Fix dpsk-r1-fp4 tp8 by reverting two commits (#13162 and #13341) (#13348)

Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>

python/sglang/srt/eplb/expert_location.py

@@ -284,17 +284,9 @@ class ExpertLocationMetadata:
     # -------------------------------- usage ------------------------------------
 
     def logical_to_all_physical(
-        self,
+        self, layer_id: int, logical_expert_id: int
-        layer_id: int,
-        logical_expert_id: int,
-        require_global_experts: bool = False,
     ) -> List[int]:
         # Use CPU copy to avoid GPU→CPU sync on every call, which is expensive in update weights scenario
-        if require_global_experts:
-            num_physical_experts = self.logical_to_all_physical_map_cpu[layer_id].shape[
-                -1
-            ]
-            return list(torch.arange(0, num_physical_experts))
         return [
             physical_expert_id
             for physical_expert_id in self.logical_to_all_physical_map_cpu[
@@ -363,10 +355,14 @@ def _compute_logical_to_all_physical_map(
                 )
 
                 # Replace by the nearest physical expert
-                if nearest_expert != -1:
+                mapped_physical_experts = logical_to_all_physical_map[layer_id][
-                    logical_to_all_physical_map[layer_id][logical_expert_id] = [
+                    logical_expert_id
-                        nearest_expert
+                ]
-                    ]
+                if (
+                    nearest_expert != -1
+                    and nearest_expert not in mapped_physical_experts
+                ):
+                    mapped_physical_experts[0] = nearest_expert
 
     logical_to_all_physical_map = _pad_nested_array(
         logical_to_all_physical_map, pad_value=-1

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -539,12 +539,9 @@ class FusedMoE(torch.nn.Module):
             # This is a shared expert.
             physical_expert_ids = [expert_id]
         else:
-            require_global_experts = getattr(
-                param, "_sglang_require_global_experts", False
-            )
             physical_expert_ids = (
                 global_expert_location_metadata.logical_to_all_physical(
-                    self.layer_id, expert_id, require_global_experts
+                    self.layer_id, expert_id
                 )
             )
 
@@ -1129,6 +1126,7 @@ class FlashInferFP4MoE(FusedMoE):
             local_expert_offset=self.moe_ep_rank * self.num_local_experts,
             local_num_experts=self.num_local_experts,
             routed_scaling_factor=self.moe_runner_config.routed_scaling_factor,
+            tile_tokens_dim=None,
             routing_method_type=RoutingMethodType.DeepSeekV3,
             do_finalize=True,
             output=symm_output,

python/sglang/srt/layers/quantization/fp8.py

@@ -1245,6 +1245,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 routed_scaling_factor=(
                     routed_scaling_factor if routed_scaling_factor is not None else 1.0
                 ),
+                tile_tokens_dim=None,
                 routing_method_type=routing_method_type,
                 use_shuffled_weight=False,
             )

python/sglang/srt/layers/quantization/modelopt_quant.py

@@ -739,6 +739,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
                         else 1.0
                     ),
                     use_routing_scales_on_input=use_routing_scales_on_input,
+                    tile_tokens_dim=None,
                     routing_method_type=routing_method_type,
                 )
 

python/sglang/srt/layers/quantization/mxfp4.py

@@ -681,6 +681,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 layer.moe_ep_rank * layer.num_local_experts,  # local_expert_offset
                 layer.num_local_experts,  # local num experts
                 None,
+                None,  # tile_tokens_dim
                 1,  # routing_method_type, renormalize
                 True,  # do finalize
                 output=symm_output,