From eb4a9837611531c59106af7e8ac877b8e5dcbe20 Mon Sep 17 00:00:00 2001
From: xyr <1615681145@qq.com>
Date: Tue, 14 Jul 2026 16:40:38 +0800
Subject: [PATCH 1/3] =?UTF-8?q?fix:=20=E4=BF=AE=E6=94=B9visible=5Fdevice?=
 =?UTF-8?q?=E5=8F=82=E6=95=B0=E6=8A=A5=E9=94=99=E5=92=8Cdense=E6=A8=A1?=
 =?UTF-8?q?=E5=9E=8B=E6=B2=A1=E6=9C=89=5Frouter=5Freplay=5Fenabled?=
 =?UTF-8?q?=E5=B1=9E=E6=80=A7=E6=8A=A5=E9=94=99?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/twinkle/model/transformers/transformers.py | 2 +-
 src/twinkle/utils/device_mesh.py               | 1 +
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/twinkle/model/transformers/transformers.py b/src/twinkle/model/transformers/transformers.py
index c8fec86d..9be48cae 100644
--- a/src/twinkle/model/transformers/transformers.py
+++ b/src/twinkle/model/transformers/transformers.py
@@ -384,7 +384,7 @@ def _router_replay_setup(self, router_replay_action, routed_experts=None, batch_
         Returns ``cleanup_fn``.
         Call *cleanup_fn* after the forward to gather recorded indices(when RECORD) and clear global state.
         """
-        if not self._router_replay_enabled:
+        if not getattr(self, '_router_replay_enabled', False):
             return lambda: None
 
         self._maybe_apply_router_replay()
diff --git a/src/twinkle/utils/device_mesh.py b/src/twinkle/utils/device_mesh.py
index 7215f7fc..aa8b1c6e 100644
--- a/src/twinkle/utils/device_mesh.py
+++ b/src/twinkle/utils/device_mesh.py
@@ -538,6 +538,7 @@ class DeviceGroup:
     ranks: Union[List[int], int]
     device_type: str
     gpus_per_worker: int = 1
+    visible_devices: Optional[Union[List[str],List[int], str, int]] = None
     _device_mesh: Dict[str, DeviceMesh] = field(default_factory=dict)
 
 

From a2f1be96ac9b14edb90f83569452dc7bf8ffddb1 Mon Sep 17 00:00:00 2001
From: xyr <1615681145@qq.com>
Date: Tue, 14 Jul 2026 17:22:22 +0800
Subject: [PATCH 2/3] =?UTF-8?q?style:=20=E4=BF=AE=E5=A4=8D=20pre-commit=20?=
 =?UTF-8?q?=E8=87=AA=E5=8A=A8=E6=A0=BC=E5=BC=8F=E5=8C=96=E9=97=AE=E9=A2=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/twinkle/kernel/builtin.py    | 9 +++++----
 src/twinkle/utils/device_mesh.py | 2 +-
 2 files changed, 6 insertions(+), 5 deletions(-)

diff --git a/src/twinkle/kernel/builtin.py b/src/twinkle/kernel/builtin.py
index faffb4e7..7675767b 100644
--- a/src/twinkle/kernel/builtin.py
+++ b/src/twinkle/kernel/builtin.py
@@ -90,10 +90,11 @@ def npu_builtin(model: nn.Module | None = None) -> dict[Any, dict[str, Any]]:
         if exists('mindspeed>=0.12.1') and exists('flash-linear-attention'):
             apply_qwen3_5_fla(model)
         else:
-            logger.warning('[NPU] [FLA] mindspeed or flash-linear-attention is not installed, or mindspeed version smaller than 0.12.1; '
-                           'FLA patch for Qwen3.5 requires mindspeed and flash-linear-attention. '
-                           'Install them with: pip install flash-linear-attention and mindspeed from source code via https://gitcode.com/Ascend/MindSpeed.'
-                           'Other NPU patches will still apply.')
+            logger.warning(
+                '[NPU] [FLA] mindspeed or flash-linear-attention is not installed, or mindspeed version smaller than 0.12.1; '
+                'FLA patch for Qwen3.5 requires mindspeed and flash-linear-attention. '
+                'Install them with: pip install flash-linear-attention and mindspeed from source code via https://gitcode.com/Ascend/MindSpeed.'
+                'Other NPU patches will still apply.')
     return bundle
 
 
diff --git a/src/twinkle/utils/device_mesh.py b/src/twinkle/utils/device_mesh.py
index aa8b1c6e..20cda0cf 100644
--- a/src/twinkle/utils/device_mesh.py
+++ b/src/twinkle/utils/device_mesh.py
@@ -538,7 +538,7 @@ class DeviceGroup:
     ranks: Union[List[int], int]
     device_type: str
     gpus_per_worker: int = 1
-    visible_devices: Optional[Union[List[str],List[int], str, int]] = None
+    visible_devices: Optional[Union[List[str], List[int], str, int]] = None
     _device_mesh: Dict[str, DeviceMesh] = field(default_factory=dict)
 
 

From 71c9bc6bc013434dbbf08fd54ad725b3d5107d7f Mon Sep 17 00:00:00 2001
From: xyr <1615681145@qq.com>
Date: Wed, 15 Jul 2026 10:11:18 +0800
Subject: [PATCH 3/3] fix: resolve flake8 formatting in builtin.py

---
 src/twinkle/kernel/builtin.py | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/src/twinkle/kernel/builtin.py b/src/twinkle/kernel/builtin.py
index 7675767b..3a6de079 100644
--- a/src/twinkle/kernel/builtin.py
+++ b/src/twinkle/kernel/builtin.py
@@ -90,11 +90,13 @@ def npu_builtin(model: nn.Module | None = None) -> dict[Any, dict[str, Any]]:
         if exists('mindspeed>=0.12.1') and exists('flash-linear-attention'):
             apply_qwen3_5_fla(model)
         else:
-            logger.warning(
-                '[NPU] [FLA] mindspeed or flash-linear-attention is not installed, or mindspeed version smaller than 0.12.1; '
-                'FLA patch for Qwen3.5 requires mindspeed and flash-linear-attention. '
-                'Install them with: pip install flash-linear-attention and mindspeed from source code via https://gitcode.com/Ascend/MindSpeed.'
-                'Other NPU patches will still apply.')
+            logger.warning('[NPU] [FLA] mindspeed or flash-linear-attention is not installed, '
+                           'or mindspeed version smaller than 0.12.1; '
+                           'FLA patch for Qwen3.5 requires mindspeed and flash-linear-attention. '
+                           'Install them with: pip install flash-linear-attention, '
+                           'and install mindspeed from source code via '
+                           'https://gitcode.com/Ascend/MindSpeed. '
+                           'Other NPU patches will still apply.')
     return bundle