Merge branch 'v0.9.2-dev-ep-bug-fix' into 'v0.9.2-dev'

fix: 修复ep的变量未定义

See merge request dcutoolkit/deeplearing/vllm!423

vllm/model_executor/models/deepseek_v2.py

@@ -421,13 +421,9 @@ class DeepseekV2MoE(nn.Module):
                             # Fix FP16 overflow
                             # See DeepseekV2DecoderLayer for more details.
                             # fp16 mode not fused quant
-                            if i_q is not None:
-                                i_q=iqis[0]
-                                i_s=iqis[1]
                             final_hidden_states = self.experts(hidden_states=hidden_states,
                                                             router_logits=router_logits,
                                                             i_q=i_q, i_s=i_s)
-                    
                         if shared_output is not None:
                             if hidden_states.dtype != torch.float16:
                                 final_hidden_states = final_hidden_states + shared_output
@@ -468,13 +464,11 @@ class DeepseekV2MoE(nn.Module):
                             assert shared_output is not None
                             final_hidden_states += (shared_output * (1. / self.routed_scaling_factor))
                     else:
-                        if i_q is not None:
-                            i_q=iqis[0]
-                            i_s=iqis[1]
+                        if iqis is not None:
+                            i_q, i_s = iqis
                         final_hidden_states = self.experts(hidden_states=hidden_states,
                                                            router_logits=router_logits,
                                                            i_q=i_q, i_s=i_s)
-
                         if shared_output is not None:
                             if hidden_states.dtype != torch.float16:
                                 final_hidden_states = final_hidden_states + shared_output
@@ -483,7 +477,6 @@ class DeepseekV2MoE(nn.Module):
                                 # See DeepseekV2DecoderLayer for more details.
                                 final_hidden_states = final_hidden_states + shared_output \
                                     * (1. / self.routed_scaling_factor)
-
         if self.tp_size > 1:
             if envs.VLLM_ENABLE_TBO:
                 final_hidden_states = self.tbo_all_reduce(final_hidden_states)
@@ -491,7 +484,6 @@ class DeepseekV2MoE(nn.Module):
                 final_hidden_states = (
                     self.experts.maybe_all_reduce_tensor_model_parallel(
                         final_hidden_states))
-
             return final_hidden_states.view(num_tokens, hidden_dim)