Drop redundant num_experts from RoutedExperts

num_experts is intrinsic to the grouped-GEMM (GroupedExperts, which sizes
w1_EFD/w2_EDF/w3_EFD); RoutedExperts is a pure wrapper (token_dispatcher +
inner_experts) and carried a duplicate. Read it from the source of truth
(routed_experts.inner_experts.num_experts) in MoE buffer sizing and the HF
export instead.

Author: wwwjn

torchtitan/experiments/rl/models/vllm_registry.py

@@ -167,7 +167,7 @@ def model_spec_to_hf_config_dict(spec: ModelSpec) -> dict[str, Any]:
 
     if moe is not None:
         # Presence required: >0 toggles MoE/EP branches.
-        hf["num_experts"] = moe.routed_experts.num_experts
+        hf["num_experts"] = moe.routed_experts.inner_experts.num_experts
         # Unused: only per-model loaders (qwen3_moe, deepseek_v2, ...) and v1/metrics/perf.py (off) read these.
         hf[
             "num_experts_per_tok"

torchtitan/models/common/config_utils.py

@@ -392,7 +392,6 @@ def make_routed_experts_config(
     inference override targets the latter.
     """
     return RoutedExperts.Config(
-        num_experts=num_experts,
         inner_experts=GroupedExperts.Config(
             dim=dim,
             hidden_dim=hidden_dim,

torchtitan/models/common/moe.py

@@ -141,13 +141,11 @@ class RoutedExperts(Module):
 
     @dataclass(kw_only=True, slots=True)
     class Config(Module.Config):
-        num_experts: int
         inner_experts: GroupedExperts.Config
         token_dispatcher: LocalTokenDispatcher.Config
 
     def __init__(self, config: Config):
         super().__init__()
-        self.num_experts = config.num_experts
         self.inner_experts = config.inner_experts.build()
         self.token_dispatcher = config.token_dispatcher.build()
 
@@ -559,9 +557,9 @@ def _init_self_buffers(self, *, buffer_device: torch.device | None = None) -> No
 
         with torch.device(buffer_device):
             self.tokens_per_expert_E = torch.zeros(
-                self.routed_experts.num_experts, dtype=torch.float32
+                self.routed_experts.inner_experts.num_experts, dtype=torch.float32
             )
             if self.load_balance_coeff is not None:
                 self.expert_bias_E = torch.zeros(
-                    self.routed_experts.num_experts, dtype=torch.float32
+                    self.routed_experts.inner_experts.num_experts, dtype=torch.float32
                 )

torchtitan/models/gpt_oss/__init__.py

@@ -174,7 +174,6 @@ def _make_gptoss_experts_config(
         "mlp2_bias_ED": partial(nn.init.trunc_normal_, std=std),
     }
     return RoutedExperts.Config(
-        num_experts=num_experts,
         inner_experts=GptOssGroupedExperts.Config(
             dim=dim,
             hidden_dim=hidden_dim,