ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Update pp llama.py to save memory (#11233)

Browse source
This commit is contained in:
Wang, Jian4 2024-06-07 13:18:16 +08:00 committed by GitHub
parent ef8e9b2ecd
commit 6f2684e5c9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 4 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
python/llm/example/GPU/Pipeline-Parallel-FastAPI/llama_models.py
View file

@ -50,8 +50,10 @@ class LlamaModel(LlamaPreTrainedModel):
self.padding_idx = config.pad_token_id self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size self.vocab_size = config.vocab_size
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) if self.pp_config.is_head:
self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
if self.pp_config.is_tail:
self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def get_input_embeddings(self): def get_input_embeddings(self):
@ -259,7 +261,6 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
if self.pp_config.is_tail: if self.pp_config.is_tail:
hidden_states = outputs[0] hidden_states = outputs[0]
logits = self.lm_head(hidden_states) logits = self.lm_head(hidden_states)
logits = logits.float()
loss = None loss = None
if labels is not None: if labels is not None: