S26-E7: reject verl with expert parallelism; require fsdp2 strategy

tests/test_rl_verl_adapter.py

@@ -447,3 +447,12 @@ def test_verl_data_cursor_refuses_what_it_cannot_honour():
         cur.seek({"sample_offset": 2})
     with pytest.raises(RuntimeError, match="twice"):
         VerlDataCursor(trainer)
+
+
+def test_verl_refuses_expert_parallel_with_e7_reason():
+    # S26 E7: verl+EP is refused with the reason and the backend to use instead.
+    from yeto.rl.adapters.verl import run_config_rules as rules
+
+    rules.check_expert_parallel(8, 1)
+    with pytest.raises(ValueError, match="E7.*Miles"):
+        rules.check_expert_parallel(8, 8)

tests/test_stage1_plan.py

@@ -250,3 +250,13 @@ def edit(doc):
         doc["trainer"]["miles"]["memory_estimate"]["host_optimizer_gib_per_node"] = 600
     out = _variant(tmp_path, edit)
     assert any("host_optimizer_gib_per_node = 600" in e for e in out.errors)
+
+
+def test_verl_backend_refuses_megatron_strategy(tmp_path):
+    # S26 E7: verl Megatron (the verl path with EP) is not wired; EP runs use Miles.
+    out = _variant(tmp_path, lambda d: d["trainer"].update(backend="verl")
+                   or d["trainer"]["verl"].update(strategy="megatron"))
+    assert any("verl.strategy 'megatron'" in e and "backend miles" in e for e in out.errors)
+    # The same strategy value is not checked for the Miles backend (its section is unused).
+    miles = _variant(tmp_path, lambda d: d["trainer"]["verl"].update(strategy="megatron"))
+    assert not any("verl.strategy" in e for e in miles.errors)

yeto/rl/adapters/verl/run_config_rules.py

@@ -19,9 +19,15 @@ def check_trainer_parallel(actor_gpus: int, tensor_parallel: int, pipeline_paral
         raise ValueError("verl backend first step: one GPU per island")
 
 
+# S26 E7 decision: the verl Megatron engine (the only verl path with expert
+# parallelism) is not wired into this adapter; MoE runs with EP go to Miles.
+EP_REFUSAL = ("verl 后端不支持专家并行（EP>1）：yeto 的 verl 适配层只接 FSDP2，"
+              "未接 verl Megatron 引擎（S26 E7 结论：不做）。EP 训练请用 Miles 后端")
+
+
 def check_expert_parallel(actor_gpus: int, expert_parallel: int) -> None:
     if expert_parallel != 1:
-        raise ValueError("verl backend first step: no expert parallelism")
+        raise ValueError(f"{EP_REFUSAL}；收到 expert_parallel={expert_parallel}")
 
 
 def check_global_batch(global_batch: int, data_parallel: int) -> None:

yeto/rl/stage1.py

@@ -65,6 +65,7 @@
 
 PLAN_SCHEMA = "yeto.stage1/v1"
 BACKENDS = ("miles", "verl")
+VERL_STRATEGY = "fsdp2"  # the only verl actor strategy the adapter runs (S26 E7)
 NEUTRAL_PENALTIES = {"repetition_penalty": 1.0, "frequency_penalty": 0.0, "presence_penalty": 0.0}
 RAY_PORT = 6379
 MEMORY_TOLERANCE = 0.02  # plan memory figures vs the estimator (both 推算)
@@ -819,6 +820,10 @@ def check(plan: Stage1Plan, out: Expanded) -> list[str]:
     if "fp8" in flat or "fp4" in flat:
         errors.append("trainer arguments enable FP8/FP4; training is bf16 only (user 10-19, report #1)")
     verl = _section(trainer_cfg, "verl")
+    if out.trainer["backend"] == "verl" and str(verl["strategy"]) != VERL_STRATEGY:
+        # S26 E7: the verl Megatron engine (EP) is not wired; EP runs use Miles.
+        errors.append(f"verl.strategy {verl['strategy']!r}: the yeto verl adapter runs {VERL_STRATEGY} "
+                      "only (verl Megatron/EP not supported, S26 E7); use trainer.backend miles for EP")
     if int(verl["fsdp_size"]) not in (0, out.trainer["cards"]) and out.trainer["cards"] % int(verl["fsdp_size"]):
         errors.append(f"verl.fsdp_size {verl['fsdp_size']} does not divide {out.trainer['cards']} cards")