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

[LoRA][1/N]Remove LoRA extra vocab (#28382)

Browse Source 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2025-11-12 03:06:21 +08:00
committed by GitHub
parent 8c32c6e4b4
commit 9d1c474704
65 changed files with 197 additions and 754 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
vllm/model_executor/models/gpt_bigcode.py
View File

@@ -207,18 +207,13 @@ class GPTBigCodeModel(nn.Module):
config = vllm_config.model_config.hf_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
self.config = config
assert not config.add_cross_attention
self.embed_dim = config.hidden_size
lora_vocab = (
(lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
if lora_config
else 0
)
self.vocab_size = config.vocab_size + lora_vocab
self.vocab_size = config.vocab_size
self.wte = VocabParallelEmbedding(
self.vocab_size, self.embed_dim, org_num_embeddings=config.vocab_size
)
@@ -290,10 +285,8 @@ class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
self.config = config
self.lora_config = lora_config
self.quant_config = quant_config
self.transformer = GPTBigCodeModel(
@@ -305,15 +298,10 @@ class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
self.lm_head = ParallelLMHead(
self.transformer.vocab_size,
self.transformer.embed_dim,
org_num_embeddings=self.config.vocab_size,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.unpadded_vocab_size = config.vocab_size
if lora_config:
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size
)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.transformer.make_empty_intermediate_tensors
)