From 83325b2f7602b1df4c5e448a7946a94979eb4455 Mon Sep 17 00:00:00 2001
From: Gahow Wang <gahow.wang@gmail.com>
Date: Sat, 11 Apr 2026 00:36:45 +0800
Subject: [PATCH] Reset new topology groups to full binary search

---
 src/aituner/llm.py      |  5 +++--
 src/aituner/store.py    |  6 ------
 tests/test_core_flow.py | 13 +++++++++++--
 3 files changed, 14 insertions(+), 10 deletions(-)

diff --git a/src/aituner/llm.py b/src/aituner/llm.py
index f6f8f4b..79ac9ea 100644
--- a/src/aituner/llm.py
+++ b/src/aituner/llm.py
@@ -301,8 +301,9 @@ def build_prompt(
         "",
         "The primary cross-topology comparison metric is request_rate_per_gpu, not raw request_rate.",
         "The proposal should beat the incumbent on request_rate_per_gpu under the 95%+ SLO target.",
-        "The evaluator uses the best feasible sampling_u from the same tp_dp_product group when it exists. For a different tp_dp_product group, it uses a scaled lower floor instead of reusing the global incumbent directly.",
-        "Do not assume a configuration with fewer GPUs must start from the global incumbent sampling_u.",
+        "The evaluator uses the best feasible sampling_u from the same tp_dp_product group when it exists.",
+        "If a tp_dp_product group has no history yet, the evaluator starts from the study's original search.low and runs a full binary search for that group.",
+        "Do not assume a configuration with fewer GPUs should inherit the global incumbent sampling_u.",
     ]
     return "\n".join(sections)
 
diff --git a/src/aituner/store.py b/src/aituner/store.py
index bb9e49e..c023e71 100644
--- a/src/aituner/store.py
+++ b/src/aituner/store.py
@@ -253,12 +253,6 @@ def _derive_search_floor(*, study: StudySpec, state: StudyState, parallel_size:
         group_incumbent.get("best_sampling_u"), (int, float)
     ):
         candidate = float(group_incumbent["best_sampling_u"])
-    elif (
-        isinstance(state.best_sampling_u, (int, float))
-        and isinstance(state.best_parallel_size, int)
-        and state.best_parallel_size > 0
-    ):
-        candidate = float(state.best_sampling_u) * float(parallel_size) / float(state.best_parallel_size)
     else:
         candidate = low
     return min(high, max(low, candidate))
diff --git a/tests/test_core_flow.py b/tests/test_core_flow.py
index 1e7b856..9e7e484 100644
--- a/tests/test_core_flow.py
+++ b/tests/test_core_flow.py
@@ -1061,6 +1061,15 @@ class CoreFlowTests(unittest.TestCase):
                 best_request_rate=3.0,
                 best_request_rate_per_gpu=0.75,
                 next_trial_index=2,
+                best_by_parallel_size={
+                    "4": {
+                        "trial_id": "trial-0001",
+                        "parallel_size": 4,
+                        "best_sampling_u": 0.375,
+                        "best_request_rate": 3.0,
+                        "best_request_rate_per_gpu": 0.75,
+                    }
+                },
                 trials=[],
             )
             proposal = Proposal.from_dict(
@@ -1112,7 +1121,7 @@ class CoreFlowTests(unittest.TestCase):
             trial, _ = store.materialize_trial(study=study, state=state, proposal=proposal)
             self.assertEqual(trial.search.low, 0.125)
 
-    def test_materialize_trial_scales_search_floor_for_different_parallel_group(self) -> None:
+    def test_materialize_trial_resets_search_floor_for_new_parallel_group(self) -> None:
         with tempfile.TemporaryDirectory() as tmp:
             tmp_path = Path(tmp)
             study_path = _write_study_assets(tmp_path)
@@ -1138,7 +1147,7 @@ class CoreFlowTests(unittest.TestCase):
                 }
             )
             trial, _ = store.materialize_trial(study=study, state=state, proposal=proposal)
-            self.assertEqual(trial.search.low, 0.2)
+            self.assertEqual(trial.search.low, study.search.low)
 
     def test_ingest_trial_results_records_failure_reason(self) -> None:
         with tempfile.TemporaryDirectory() as tmp: