test(session): keep training-session sampling tests unbounded

tests/fast/rollout/session/test_request_args.py

@@ -249,6 +249,8 @@ def resolve_request_args(self, request_args, *, turn_args):
 
 @pytest.mark.parametrize("evaluation", [False, True])
 def test_session_sampling_defaults_fill_only_omitted_fields(evaluation):
+    # Training defaults stay unbounded; see SAMPLING in tests/fast/router/test_session_evaluation.py.
+    defaults = {"temperature": 0.6, "top_p": 0.9 if evaluation else 1.0, "top_k": 20 if evaluation else -1}
     temperature = 0.1 if evaluation else 0.6
     client_args = {"temperature": temperature, "top_p": None, "top_k": -1}
     original = deepcopy(client_args)
@@ -258,11 +260,11 @@ def test_session_sampling_defaults_fill_only_omitted_fields(evaluation):
         config=make_session_server_config(),
         turn_args=None,
         evaluation=evaluation,
-        sampling_defaults={"temperature": 0.6, "top_p": 0.9, "top_k": 20},
+        sampling_defaults=defaults,
     )
     assert {key: prepared.body[key] for key in ("temperature", "top_p", "top_k")} == {
         "temperature": temperature,
-        "top_p": 0.9,
+        "top_p": defaults["top_p"],
         "top_k": -1,
     }
     assert client_args == original

tests/fast/router/test_session_evaluation.py

@@ -183,18 +183,22 @@ async def test_concurrent_train_and_eval_do_not_share_policy(env):
     assert sorted(request["return_indexer_topk"] for request in env.backend.requests) == [False, True]
 
 
-SAMPLING = {"temperature": 0.6, "top_p": 0.9, "top_k": 20}
+# Training sessions stay unbounded: sampling-support replay rejects bounded training sampling
+# (top_p < 1 or top_k > 0) on a run that does not capture rollout sampling support.
+SAMPLING = {"temperature": 0.6, "top_p": 1.0, "top_k": -1}
+EVAL_SAMPLING = {"temperature": 0.6, "top_p": 0.9, "top_k": 20}
 
 
 @pytest.mark.parametrize("evaluation", [False, True])
 async def test_creation_sampling_defaults_fill_only_omitted_fields(env, evaluation):
-    sid = await _create(env, json.dumps({**SAMPLING, "evaluation": evaluation}).encode())
+    defaults = EVAL_SAMPLING if evaluation else SAMPLING
+    sid = await _create(env, json.dumps({**defaults, "evaluation": evaluation}).encode())
     await _chat(env, sid, [USER])
-    temperature = 0.1 if evaluation else SAMPLING["temperature"]
+    temperature = 0.1 if evaluation else defaults["temperature"]
     await _chat(env, sid, [USER, ASSISTANT, TOOL], temperature=temperature, top_p=None, top_k=-1)
     omitted, explicit = env.backend.requests[-2:]
-    assert {key: omitted[key] for key in SAMPLING} == SAMPLING
-    assert {key: explicit[key] for key in SAMPLING} == {**SAMPLING, "temperature": temperature, "top_k": -1}
+    assert {key: omitted[key] for key in defaults} == defaults
+    assert {key: explicit[key] for key in defaults} == {**defaults, "temperature": temperature, "top_k": -1}
 
 
 @pytest.mark.parametrize("saved,requested", [(0.6, 0.1), (0.6, 0.0), (0.0, 0.6)])
@@ -212,9 +216,9 @@ async def test_training_temperature_mismatch_is_rejected_before_forwarding(env,
             f"temperature={requested!r} does not match the training session temperature={saved!r}"
         )
         assert len(env.backend.requests) == request_count
-        await _chat(env, sid, messages, temperature=saved, top_p=0.5, top_k=-1)
+        await _chat(env, sid, messages, temperature=saved, top_p=1.0, top_k=-1)
         wire = env.backend.requests[-1]
-        assert (wire["temperature"], wire["top_p"], wire["top_k"]) == (saved, 0.5, -1)
+        assert (wire["temperature"], wire["top_p"], wire["top_k"]) == (saved, 1.0, -1)
 
 
 async def test_training_null_temperature_uses_registered_value(env):
@@ -240,7 +244,7 @@ async def test_an_integer_temperature_is_accepted_as_a_float_default(env):
 
 async def test_an_integral_float_top_k_is_stored_as_an_int(env):
     """An eval dataset YAML can spell top_k as 40.0; the engine must still receive an integer."""
-    sid = await _create(env, b'{"top_k": 40.0}')
+    sid = await _create(env, b'{"evaluation": true, "top_k": 40.0}')
     await _chat(env, sid, [USER])
     top_k = env.backend.requests[-1]["top_k"]
     assert top_k == 40 and isinstance(top_k, int)