DeepSeek V2/V3/R1 only place `lm_head` on last pp rank (#13833)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

vllm/model_executor/models/deepseek_v2.py

@@ -636,9 +636,12 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP):
         self.quant_config = quant_config
         self.model = DeepseekV2Model(vllm_config=vllm_config,
                                      prefix=maybe_prefix(prefix, "model"))
-        self.lm_head = ParallelLMHead(config.vocab_size,
-                                      config.hidden_size,
-                                      quant_config=quant_config)
+        if get_pp_group().is_last_rank:
+            self.lm_head = ParallelLMHead(config.vocab_size,
+                                          config.hidden_size,
+                                          quant_config=quant_config)
+        else:
+            self.lm_head = PPMissingLayer()
         self.logits_processor = LogitsProcessor(config.vocab_size)
         self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (