]):
+    def on_eval(self, metrics: dict[str, float]) -> bool:
         actual_value = metrics.get(self.args.ci_metric_checker_key)
         assert actual_value is not None, f"{metrics=} {self.args.ci_metric_checker_key=}"
 
@@ -22,6 +22,7 @@ def on_eval(self, metrics: dict[str, float]):
         logger.info(f"[MetricChecker] {check_success=} {actual_value=} {self.args.ci_metric_checker_threshold=}")
 
         self._exists_check_success |= check_success
+        return self._exists_check_success
 
     def dispose(self):
         assert self._exists_check_success, "[MetricChecker] accuracy check failed"
diff --git a/tests/e2e/ft/README.md b/tests/e2e/ft/README.md
index 9d0413a5d0..35a9b7df28 100644
--- a/tests/e2e/ft/README.md
+++ b/tests/e2e/ft/README.md
@@ -270,8 +270,9 @@ Entry `test_trainer_ft_realistic_gsm8k.py`, no mode variants. Runs the same exte
 
 ```
 Type: non-comparison (no baseline run; reference = the baseline test's wandb curves)
-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)
+Recipe: Qwen2.5-0.5B, GRPO, up to 250 rollouts; CI stops after the first eval reaches
+        the 0.55 threshold; 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)
 
diff --git a/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py b/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py
index 5929d222d6..472dc6b58f 100644
--- a/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py
+++ b/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py
@@ -156,6 +156,7 @@ def _get_gsm8k_train_args(*, seed: int, num_rollout: int, metric_threshold: floa
         "--ci-disable-kl-checker "
         "--ci-metric-checker-key eval/gsm8k "
         f"--ci-metric-checker-threshold {metric_threshold} "
+        "--ci-metric-checker-stop-on-success "
     )
 
     misc_args = (
diff --git a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
index 21f42fc543..c5d2b0640f 100644
--- a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
+++ b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
@@ -6,7 +6,7 @@
 from tests.ci.ci_register import register_cuda_ci
 from tests.e2e.ft.conftest_ft.scenario_ft_random import run_ci
 
-register_cuda_ci(est_time=2400, suite="stage-c-8-gpu-h200", labels=["ft-long"])
+register_cuda_ci(est_time=3600, suite="stage-c-8-gpu-h200", labels=["ft-long"])
 
 _MODE: str = "dp2_cp2_real_rollout"
 
diff --git a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
index e1a46fc8bf..2bd259a47d 100644
--- a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
+++ b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
@@ -6,7 +6,7 @@
 from tests.ci.ci_register import register_cuda_ci
 from tests.e2e.ft.conftest_ft.scenario_ft_random import run_ci
 
-register_cuda_ci(est_time=1800, suite="stage-c-8-gpu-h200", labels=["ft-long"])
+register_cuda_ci(est_time=3600, suite="stage-c-8-gpu-h200", labels=["ft-long"])
 
 _MODE: str = "dp2_cp2_tp2_ep2"
 
diff --git a/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py b/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
index 356ffdf5c0..70656fdfb6 100644
--- a/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
+++ b/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
@@ -7,7 +7,7 @@
 from tests.e2e.ft.conftest_ft.scenario_realistic_gsm8k import run_ci
 
 register_cuda_ci(
-    est_time=9000,
+    est_time=16500,
     suite="stage-c-8-gpu-h200",
     labels=["ft-long", "long"],
 )
diff --git a/tests/fast/ray/rollout/conftest.py b/tests/fast/ray/rollout/conftest.py
index 7e5e94fdbd..f7ab58bb8e 100644
--- a/tests/fast/ray/rollout/conftest.py
+++ b/tests/fast/ray/rollout/conftest.py
@@ -107,6 +107,9 @@ def make_args(**overrides: Any) -> Namespace:
         save=None,
         # CI
         ci_test=False,
+        ci_metric_checker_key=None,
+        ci_metric_checker_threshold=None,
+        ci_metric_checker_stop_on_success=False,
         # dumper (sglang debug dumper integration)
         dumper_enable=False,
         dumper_inference=False,
diff --git a/tests/fast/ray/rollout/real_ray/test_rollout_manager.py b/tests/fast/ray/rollout/real_ray/test_rollout_manager.py
index 962a53099d..7d6b15151d 100644
--- a/tests/fast/ray/rollout/real_ray/test_rollout_manager.py
+++ b/tests/fast/ray/rollout/real_ray/test_rollout_manager.py
@@ -561,6 +561,32 @@ def fake_eval_fn(input):
         assert isinstance(captured[0], RolloutFnEvalInput)
         assert captured[0].rollout_id == 10
 
+    async def test_reports_stop_after_metric_success(
+        self,
+        ray_local_mode,
+        placement_group_factory,
+        tmp_path,
+        patch_low_level,
+        monkeypatch,
+    ):
+        import miles.ray.rollout.rollout_manager as rmgr
+
+        args = _make_test_args(tmp_path, models=[("actor", True)])
+        args.ci_test = True
+        args.ci_metric_checker_key = "eval/gsm8k"
+        args.ci_metric_checker_threshold = 0.55
+        args.ci_metric_checker_stop_on_success = True
+        pg = placement_group_factory(2)
+
+        manager = _make_manager(args, pg)
+        manager.eval_generate_rollout = lambda input: RolloutFnEvalOutput(
+            data={"gsm8k": {"rewards": [1.0]}},
+            metrics={},
+        )
+        monkeypatch.setattr(rmgr, "log_eval_rollout_data", lambda *a, **kw: {"eval/gsm8k": 0.56})
+
+        assert await manager.eval(rollout_id=10) is True
+
     async def test_skipped_in_debug_train_only_mode(
         self,
         ray_local_mode,
diff --git a/tests/fast/utils/test_arguments.py b/tests/fast/utils/test_arguments.py
index 20a7bfb758..8f2d48f3fc 100644
--- a/tests/fast/utils/test_arguments.py
+++ b/tests/fast/utils/test_arguments.py
@@ -149,6 +149,15 @@ def test_recompute_logprobs_via_prefill_flag_is_parsed():
     assert args.recompute_logprobs_via_prefill is True
 
 
+def test_ci_metric_checker_stop_on_success_flag_is_parsed():
+    parser = argparse.ArgumentParser()
+    get_miles_extra_args_provider()(parser)
+
+    args = parser.parse_args(["--ci-metric-checker-stop-on-success"] + REQUIRED_ARGS)
+
+    assert args.ci_metric_checker_stop_on_success is True
+
+
 def test_custom_megatron_post_save_hook_path_is_parsed():
     parser = argparse.ArgumentParser()
     get_miles_extra_args_provider()(parser)
diff --git a/tests/fast/utils/test_metric_checker.py b/tests/fast/utils/test_metric_checker.py
new file mode 100644
index 0000000000..059671480b
--- /dev/null
+++ b/tests/fast/utils/test_metric_checker.py
@@ -0,0 +1,33 @@
+from types import SimpleNamespace
+
+import pytest
+
+from miles.utils.metric_checker import MetricChecker
+
+
+def test_on_eval_reports_cumulative_success() -> None:
+    checker = MetricChecker(
+        SimpleNamespace(
+            ci_metric_checker_key="eval/gsm8k",
+            ci_metric_checker_threshold=0.55,
+        )
+    )
+
+    assert checker.on_eval({"eval/gsm8k": 0.54}) is False
+    assert checker.on_eval({"eval/gsm8k": 0.56}) is True
+    assert checker.on_eval({"eval/gsm8k": 0.53}) is True
+    checker.dispose()
+
+
+def test_dispose_rejects_runs_without_a_successful_eval() -> None:
+    checker = MetricChecker(
+        SimpleNamespace(
+            ci_metric_checker_key="eval/gsm8k",
+            ci_metric_checker_threshold=0.55,
+        )
+    )
+
+    checker.on_eval({"eval/gsm8k": 0.54})
+
+    with pytest.raises(AssertionError, match=r"accuracy check failed"):
+        checker.dispose()
diff --git a/train.py b/train.py
index 06f8ac81d1..1d1fcc7a3d 100644
--- a/train.py
+++ b/train.py
@@ -90,7 +90,9 @@ async def save(rollout_id):
     # note that for async training, one can change the position of the sync operation(ray.get).
     for rollout_id in range(args.start_rollout_id, args.num_rollout):
         if args.eval_interval is not None and rollout_id == args.start_rollout_id and not args.skip_eval_before_train:
-            await rollout_manager.eval.remote(rollout_id)
+            if await rollout_manager.eval.remote(rollout_id):
+                logger.info("CI metric threshold reached at rollout_id=%d; stopping training", rollout_id)
+                break
 
         rollout_data_ref = await rollout_manager.generate.remote(rollout_id)
 
@@ -121,7 +123,9 @@ async def save(rollout_id):
             await rollout_manager.onload_kv.remote()
 
         if should_run_periodic_action(rollout_id, args.eval_interval, num_rollout_per_epoch):
-            await rollout_manager.eval.remote(rollout_id)
+            if await rollout_manager.eval.remote(rollout_id):
+                logger.info("CI metric threshold reached at rollout_id=%d; stopping training", rollout_id)
+                break
 
         if (
             args.debug_exit_after_rollout is not None
diff --git a/train_async.py b/train_async.py
index 239265f69c..7900eef163 100644
--- a/train_async.py
+++ b/train_async.py
@@ -53,7 +53,10 @@ async def train(args):
         )
 
     if args.eval_interval is not None and args.start_rollout_id == 0 and not args.skip_eval_before_train:
-        await rollout_manager.eval.remote(0)
+        if await rollout_manager.eval.remote(0):
+            logger.info("CI metric threshold reached at rollout_id=0; stopping training")
+            await rollout_manager.dispose.remote()
+            return
 
     # async train loop.
     rollout_data_next_future = rollout_manager.generate.remote(args.start_rollout_id)
@@ -93,7 +96,9 @@ async def train(args):
             await actor_model.update_weights(rollout_id=rollout_id)
 
         if should_run_periodic_action(rollout_id, args.eval_interval, num_rollout_per_epoch):
-            await rollout_manager.eval.remote(rollout_id)
+            if await rollout_manager.eval.remote(rollout_id):
+                logger.info("CI metric threshold reached at rollout_id=%d; stopping training", rollout_id)
+                break
 
         if (
             args.debug_exit_after_rollout is not None

From 5aec78c4382b3c470918cbd6185eaab2a23359a7 Mon Sep 17 00:00:00 2001
From: Jiajun Li <guapisolo@gmail.com>
Date: Wed, 15 Jul 2026 20:32:44 +0000
Subject: [PATCH 3/4] ci: remove standalone metric checker CPU test

The standalone metric checker test is implicitly collected by stage-a-cpu and is not useful for this FT soak change. Drop it from the PR without changing the metric checker implementation or the remaining stop-on-success coverage.
---
 tests/fast/utils/test_metric_checker.py | 33 -------------------------
 1 file changed, 33 deletions(-)
 delete mode 100644 tests/fast/utils/test_metric_checker.py

diff --git a/tests/fast/utils/test_metric_checker.py b/tests/fast/utils/test_metric_checker.py
deleted file mode 100644
index 059671480b..0000000000
--- a/tests/fast/utils/test_metric_checker.py
+++ /dev/null
@@ -1,33 +0,0 @@
-from types import SimpleNamespace
-
-import pytest
-
-from miles.utils.metric_checker import MetricChecker
-
-
-def test_on_eval_reports_cumulative_success() -> None:
-    checker = MetricChecker(
-        SimpleNamespace(
-            ci_metric_checker_key="eval/gsm8k",
-            ci_metric_checker_threshold=0.55,
-        )
-    )
-
-    assert checker.on_eval({"eval/gsm8k": 0.54}) is False
-    assert checker.on_eval({"eval/gsm8k": 0.56}) is True
-    assert checker.on_eval({"eval/gsm8k": 0.53}) is True
-    checker.dispose()
-
-
-def test_dispose_rejects_runs_without_a_successful_eval() -> None:
-    checker = MetricChecker(
-        SimpleNamespace(
-            ci_metric_checker_key="eval/gsm8k",
-            ci_metric_checker_threshold=0.55,
-        )
-    )
-
-    checker.on_eval({"eval/gsm8k": 0.54})
-
-    with pytest.raises(AssertionError, match=r"accuracy check failed"):
-        checker.dispose()

From c918ba22adef581b140389c896d89812bcbea9b7 Mon Sep 17 00:00:00 2001
From: Jiajun Li <guapisolo@gmail.com>
Date: Wed, 15 Jul 2026 20:45:31 +0000
Subject: [PATCH 4/4] ci: limit FT soak enablement to two random tests

FT soak tests must complete their configured rollout counts instead of stopping after a metric passes. Remove the early-stop flag and its wiring, restore the original estimates, and keep realistic GSM8K disabled so this PR only enables the two random FT cases.
---
 miles/ray/rollout/rollout_manager.py          |  6 ++---
 miles/utils/arguments.py                      | 13 ----------
 miles/utils/metric_checker.py                 |  3 +--
 tests/e2e/ft/README.md                        |  5 ++--
 .../conftest_ft/scenario_realistic_gsm8k.py   |  1 -
 ..._trainer_ft_random_dp2_cp2_real_rollout.py |  2 +-
 .../test_trainer_ft_random_dp2_cp2_tp2_ep2.py |  2 +-
 .../e2e/ft/test_trainer_ft_realistic_gsm8k.py |  2 +-
 tests/fast/ray/rollout/conftest.py            |  3 ---
 .../rollout/real_ray/test_rollout_manager.py  | 26 -------------------
 tests/fast/utils/test_arguments.py            |  9 -------
 train.py                                      |  8 ++----
 train_async.py                                |  9 ++-----
 13 files changed, 12 insertions(+), 77 deletions(-)

diff --git a/miles/ray/rollout/rollout_manager.py b/miles/ray/rollout/rollout_manager.py
index 8b96bfd173..ecaa9f41e5 100644
--- a/miles/ray/rollout/rollout_manager.py
+++ b/miles/ray/rollout/rollout_manager.py
@@ -133,7 +133,7 @@ async def generate(self, rollout_id):
     async def eval(self, rollout_id):
         if self.args.debug_train_only:
             # if debug train only, we don't generate evaluation data
-            return False
+            return
         self._health_monitoring_resume()
 
         if self.use_experimental_refactor:
@@ -148,9 +148,7 @@ async def eval(self, rollout_id):
         save_debug_rollout_data(self.args, data, rollout_id=rollout_id, evaluation=True)
         metrics = log_eval_rollout_data(rollout_id, self.args, data, result.metrics)
         if self._metric_checker is not None:
-            check_succeeded = self._metric_checker.on_eval(metrics)
-            return self.args.ci_metric_checker_stop_on_success and check_succeeded
-        return False
+            self._metric_checker.on_eval(metrics)
 
     async def _get_rollout_data(self, rollout_id):
         if self.args.load_debug_rollout_data:
diff --git a/miles/utils/arguments.py b/miles/utils/arguments.py
index f57baad918..53dcc34f8e 100644
--- a/miles/utils/arguments.py
+++ b/miles/utils/arguments.py
@@ -1993,11 +1993,6 @@ def add_ci_arguments(parser):
                 type=float,
                 default=None,
             )
-            parser.add_argument(
-                "--ci-metric-checker-stop-on-success",
-                action="store_true",
-                help="Stop training after a CI metric evaluation reaches its threshold.",
-            )
             parser.add_argument(
                 "--ci-save-grad-norm",
                 type=str,
@@ -2271,14 +2266,6 @@ def miles_validate_args(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.ci_metric_checker_stop_on_success and (
-        not args.ci_test or args.ci_metric_checker_key is None or args.ci_metric_checker_threshold is None
-    ):
-        raise ValueError(
-            "--ci-metric-checker-stop-on-success requires --ci-test, "
-            "--ci-metric-checker-key, and --ci-metric-checker-threshold"
-        )
-
     if "train" in args.ft_components:
         args.indep_dp = True
         args.delay_split_train_data_by_dp = True
diff --git a/miles/utils/metric_checker.py b/miles/utils/metric_checker.py
index 07811862a4..c4921d3b09 100644
--- a/miles/utils/metric_checker.py
+++ b/miles/utils/metric_checker.py
@@ -14,7 +14,7 @@ def __init__(self, args):
         self.args = args
         self._exists_check_success = False
 
-    def on_eval(self, metrics: dict[str, float]) -> bool:
+    def on_eval(self, metrics: dict[str, float]):
         actual_value = metrics.get(self.args.ci_metric_checker_key)
         assert actual_value is not None, f"{metrics=} {self.args.ci_metric_checker_key=}"
 
@@ -22,7 +22,6 @@ def on_eval(self, metrics: dict[str, float]) -> bool:
         logger.info(f"[MetricChecker] {check_success=} {actual_value=} {self.args.ci_metric_checker_threshold=}")
 
         self._exists_check_success |= check_success
-        return self._exists_check_success
 
     def dispose(self):
         assert self._exists_check_success, "[MetricChecker] accuracy check failed"
diff --git a/tests/e2e/ft/README.md b/tests/e2e/ft/README.md
index 35a9b7df28..9d0413a5d0 100644
--- a/tests/e2e/ft/README.md
+++ b/tests/e2e/ft/README.md
@@ -270,9 +270,8 @@ Entry `test_trainer_ft_realistic_gsm8k.py`, no mode variants. Runs the same exte
 
 ```
 Type: non-comparison (no baseline run; reference = the baseline test's wandb curves)
-Recipe: Qwen2.5-0.5B, GRPO, up to 250 rollouts; CI stops after the first eval reaches
-        the 0.55 threshold; parallelism mirrors dp2_cp2_real_rollout (2 cells x CP2
-        on 4 train GPUs + 4 rollout engines x 1 GPU, disaggregated)
+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)
 
diff --git a/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py b/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py
index 472dc6b58f..5929d222d6 100644
--- a/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py
+++ b/tests/e2e/ft/conftest_ft/scenario_realistic_gsm8k.py
@@ -156,7 +156,6 @@ def _get_gsm8k_train_args(*, seed: int, num_rollout: int, metric_threshold: floa
         "--ci-disable-kl-checker "
         "--ci-metric-checker-key eval/gsm8k "
         f"--ci-metric-checker-threshold {metric_threshold} "
-        "--ci-metric-checker-stop-on-success "
     )
 
     misc_args = (
diff --git a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
index c5d2b0640f..21f42fc543 100644
--- a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
+++ b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_real_rollout.py
@@ -6,7 +6,7 @@
 from tests.ci.ci_register import register_cuda_ci
 from tests.e2e.ft.conftest_ft.scenario_ft_random import run_ci
 
-register_cuda_ci(est_time=3600, suite="stage-c-8-gpu-h200", labels=["ft-long"])
+register_cuda_ci(est_time=2400, suite="stage-c-8-gpu-h200", labels=["ft-long"])
 
 _MODE: str = "dp2_cp2_real_rollout"
 
diff --git a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
index 2bd259a47d..e1a46fc8bf 100644
--- a/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
+++ b/tests/e2e/ft/test_trainer_ft_random_dp2_cp2_tp2_ep2.py
@@ -6,7 +6,7 @@
 from tests.ci.ci_register import register_cuda_ci
 from tests.e2e.ft.conftest_ft.scenario_ft_random import run_ci
 
-register_cuda_ci(est_time=3600, suite="stage-c-8-gpu-h200", labels=["ft-long"])
+register_cuda_ci(est_time=1800, suite="stage-c-8-gpu-h200", labels=["ft-long"])
 
 _MODE: str = "dp2_cp2_tp2_ep2"
 
diff --git a/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py b/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
index 541eeb1824..a1077f4a54 100644
--- a/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
+++ b/tests/e2e/ft/test_trainer_ft_realistic_gsm8k.py
@@ -7,7 +7,7 @@
 from tests.e2e.ft.conftest_ft.scenario_realistic_gsm8k import run_ci
 
 register_cuda_ci(
-    est_time=16500,
+    est_time=9000,
     suite="stage-c-8-gpu-h200",
     labels=["ft-long"],
     disabled="FT soak tests pending CI infra support",
diff --git a/tests/fast/ray/rollout/conftest.py b/tests/fast/ray/rollout/conftest.py
index f7ab58bb8e..7e5e94fdbd 100644
--- a/tests/fast/ray/rollout/conftest.py
+++ b/tests/fast/ray/rollout/conftest.py
@@ -107,9 +107,6 @@ def make_args(**overrides: Any) -> Namespace:
         save=None,
         # CI
         ci_test=False,
-        ci_metric_checker_key=None,
-        ci_metric_checker_threshold=None,
-        ci_metric_checker_stop_on_success=False,
         # dumper (sglang debug dumper integration)
         dumper_enable=False,
         dumper_inference=False,
diff --git a/tests/fast/ray/rollout/real_ray/test_rollout_manager.py b/tests/fast/ray/rollout/real_ray/test_rollout_manager.py
index 7d6b15151d..962a53099d 100644
--- a/tests/fast/ray/rollout/real_ray/test_rollout_manager.py
+++ b/tests/fast/ray/rollout/real_ray/test_rollout_manager.py
@@ -561,32 +561,6 @@ def fake_eval_fn(input):
         assert isinstance(captured[0], RolloutFnEvalInput)
         assert captured[0].rollout_id == 10
 
-    async def test_reports_stop_after_metric_success(
-        self,
-        ray_local_mode,
-        placement_group_factory,
-        tmp_path,
-        patch_low_level,
-        monkeypatch,
-    ):
-        import miles.ray.rollout.rollout_manager as rmgr
-
-        args = _make_test_args(tmp_path, models=[("actor", True)])
-        args.ci_test = True
-        args.ci_metric_checker_key = "eval/gsm8k"
-        args.ci_metric_checker_threshold = 0.55
-        args.ci_metric_checker_stop_on_success = True
-        pg = placement_group_factory(2)
-
-        manager = _make_manager(args, pg)
-        manager.eval_generate_rollout = lambda input: RolloutFnEvalOutput(
-            data={"gsm8k": {"rewards": [1.0]}},
-            metrics={},
-        )
-        monkeypatch.setattr(rmgr, "log_eval_rollout_data", lambda *a, **kw: {"eval/gsm8k": 0.56})
-
-        assert await manager.eval(rollout_id=10) is True
-
     async def test_skipped_in_debug_train_only_mode(
         self,
         ray_local_mode,
diff --git a/tests/fast/utils/test_arguments.py b/tests/fast/utils/test_arguments.py
index 8f2d48f3fc..20a7bfb758 100644
--- a/tests/fast/utils/test_arguments.py
+++ b/tests/fast/utils/test_arguments.py
@@ -149,15 +149,6 @@ def test_recompute_logprobs_via_prefill_flag_is_parsed():
     assert args.recompute_logprobs_via_prefill is True
 
 
-def test_ci_metric_checker_stop_on_success_flag_is_parsed():
-    parser = argparse.ArgumentParser()
-    get_miles_extra_args_provider()(parser)
-
-    args = parser.parse_args(["--ci-metric-checker-stop-on-success"] + REQUIRED_ARGS)
-
-    assert args.ci_metric_checker_stop_on_success is True
-
-
 def test_custom_megatron_post_save_hook_path_is_parsed():
     parser = argparse.ArgumentParser()
     get_miles_extra_args_provider()(parser)
diff --git a/train.py b/train.py
index 1d1fcc7a3d..06f8ac81d1 100644
--- a/train.py
+++ b/train.py
@@ -90,9 +90,7 @@ async def save(rollout_id):
     # note that for async training, one can change the position of the sync operation(ray.get).
     for rollout_id in range(args.start_rollout_id, args.num_rollout):
         if args.eval_interval is not None and rollout_id == args.start_rollout_id and not args.skip_eval_before_train:
-            if await rollout_manager.eval.remote(rollout_id):
-                logger.info("CI metric threshold reached at rollout_id=%d; stopping training", rollout_id)
-                break
+            await rollout_manager.eval.remote(rollout_id)
 
         rollout_data_ref = await rollout_manager.generate.remote(rollout_id)
 
@@ -123,9 +121,7 @@ async def save(rollout_id):
             await rollout_manager.onload_kv.remote()
 
         if should_run_periodic_action(rollout_id, args.eval_interval, num_rollout_per_epoch):
-            if await rollout_manager.eval.remote(rollout_id):
-                logger.info("CI metric threshold reached at rollout_id=%d; stopping training", rollout_id)
-                break
+            await rollout_manager.eval.remote(rollout_id)
 
         if (
             args.debug_exit_after_rollout is not None
diff --git a/train_async.py b/train_async.py
index 7900eef163..239265f69c 100644
--- a/train_async.py
+++ b/train_async.py
@@ -53,10 +53,7 @@ async def train(args):
         )
 
     if args.eval_interval is not None and args.start_rollout_id == 0 and not args.skip_eval_before_train:
-        if await rollout_manager.eval.remote(0):
-            logger.info("CI metric threshold reached at rollout_id=0; stopping training")
-            await rollout_manager.dispose.remote()
-            return
+        await rollout_manager.eval.remote(0)
 
     # async train loop.
     rollout_data_next_future = rollout_manager.generate.remote(args.start_rollout_id)
@@ -96,9 +93,7 @@ async def train(args):
             await actor_model.update_weights(rollout_id=rollout_id)
 
         if should_run_periodic_action(rollout_id, args.eval_interval, num_rollout_per_epoch):
-            if await rollout_manager.eval.remote(rollout_id):
-                logger.info("CI metric threshold reached at rollout_id=%d; stopping training", rollout_id)
-                break
+            await rollout_manager.eval.remote(rollout_id)
 
         if (
             args.debug_exit_after_rollout is not None