[Feature] use phi permute/unpermute & rm swiglu (#6361)

* tp文字输出正常

* B eb5 mini文字输出正常

* eb5mini ep B卡 文字输出正常

* default use phi moe op

* stash

* tp H卡正常

* ep ok

* rm debug

* rm debug tool

* rm del ffn_out

* rm swiglu

* add envs to swiglu

* merge dev

* fix ci baseline

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix ci baseline 2

---------

Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

fastdeploy/model_executor/models/glm4_moe.py

@@ -25,6 +25,7 @@ from paddleformers.transformers import PretrainedModel
 from paddleformers.utils.log import logger
 
 from fastdeploy.config import FDConfig
+from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce
 from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.model_executor.graph_optimization.decorator import (
     support_graph_optimization,
@@ -161,6 +162,7 @@ class Glm4Moe(nn.Layer):
 
         self.experts = FusedMoE(
             fd_config,
+            reduce_results=False,
             renormalize=self.norm_topk_prob,
             moe_intermediate_size=fd_config.model_config.moe_intermediate_size,
             num_experts=fd_config.model_config.n_routed_experts,
@@ -181,14 +183,21 @@ class Glm4Moe(nn.Layer):
                 intermediate_size=shared_experts_intermediate_size,
                 layer_id=layer_id,
                 prefix=f"{prefix}.shared_experts",
+                reduce_results=False,
             )
 
     def forward(self, x, forward_meta: ForwardMeta = None):
+        # Both experts and shared_experts return partial sums (no all-reduce).
+        # Combine them first, then do a single all-reduce — eliminating one
+        # collective communication compared to the naive sequential approach.
+        # NOTE: only valid for pure-TP mode (use_ep=False). In EP or EP+TP modes
+        # FusedMoE uses all-to-all internally and already produces a full result,
+        # so the extra all-reduce must be skipped to avoid double-reduction.
         out = self.experts(x, self.gate, forward_meta)
         if self.n_shared_experts > 0:
-            shared_experts_out = self.shared_experts(x)
-            out = out + shared_experts_out
-
+            out = out + self.shared_experts(x)
+        if self.use_tp and not self.use_ep:
+            out = tensor_model_parallel_all_reduce(out, self.tp_group)
         return out
 
 
@@ -535,7 +544,9 @@ class Glm4MoeForCausalLM(ModelForCasualLM):
         forward_meta: ForwardMeta,
     ):
         """ """
+        paddle.cuda.nvtx.range_push("GLM4_MOE_BF")
         hidden_states = self.model(ids_remove_padding=ids_remove_padding, forward_meta=forward_meta)
+        paddle.cuda.nvtx.range_pop()
 
         return hidden_states