fix mypy check

Author: RangiLyu

xtuner/v1/float8/float8_handler.py

@@ -36,8 +36,8 @@ def default_grouped_linear_filter_fn(mod: nn.Module, fqn: str):
 
 # handler 要跟 Engine 一一对应？
 class Float8Handler:
-    scaling_granularity_gemm: ScalingGranularity
-    scaling_granularity_grouped_gemm: ScalingGranularity
+    scaling_granularity_gemm: Optional[ScalingGranularity]
+    scaling_granularity_grouped_gemm: Optional[ScalingGranularity]
     fsdp_mesh: Optional[DeviceMesh] = None
     tilewise_reduce_mesh_devided_64: Optional[DeviceMesh] = None
     tilewise_reduce_mesh_mapping: Dict[Tuple[int, int], DeviceMesh] = {}
@@ -64,9 +64,9 @@ def __init__(
         assert scaling_granularity_gemm in (ScalingGranularity.TILEWISE, ScalingGranularity.TENSORWISE) or (
             scaling_granularity_gemm is None
         ), "scaling_granularity_gemm must be TILEWISE or TENSORWISE."
-        assert scaling_granularity_grouped_gemm in (ScalingGranularity.TILEWISE, ScalingGranularity.TENSORWISE), (
-            "scaling_granularity_grouped_gemm must be TILEWISE or TENSORWISE."
-        )
+        assert scaling_granularity_grouped_gemm in (ScalingGranularity.TILEWISE, ScalingGranularity.TENSORWISE) or (
+            scaling_granularity_grouped_gemm is None
+        ), "scaling_granularity_grouped_gemm must be TILEWISE or TENSORWISE."
 
         self.scaling_granularity_gemm = scaling_granularity_gemm
         self.scaling_granularity_grouped_gemm = scaling_granularity_grouped_gemm

xtuner/v1/ray/rollout/lmdeploy.py

@@ -234,7 +234,7 @@ def _transform_rollout_config_to_server_configs(self) -> Namespace:
         lmdeploy_config_kwargs["uvicorn_log_level"] = lmdeploy_config_kwargs.pop("uvicorn_log_level", "ERROR")
         lmdeploy_config_kwargs["tm_log_level"] = lmdeploy_config_kwargs.pop("tm_log_level", "ERROR")
 
-        extra_engine_config = {}
+        extra_engine_config: dict[str, Any] = {}
         if backend == "pytorch" and self.config.enable_return_routed_experts:
             extra_engine_config["enable_return_routed_experts"] = True
         if backend == "pytorch" and self.config.router_n_groups:

xtuner/v1/rl/base/worker.py

@@ -410,8 +410,8 @@ def fit(self, data_batches: list[WorkerInputItem], rollout_idx: int):
 
         # old logprobs are inplaced updated in compute_actor_logprobs
         if isinstance(self.config.model_cfg, BaseComposeConfig):
-            if self._engine.llm_float8_handler is not None and self._engine.llm_float8_handler.enabled:
-                self._engine.llm_float8_handler.precompute_float8_dynamic_scale_for_fsdp(
+            if self._engine.llm_float8_handler is not None and self._engine.llm_float8_handler.enabled:  # type: ignore [attr-defined]
+                self._engine.llm_float8_handler.precompute_float8_dynamic_scale_for_fsdp(  # type: ignore [attr-defined]
                     self._engine.model.language_model
                 )
         else: