[Feature] support glm tp+dp+ep

fastdeploy/model_executor/models/glm4_moe.py

@@ -25,7 +25,6 @@
 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,
@@ -160,7 +159,6 @@ def __init__(
 
         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,
@@ -174,23 +172,21 @@ def __init__(
             weight_key_map=weight_key_map,
         )
 
-        shared_experts_intermediate_size = self.n_shared_experts * fd_config.model_config.moe_intermediate_size
-
-        self.shared_experts = Glm4MoeMLP(
-            fd_config=fd_config,
-            intermediate_size=shared_experts_intermediate_size,
-            layer_id=layer_id,
-            prefix=f"{prefix}.shared_experts",
-            reduce_results=False,
-        )
+        if self.n_shared_experts > 0:
+            shared_experts_intermediate_size = self.n_shared_experts * fd_config.model_config.moe_intermediate_size
+            self.shared_experts = Glm4MoeMLP(
+                fd_config=fd_config,
+                intermediate_size=shared_experts_intermediate_size,
+                layer_id=layer_id,
+                prefix=f"{prefix}.shared_experts",
+            )
 
     def forward(self, x, forward_meta: ForwardMeta = None):
-        shared_experts_out = self.shared_experts(x)
         out = self.experts(x, self.gate, forward_meta)
-        out = out + shared_experts_out
-        # We do to TP all reduce after the sum of experts.
-        if self.tensor_parallel_size > 1:
-            out = tensor_model_parallel_all_reduce(out, self.tp_group)
+        if self.n_shared_experts > 0:
+            shared_experts_out = self.shared_experts(x)
+            out = out + shared_experts_out
+
         return out
 
 

tests/e2e/4cards_cases/test_GLM_45_AIR_tp4.py

@@ -181,7 +181,7 @@ def test_lm_head_fp32(api_url, headers, consistent_payload):
     # 校验返回内容与概率信息
     assert (
         resp_json["choices"][0]["message"]["content"]
-        == "\n<think>这个问题是关于牛顿的三大运动定律的。牛顿的三大运动定律是经典"
+        == "\n<think>我需要回答牛顿的三大运动定律是什么。牛顿的三大运动定律是经典"
     ), f"The response content is not as expected {resp_json['choices'][0]['message']['content']}."
 
 

tests/e2e/utils/rollout_routing_replay_test_utils.py

@@ -154,9 +154,11 @@ def check_routing_replay_chat_completion(openai_client, moe_layer_num: int, mode
     cur_save_routing_path = f"./R3_tmp/routing_replay_output_{model_name}/"
     model_path = os.getenv("MODEL_PATH")
     if model_path:
-        baseline_path = os.path.join(model_path, f"R3_BaseLine_dev_uint8/routing_replay_output_baseline_{model_name}")
+        baseline_path = os.path.join(
+            model_path, f"R3_BaseLine_dev_uint8_0205/routing_replay_output_baseline_{model_name}"
+        )
     else:
-        baseline_path = f"./R3_BaseLine_dev_uint8/routing_replay_output_baseline_{model_name}"
+        baseline_path = f"./R3_BaseLine_dev_uint8_0205/routing_replay_output_baseline_{model_name}"
     stream_baseline_path = os.path.join(baseline_path, "r3_chat_completion_stream")
 
     nonstream_baseline_path = os.path.join(baseline_path, "r3_chat_completion_nonstream")