biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-02-26 01:24:57 +00:00
committed by GitHub
parent 07c4353057
commit 24679788ed
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/deepseek_v2.py
View File

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