Merge branch 'v0.9.2-dev-ds' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.9.2-dev-ds

vllm/model_executor/models/deepseek_v2.py

@@ -43,8 +43,8 @@ from vllm.distributed import (get_ep_group, get_pp_group, get_dp_group,
                               get_tensor_model_parallel_world_size)
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.fused_moe import FusedMoE
-from vllm.model_executor.layers.fused_moe.ep_moe.layer import EPMoE
-from vllm.model_executor.layers.fused_moe.ep_moe.ep_moe_utlis import EPSharedExperts
+from vllm.model_executor.layers.fused_moe.mori_moe.layer import MoriMoE
+from vllm.model_executor.layers.fused_moe.mori_moe.ep_moe_utlis import EPSharedExperts
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                MergedColumnParallelLinear,
@@ -167,9 +167,10 @@ class DeepseekV2MoE(nn.Module):
                                     self.n_local_physical_experts)
 
         dp_size = get_dp_group().world_size
-        self.use_mori_ep = envs.VLLM_USE_MORI_EP and dp_size > 1 and parallel_config.enable_expert_parallel
+        self.use_mori_ep = parallel_config.enable_expert_parallel and dp_size > 1 and envs.VLLM_ALL2ALL_BACKEND == 'mori'
+        self.enable_expert_parallel = parallel_config.enable_expert_parallel
 
-        moe_cls = FusedMoE if not self.use_mori_ep else EPMoE
+        moe_cls = FusedMoE if not self.use_mori_ep else MoriMoE
         self.experts = moe_cls(
             num_experts=config.n_routed_experts,
             top_k=config.num_experts_per_tok,
@@ -224,7 +225,7 @@ class DeepseekV2MoE(nn.Module):
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
 
-        if not self.use_mori_ep:
+        if not self.enable_expert_parallel:
             if envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD:
                 final_hidden_states = self.experts(
                         hidden_states=hidden_states,
@@ -248,6 +249,20 @@ class DeepseekV2MoE(nn.Module):
                         # See DeepseekV2DecoderLayer for more details.
                         final_hidden_states = final_hidden_states + shared_output \
                             * (1. / self.routed_scaling_factor)
+        else:
+            if not self.use_mori_ep:
+                final_hidden_states = self.experts(
+                        hidden_states=hidden_states,
+                        router_logits=router_logits)
+                    
+                if shared_output is not None:
+                    if hidden_states.dtype != torch.float16:
+                        final_hidden_states = final_hidden_states + shared_output
+                    else:
+                        # Fix FP16 overflow
+                        # See DeepseekV2DecoderLayer for more details.
+                        final_hidden_states = final_hidden_states + shared_output \
+                            * (1. / self.routed_scaling_factor)
             else:
                 final_hidden_states = self.experts(hidden_states=hidden_states,
                                                 router_logits=router_logits)
@@ -927,7 +942,7 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
 
         parallel_config = vllm_config.parallel_config
         dp_size = get_dp_group().world_size
-        self.use_mori_ep = envs.VLLM_USE_MORI_EP and dp_size > 1 and parallel_config.enable_expert_parallel
+        self.use_mori_ep = envs.VLLM_ALL2ALL_BACKEND == 'mori' and dp_size > 1 and parallel_config.enable_expert_parallel
 
     def set_eplb_state(
         self,

vllm/v1/spec_decode/eagle.py

@@ -515,6 +515,9 @@ class EagleProposer:
 
         if self.dp_size > 1 and self.enable_expert_parallel and self.num_speculative_tokens > 1:
             for _ in range(self.num_speculative_tokens - 1):
+                with set_forward_context(attn_metadata,
+                                 self.vllm_config,
+                                 num_tokens=num_tokens):
                     self.model(
                         self.input_ids[:num_tokens],
                         self.positions[:num_tokens],

vllm/v1/worker/gpu_model_runner.py

@@ -323,9 +323,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         # from the KV cache of `shared_kv_cache_layers[layer_name]`.
         self.shared_kv_cache_layers: dict[str, str] = {}
 
-        dp_size = self.vllm_config.parallel_config.data_parallel_size
-        self.use_mori_ep = envs.VLLM_USE_MORI_EP and dp_size > 1 and parallel_config.enable_expert_parallel
-
     def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
         """
         Update the order of requests in the batch based on the attention
@@ -1238,7 +1235,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         # TODO(tms) : There are many cases where padding is enabled for
         # prefills, causing unnecessary and excessive padding of activations.
 
-        if dp_size == 1 or self.vllm_config.model_config.enforce_eager or self.use_mori_ep:
+        if dp_size == 1 or self.vllm_config.model_config.enforce_eager or envs.VLLM_ALL2ALL_BACKEND == 'naive':
             # Early exit.
             return 0, None