From ec452d79cac764c447fab74601bd9b31f1551a77 Mon Sep 17 00:00:00 2001
From: Tianyi Zhang <123608656+tianyi-zhang-02@users.noreply.github.com>
Date: Thu, 6 Aug 2026 00:21:47 -0700
Subject: [PATCH] fix(sglang): expose cfg on SGLangGeneration

create_weight_synchronizer reads generation.cfg unconditionally, before any
backend dispatch. SGLangGeneration stored the generation config only as
self.sglang_cfg and defined no cfg attribute, so backend=sglang raised
AttributeError there instead of getting its HTTPWeightSynchronizer.

VllmGeneration, TRTLLMGeneration and MegatronGeneration all expose cfg;
megatron_generation.py documents it as the GenerationInterface contract.
sglang_cfg already is that same config object, so alias it with a read-only
property rather than keep a second reference that could drift.

Signed-off-by: Tianyi Zhang <123608656+tianyi-zhang-02@users.noreply.github.com>
---
 .../generation/sglang/sglang_generation.py    | 11 +++++++
 .../weight_sync/test_weight_synchronizer.py   | 30 +++++++++++++++++++
 2 files changed, 41 insertions(+)

diff --git a/nemo_rl/models/generation/sglang/sglang_generation.py b/nemo_rl/models/generation/sglang/sglang_generation.py
index 9aa21547a6..6c7ac384e4 100644
--- a/nemo_rl/models/generation/sglang/sglang_generation.py
+++ b/nemo_rl/models/generation/sglang/sglang_generation.py
@@ -122,6 +122,17 @@ def __init__(
         if init_handles:
             ray.get(init_handles)
 
+    @property
+    def cfg(self) -> SGLangConfig:
+        """Full generation config, matching the ``GenerationInterface`` contract.
+
+        ``sglang_cfg`` already is that config, so alias it rather than keep a
+        second reference. Backend-agnostic callers such as
+        ``nemo_rl.weight_sync.factory.create_weight_synchronizer`` read
+        ``generation.cfg`` regardless of backend.
+        """
+        return self.sglang_cfg
+
     # ------------------------------------------------------------------
     # Engine topology properties (formerly ``ServerGroup``)
     # ------------------------------------------------------------------
diff --git a/tests/unit/weight_sync/test_weight_synchronizer.py b/tests/unit/weight_sync/test_weight_synchronizer.py
index e799fc56bd..6cbda470d8 100644
--- a/tests/unit/weight_sync/test_weight_synchronizer.py
+++ b/tests/unit/weight_sync/test_weight_synchronizer.py
@@ -444,6 +444,36 @@ def test_colocated_sglang_returns_http(self):
         )
         assert isinstance(sync, HTTPWeightSynchronizer)
 
+    def test_colocated_sglang_accepts_real_generation_object(self):
+        """A real SGLangGeneration must expose ``cfg`` like the other backends.
+
+        ``create_weight_synchronizer`` reads ``generation.cfg`` before any
+        backend dispatch. The MagicMock double used above answers ``cfg``
+        automatically, so only the real class can catch a missing attribute.
+        """
+        from nemo_rl.models.generation.sglang.sglang_generation import (
+            SGLangGeneration,
+        )
+
+        # ``__new__`` skips ``__init__``, which would need Ray, a router and
+        # live engines; ``cfg`` only depends on ``sglang_cfg``.
+        gen = SGLangGeneration.__new__(SGLangGeneration)
+        gen.sglang_cfg = {"backend": "sglang", "model_name": "dummy"}
+        # ``__del__`` calls ``shutdown()``, which reads these four attributes.
+        gen.all_engines = []
+        gen._router_actor = None
+        gen._http_client = None
+        gen._async_loop = None
+        assert gen.cfg is gen.sglang_cfg
+
+        sync = create_weight_synchronizer(
+            policy=_mock_policy(),
+            generation=gen,
+            generation_backend=SGLANG_BACKEND,
+            colocated=True,
+        )
+        assert isinstance(sync, HTTPWeightSynchronizer)
+
     def test_colocated_megatron_returns_ipc(self):
         policy = _mock_policy()
         gen = _mock_generation()